initial commit
Browse files- README.md +0 -0
- app.py +159 -0
- easyeditor/__init__.py +2 -0
- easyeditor/models/README.md +6 -0
- easyeditor/models/__init__.py +1 -0
- easyeditor/models/__pycache__/__init__.cpython-39.pyc +0 -0
- easyeditor/models/grace/GRACE.py +445 -0
- easyeditor/models/grace/__init__.py +2 -0
- easyeditor/models/grace/__pycache__/GRACE.cpython-39.pyc +0 -0
- easyeditor/models/grace/__pycache__/__init__.cpython-39.pyc +0 -0
- easyeditor/models/grace/__pycache__/grace_hparams.cpython-39.pyc +0 -0
- easyeditor/models/grace/__pycache__/grace_main.cpython-39.pyc +0 -0
- easyeditor/models/grace/__pycache__/metrics.cpython-39.pyc +0 -0
- easyeditor/models/grace/__pycache__/utils.cpython-39.pyc +0 -0
- easyeditor/models/grace/grace_hparams.py +48 -0
- easyeditor/models/grace/grace_main.py +40 -0
- easyeditor/models/grace/metrics.py +59 -0
- easyeditor/models/grace/utils.py +86 -0
- easyeditor/util/__init__.py +2 -0
- easyeditor/util/__pycache__/__init__.cpython-39.pyc +0 -0
- easyeditor/util/__pycache__/hparams.cpython-39.pyc +0 -0
- easyeditor/util/__pycache__/logit_lens.cpython-39.pyc +0 -0
- easyeditor/util/__pycache__/nethook.cpython-39.pyc +0 -0
- easyeditor/util/alg_dict.py +45 -0
- easyeditor/util/alg_train_dict.py +9 -0
- easyeditor/util/generate.py +171 -0
- easyeditor/util/globals.py +43 -0
- easyeditor/util/hparams.py +46 -0
- easyeditor/util/logit_lens.py +97 -0
- easyeditor/util/nethook.py +451 -0
- easyeditor/util/perplexity.py +24 -0
- easyeditor/util/runningstats.py +1883 -0
- hparams/GRACE/README.md +19 -0
- hparams/GRACE/gpt2.yaml +19 -0
- hparams/config.yaml +6 -0
- requirements.txt +25 -0
- utils.py +42 -0
README.md
CHANGED
File without changes
|
app.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from utils import *
|
3 |
+
from transformers import pipeline
|
4 |
+
|
5 |
+
css = """
|
6 |
+
|
7 |
+
"""
|
8 |
+
|
9 |
+
ori_model = None
|
10 |
+
edit_model = None
|
11 |
+
# input=None
|
12 |
+
|
13 |
+
def slowly_reverse(word, progress=gr.Progress()):
|
14 |
+
progress(0, desc="Starting")
|
15 |
+
time.sleep(1)
|
16 |
+
progress(0.05)
|
17 |
+
new_string = ""
|
18 |
+
for letter in progress.tqdm(word, desc="Editing"):
|
19 |
+
time.sleep(0.25)
|
20 |
+
new_string = letter + new_string
|
21 |
+
return new_string
|
22 |
+
|
23 |
+
with gr.Blocks(css=css,theme=gr.themes.Soft(text_size="sm")) as demo:
|
24 |
+
with gr.Row(equal_height=True):
|
25 |
+
gr.HTML(
|
26 |
+
"""
|
27 |
+
<div>
|
28 |
+
<h1>🔧EasyEdit: An Easy-to-use Knowledge Editing Framework for Large Language Models</h1>
|
29 |
+
|
30 |
+
<p>
|
31 |
+
📑[<a href="https://huggingface.co/papers/2308.07269">Paper</a>]
|
32 |
+
👨💻[<a href="https://github.com/zjunlp/EasyEdit" target="_blank"><span class="icon"><i class="fab fa-github"></i></span>Code</a>]
|
33 |
+
📄[<a href="https://zjunlp.gitbook.io/easyedit">Docs</a>]
|
34 |
+
🤗[<a href="https://huggingface.co/spaces/zjunlp/EasyEdit" target="_blank">Demo</a>]
|
35 |
+
[<a href="https://arxiv.org/abs/2211.11031">via GRACE</a>]
|
36 |
+
</p>
|
37 |
+
</div>
|
38 |
+
"""
|
39 |
+
)
|
40 |
+
# gr.HTML("""<div style="text-align: center; margin: 0 auto;"><p><h1> Knowledge Editing</h1></div>""")
|
41 |
+
|
42 |
+
# with gr.Row():
|
43 |
+
# gr.Markdown("<p align='center'><a href='https://github.com/zjunlp/EasyEdit'>🔧https://github.com/zjunlp/EasyEdit</a></p>")
|
44 |
+
|
45 |
+
with gr.Row():
|
46 |
+
gr.Markdown("#### Knowledge editing aims to subtly inject/edit updated knowledge or adjust undesirable behaviors, while minimizing the impact on unrelated inputs.")
|
47 |
+
|
48 |
+
|
49 |
+
with gr.Row():
|
50 |
+
prompt = gr.Textbox(label="Edit Prompt")
|
51 |
+
target_new = gr.Textbox(label="Edit Target New")
|
52 |
+
with gr.Row():
|
53 |
+
num_steps = gr.Slider(10, 100, value=20, step=1, label='Edit Steps')
|
54 |
+
replacement = gr.Dropdown(
|
55 |
+
choices=["replace_last", "replace_all", "replace_prompt"],
|
56 |
+
value="replace_last",
|
57 |
+
label="Replacement",
|
58 |
+
)
|
59 |
+
with gr.Row():
|
60 |
+
button4clear = gr.Button("Clear")
|
61 |
+
button4edit = gr.Button("Edit",variant="primary")
|
62 |
+
with gr.Row():
|
63 |
+
examples = gr.Examples(
|
64 |
+
examples=[
|
65 |
+
["Who is the architect for Toodyay Fire Station?","Wong Tung & Sons"],
|
66 |
+
["Who is Claire Clairmont\'s sister?","Clairmont-Mayer"],
|
67 |
+
["Which fictional universe is Chlorophyll Kid part of?","Image Universe"]
|
68 |
+
],
|
69 |
+
examples_per_page=3,
|
70 |
+
inputs=[prompt,target_new],
|
71 |
+
)
|
72 |
+
# with gr.Row():
|
73 |
+
# input_text = gr.Textbox(label="Status Information",value="Model editing may take about a minute, please be patient.")
|
74 |
+
with gr.Row():
|
75 |
+
gr.HTML(
|
76 |
+
"""
|
77 |
+
<h3>Reliability Evaluation</h3>
|
78 |
+
"""
|
79 |
+
)
|
80 |
+
with gr.Row():
|
81 |
+
input = gr.Textbox(label="Input Text")
|
82 |
+
with gr.Row():
|
83 |
+
with gr.Column():
|
84 |
+
button4gen_ori=gr.HighlightedText(
|
85 |
+
label="original output",
|
86 |
+
combine_adjacent=True,
|
87 |
+
show_legend=False,
|
88 |
+
color_map={"output": "yellow"},
|
89 |
+
)
|
90 |
+
with gr.Column():
|
91 |
+
button4gen_edit=gr.HighlightedText(
|
92 |
+
label="edited output",
|
93 |
+
combine_adjacent=True,
|
94 |
+
show_legend=False,
|
95 |
+
color_map={"output": "yellow"},
|
96 |
+
)
|
97 |
+
with gr.Row():
|
98 |
+
button4gen = gr.Button("Generate",variant="primary")
|
99 |
+
|
100 |
+
with gr.Row():
|
101 |
+
gr.HTML(
|
102 |
+
"""
|
103 |
+
<h3>Locality Evaluation</h3>
|
104 |
+
"""
|
105 |
+
)
|
106 |
+
with gr.Row():
|
107 |
+
loc_input = gr.Dropdown(
|
108 |
+
choices=[
|
109 |
+
"who sang the theme song for laverne and shirley",
|
110 |
+
"when does the last episode of adventure time air",
|
111 |
+
"who plays alec ramsay in the black stallion",
|
112 |
+
"where did an independence movement occur because of the congress of vienna",
|
113 |
+
"where is the ucla usc game being played"
|
114 |
+
],
|
115 |
+
value="where is the ucla usc game being played",
|
116 |
+
label="Unrelated Input Text",
|
117 |
+
)
|
118 |
+
with gr.Row():
|
119 |
+
with gr.Column():
|
120 |
+
button4gen_loc_ori=gr.HighlightedText(
|
121 |
+
label="original output",
|
122 |
+
combine_adjacent=True,
|
123 |
+
show_legend=False,
|
124 |
+
color_map={"output": "green"},
|
125 |
+
)
|
126 |
+
with gr.Column():
|
127 |
+
button4gen_loc_edit=gr.HighlightedText(
|
128 |
+
label="edited output",
|
129 |
+
combine_adjacent=True,
|
130 |
+
show_legend=False,
|
131 |
+
color_map={"output": "green"},
|
132 |
+
)
|
133 |
+
with gr.Row():
|
134 |
+
button4locgen = gr.Button("Generate",variant="primary")
|
135 |
+
|
136 |
+
button4clear.click(lambda: ("", ""), outputs=[prompt,target_new])
|
137 |
+
button4edit.click(fn=edit, inputs=[prompt,target_new, num_steps, replacement], outputs=input)
|
138 |
+
button4gen.click(fn=generate, inputs=[input, target_new], outputs=[button4gen_ori, button4gen_edit])
|
139 |
+
button4locgen.click(fn=generate, inputs=loc_input, outputs=[button4gen_loc_ori, button4gen_loc_edit])
|
140 |
+
|
141 |
+
|
142 |
+
with gr.Accordion("Citation", open=False):
|
143 |
+
gr.Markdown(
|
144 |
+
"""
|
145 |
+
```bibtex
|
146 |
+
@misc{wang2023easyedit,
|
147 |
+
title={EasyEdit: An Easy-to-use Knowledge Editing Framework for Large Language Models},
|
148 |
+
author={Peng Wang and Ningyu Zhang and Xin Xie and Yunzhi Yao and Bozhong Tian and Mengru Wang and Zekun Xi and Siyuan Cheng and Kangwei Liu and Guozhou Zheng and Huajun Chen},
|
149 |
+
year={2023},
|
150 |
+
eprint={2308.07269},
|
151 |
+
archivePrefix={arXiv},
|
152 |
+
primaryClass={cs.CL}
|
153 |
+
}
|
154 |
+
```
|
155 |
+
"""
|
156 |
+
)
|
157 |
+
|
158 |
+
|
159 |
+
demo.launch()
|
easyeditor/__init__.py
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
from .models import *
|
2 |
+
from .util import *
|
easyeditor/models/README.md
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
We compare ROME against several open sourced state-of-the-art model editors. All are implemented in their respective folders. Implementations other than FT/FT+L are adapted from third parties.
|
2 |
+
- Fine-Tuning (`ft`): Direct fine-tuning.
|
3 |
+
- Constrained Fine-Tuning (`ft`): FT with $L_\infty$ norm constraint. Inspired by Zhu et al. [[Paper]](https://arxiv.org/abs/2012.00363)
|
4 |
+
- Knowledge Neurons (`kn`): Dai et al. [[Code]](https://github.com/EleutherAI/knowledge-neurons) [[Paper]](https://arxiv.org/abs/2104.08696)
|
5 |
+
- Knowledge Editor (`efk`): De Cao et al. [[Code]](https://github.com/eric-mitchell/mend) [[Paper]](https://arxiv.org/abs/2104.08164)
|
6 |
+
- Model Editor Networks with Gradient Decomposition (`mend`): Mitchell et al. [[Code]](https://github.com/eric-mitchell/mend) [[Paper]](https://arxiv.org/abs/2110.11309)
|
easyeditor/models/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .grace import *
|
easyeditor/models/__pycache__/__init__.cpython-39.pyc
ADDED
Binary file (172 Bytes). View file
|
|
easyeditor/models/grace/GRACE.py
ADDED
@@ -0,0 +1,445 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# import torch
|
2 |
+
# from .utils import parent_module, brackets_to_periods
|
3 |
+
# import transformers
|
4 |
+
# import os
|
5 |
+
# os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
|
6 |
+
|
7 |
+
# def euc(query, key):
|
8 |
+
# # Euclidean distance
|
9 |
+
# if len(key.shape) < 2:
|
10 |
+
# key = key.view(1, -1)
|
11 |
+
# return torch.cdist(key, query, p=2)
|
12 |
+
|
13 |
+
# def perturb_values(chosen_value, num_pert, device):
|
14 |
+
# # Create a bunch of noised versions of the value, then create batch, then train value
|
15 |
+
# chosen_value = chosen_value
|
16 |
+
# noise = torch.normal(0, 1, chosen_value.shape, device=device)
|
17 |
+
# noise[0] = noise[0]*0
|
18 |
+
# noise.requires_grad = True
|
19 |
+
# chosen_value = chosen_value + noise
|
20 |
+
# return chosen_value
|
21 |
+
|
22 |
+
# class GRACE(torch.nn.Module):
|
23 |
+
# def __init__(self, config, model, device):
|
24 |
+
# super(GRACE, self).__init__()
|
25 |
+
# self.config = config
|
26 |
+
# self.log_dict = {}
|
27 |
+
# self.model = model
|
28 |
+
# # self.tokenizer = model.tokenizer
|
29 |
+
# layer = config.inner_params[0]
|
30 |
+
# self.device = device
|
31 |
+
|
32 |
+
# # --- ensure proper formatting (GRACE edits ~layers~ not weights matrices) ---
|
33 |
+
# suffixes = [".weight", ".bias"]
|
34 |
+
# self.layer = layer.rsplit(".", 1)[0] if any(layer.endswith(x) for x in suffixes) else layer
|
35 |
+
|
36 |
+
# for n, p in self.model.named_parameters():
|
37 |
+
# p.requires_grad = False
|
38 |
+
|
39 |
+
# if isinstance(self.model, transformers.models.gpt2.modeling_gpt2.GPT2LMHeadModel):
|
40 |
+
# transpose = False
|
41 |
+
# else:
|
42 |
+
# transpose = True
|
43 |
+
|
44 |
+
# # --- Add GRACE to chosen layers ---
|
45 |
+
# edit_module = parent_module(self.model, brackets_to_periods(self.layer))
|
46 |
+
# layer_name = self.layer.rsplit(".", 1)[-1]
|
47 |
+
# original_layer = getattr(edit_module, layer_name)
|
48 |
+
|
49 |
+
# if type(original_layer) is not GRACEAdapter:
|
50 |
+
# setattr(edit_module, layer_name, GRACEAdapter(config, original_layer, transpose=transpose).to(self.device))
|
51 |
+
|
52 |
+
# def __call__(self, **kwargs):
|
53 |
+
# # if self.config.task == "hallucination":
|
54 |
+
# # print(kwargs)
|
55 |
+
# # key_id = (kwargs["labels"] == -100).sum() - 1
|
56 |
+
# # setattr(eval(f"self.model.{self.layer}"), "key_id", key_id) # Tell GRACE which token to use for its query (default is the last token)
|
57 |
+
# return self.model(**kwargs)
|
58 |
+
|
59 |
+
# def generate(self, *args, **kwargs):
|
60 |
+
# setattr(eval(f"self.model.{self.layer}"), "key_id", -1)
|
61 |
+
# return self.model.generate(*args, **kwargs)
|
62 |
+
|
63 |
+
# def edit(self, config, tokens):
|
64 |
+
# key_id = (tokens["labels"] == -100).sum() - 1
|
65 |
+
# setattr(eval(f"self.model.{self.layer}"), "key_id", key_id)
|
66 |
+
|
67 |
+
# # --- pass edit label, training mode, and key_id into GRACE ---
|
68 |
+
# setattr(eval(f"self.model.{self.layer}"), "training", True)
|
69 |
+
# setattr(eval(f"self.model.{self.layer}"), "edit_label", tokens["labels"])
|
70 |
+
|
71 |
+
# self.losses = []
|
72 |
+
# # --- train GRACE value ---
|
73 |
+
# for i in range(config.n_iter):
|
74 |
+
# # --- insert iteration into each layer (only initiate keys on iteration 1) ---
|
75 |
+
# setattr(eval(f"self.model.{self.layer}"), "iter", i)
|
76 |
+
|
77 |
+
# # --- pass tokens through model (including through the GRACE layer) ---
|
78 |
+
# outputs = self.model(**tokens)
|
79 |
+
# if i == 0:
|
80 |
+
# # --- we only need to create an optimizer for the first iteration (but forward pass instantiates the key, so optimzer is passed after first inference) ---
|
81 |
+
# optimizer = torch.optim.Adam(self.model.parameters(), config.edit_lr)
|
82 |
+
# loss = outputs.loss
|
83 |
+
# loss.backward()
|
84 |
+
# optimizer.step()
|
85 |
+
# optimizer.zero_grad()
|
86 |
+
# self.losses.append(loss.detach().cpu().numpy())
|
87 |
+
|
88 |
+
# self.loss = loss # Log final loss
|
89 |
+
|
90 |
+
# # --- pull out info we want to log from the GRACE layer ---
|
91 |
+
# setattr(eval(f"self.model.{self.layer}"), "training", False)
|
92 |
+
# chosen_key = getattr(eval(f"self.model.{self.layer}"), "chosen_key")
|
93 |
+
# nkeys = len(getattr(eval(f"self.model.{self.layer}"), "keys"))
|
94 |
+
|
95 |
+
# self.log_dict["chosen_key"] = chosen_key
|
96 |
+
# self.log_dict["nkeys"] = nkeys
|
97 |
+
|
98 |
+
# class GRACEAdapter(torch.nn.Module):
|
99 |
+
# def __init__(self, config, layer, transpose):
|
100 |
+
# super(GRACEAdapter, self).__init__()
|
101 |
+
|
102 |
+
# self.layer = layer
|
103 |
+
# self.weight = self.layer.weight
|
104 |
+
# self.init_epsilon = config.eps
|
105 |
+
# self.dist_fn = config.dist_fn
|
106 |
+
# self.replacement = config.replacement
|
107 |
+
# self.device = layer.weight.device
|
108 |
+
# self.config = config
|
109 |
+
# self.num_pert = config.num_pert
|
110 |
+
# self.key_id = -1
|
111 |
+
# self.ensure_replace_token_loc = False
|
112 |
+
|
113 |
+
# if transpose:
|
114 |
+
# self.key_shape = layer.weight.shape[1]
|
115 |
+
# self.value_shape = layer.weight.shape[0]
|
116 |
+
# else:
|
117 |
+
# self.key_shape = layer.weight.shape[0]
|
118 |
+
# self.value_shape = layer.weight.shape[1]
|
119 |
+
# self.training = False
|
120 |
+
|
121 |
+
# def add_key(self, new_key, new_value):
|
122 |
+
# keys = torch.vstack([self.keys, new_key.detach()]) # Add new key to list of keys
|
123 |
+
|
124 |
+
# values = torch.nn.Parameter(torch.vstack([self.values, new_value]), requires_grad=True) # Add new value to list of values
|
125 |
+
|
126 |
+
# new_epsilon = torch.tensor(self.init_epsilon, device=self.device).view(1)
|
127 |
+
# epsilons = torch.vstack([self.epsilons, new_epsilon]) # Add new epsilon to list of epsilons
|
128 |
+
|
129 |
+
# key_labels = self.key_labels + [self.edit_label] # Add new key_label to list of key_labels
|
130 |
+
|
131 |
+
# return keys, values, epsilons, key_labels
|
132 |
+
|
133 |
+
# def init_key_value(self, query, value):
|
134 |
+
# key = query.detach()
|
135 |
+
# epsilon = torch.tensor(self.init_epsilon, device=self.device, requires_grad=False).view(1)
|
136 |
+
# key_label = [self.edit_label]
|
137 |
+
# return key, value, epsilon, key_label
|
138 |
+
|
139 |
+
# def label_match(self, edit_label, key_label):
|
140 |
+
# return edit_label.float().mean() == key_label.float().mean()
|
141 |
+
|
142 |
+
# def split_epsilons_in_half(self, nearest_key, smallest_distance):
|
143 |
+
# self.epsilons[nearest_key] = (smallest_distance / 2) - 1e-5 # Cut nearest epsilon in half
|
144 |
+
# self.epsilons[-1] = smallest_distance / 2 # Cut new epsilon in half
|
145 |
+
|
146 |
+
# def forward(self, *args):
|
147 |
+
# # Run layer forward and save what it would have returned for this instance
|
148 |
+
# layer_out = self.layer(*args)
|
149 |
+
|
150 |
+
# ### If training, we need to modify the codebook
|
151 |
+
# if (not self.training) & ('keys' not in self.__dict__):
|
152 |
+
# # If it's not training time and we haven't added any keys yet (this is before doing any editing)
|
153 |
+
# # print(self.__dict__)
|
154 |
+
# return layer_out
|
155 |
+
# else:
|
156 |
+
# if not self.training and not self.ensure_replace_token_loc and self.key_id == -1:
|
157 |
+
# token_to_edit = args[0].shape[1]-1
|
158 |
+
# self.key_id = args[0].shape[1]-1
|
159 |
+
# self.ensure_replace_token_loc = True
|
160 |
+
# else:
|
161 |
+
# token_to_edit = min(self.key_id, args[0].shape[1]-1) # args[0].shape[1] - 1 is sequence length
|
162 |
+
# query = args[0][:, token_to_edit, :] # Just use activation for last token
|
163 |
+
# if self.config.val_init == "cold":
|
164 |
+
# new_value = torch.nn.Parameter(torch.rand(1, self.value_shape, requires_grad=True, device=self.device))
|
165 |
+
# elif self.config.val_init == "warm":
|
166 |
+
# new_value = torch.nn.Parameter(layer_out[:, token_to_edit, :].detach(), requires_grad=True)
|
167 |
+
|
168 |
+
# if 'keys' not in self.__dict__:
|
169 |
+
# # If no keys exist, initialize keys, values, epsilons, and key labels
|
170 |
+
# self.keys, self.values, self.epsilons, self.key_labels = self.init_key_value(query, new_value)
|
171 |
+
# elif self.iter == 0:
|
172 |
+
# # Keys exist, so we have decide whether or not to update them (the fact that we've made it to this point means there was an error!)
|
173 |
+
|
174 |
+
# # --- search through keys for a match for query ---
|
175 |
+
# dists = torch.cdist(self.keys, query, p=2).view(-1, len(query))
|
176 |
+
# smallest_distance, nearest_key = dists.min(0)
|
177 |
+
|
178 |
+
# if smallest_distance > (self.init_epsilon + self.epsilons[nearest_key]):
|
179 |
+
# # If there's no close key, make a new key
|
180 |
+
# self.keys, self.values, self.epsilons, self.key_labels = self.add_key(query, new_value)
|
181 |
+
# else:
|
182 |
+
# # If there is a close key, we need to handle conflicts
|
183 |
+
# if not self.label_match(self.edit_label, self.key_labels[nearest_key]):
|
184 |
+
# self.keys, self.values, self.epsilons, self.key_labels = self.add_key(query, new_value)
|
185 |
+
# self.split_epsilons_in_half(nearest_key, smallest_distance)
|
186 |
+
# else:
|
187 |
+
# # If the current label is the SAME as the nearest label, just make the nearest epsilon bigger
|
188 |
+
# if smallest_distance > self.epsilons[nearest_key]:
|
189 |
+
# if self.config.eps_expand== "coverage":
|
190 |
+
# self.epsilons[nearest_key] = smallest_distance # Replace nearest epsilon with dist between old key and new key
|
191 |
+
# elif self.config.eps_expand == "moving_average":
|
192 |
+
# a = 0.5
|
193 |
+
# self.keys[nearest_key] = a*self.keys[nearest_key] + (1-a)*query # Move old key to be halfway between
|
194 |
+
# self.epsilons[nearest_key] = smallest_distance
|
195 |
+
# # self.epsilons[nearest_key] = smallest_distance + self.init_epsilon
|
196 |
+
# else:
|
197 |
+
# # If not iter 0, we don't need to change keys, we just need to learn the value
|
198 |
+
# pass
|
199 |
+
# # print(token_to_edit)
|
200 |
+
# # compute distance from query to all keys and find the closest keys
|
201 |
+
# dists = torch.cdist(self.keys, query, p=2).view(-1, len(query))
|
202 |
+
# smallest_dist, self.chosen_key = dists.min(0)
|
203 |
+
# smallest_dist = smallest_dist.view(-1, 1)
|
204 |
+
# chosen_value = self.values[self.chosen_key]
|
205 |
+
# eps = self.epsilons[self.chosen_key].view(-1, 1)
|
206 |
+
|
207 |
+
# if (self.config.val_train == "adv") and (self.training):
|
208 |
+
# chosen_value = perturb_values(chosen_value, self.num_pert, self.device)
|
209 |
+
|
210 |
+
# if self.replacement == "replace_all":
|
211 |
+
# layer_out = torch.where((smallest_dist <= eps).view(-1, 1, 1), chosen_value.unsqueeze(1).repeat_interleave(layer_out.shape[1], 1), layer_out)
|
212 |
+
# elif self.replacement == "replace_last":
|
213 |
+
# layer_out[:, token_to_edit] = torch.where((smallest_dist <= eps), chosen_value, layer_out[:, token_to_edit])
|
214 |
+
# elif self.replacement == "replace_prompt":
|
215 |
+
# layer_out[:, :token_to_edit] = torch.where((smallest_dist <= eps), chosen_value, layer_out[:, :token_to_edit])
|
216 |
+
# else:
|
217 |
+
# print("token replacement choice not found")
|
218 |
+
# return layer_out
|
219 |
+
import copy
|
220 |
+
|
221 |
+
import torch
|
222 |
+
from .utils import parent_module, brackets_to_periods
|
223 |
+
import transformers
|
224 |
+
import os
|
225 |
+
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
|
226 |
+
|
227 |
+
def euc(query, key):
|
228 |
+
# Euclidean distance
|
229 |
+
if len(key.shape) < 2:
|
230 |
+
key = key.view(1, -1)
|
231 |
+
return torch.cdist(key, query, p=2)
|
232 |
+
|
233 |
+
def perturb_values(chosen_value, num_pert, device):
|
234 |
+
# Create a bunch of noised versions of the value, then create batch, then train value
|
235 |
+
chosen_value = chosen_value
|
236 |
+
noise = torch.normal(0, 1, chosen_value.shape, device=device)
|
237 |
+
noise[0] = noise[0]*0
|
238 |
+
noise.requires_grad = True
|
239 |
+
chosen_value = chosen_value + noise
|
240 |
+
return chosen_value
|
241 |
+
|
242 |
+
class GRACE(torch.nn.Module):
|
243 |
+
def __init__(self, config, model, device):
|
244 |
+
super(GRACE, self).__init__()
|
245 |
+
self.config = config
|
246 |
+
self.log_dict = {}
|
247 |
+
self.model = model
|
248 |
+
self.config = config
|
249 |
+
# self.tokenizer = model.tokenizer
|
250 |
+
layer = config.inner_params[0]
|
251 |
+
self.device = device
|
252 |
+
self.original_layer = None
|
253 |
+
|
254 |
+
# --- ensure proper formatting (GRACE edits ~layers~ not weights matrices) ---
|
255 |
+
suffixes = [".weight", ".bias"]
|
256 |
+
self.layer = layer.rsplit(".", 1)[0] if any(layer.endswith(x) for x in suffixes) else layer
|
257 |
+
|
258 |
+
for n, p in self.model.named_parameters():
|
259 |
+
p.requires_grad = False
|
260 |
+
|
261 |
+
if isinstance(self.model, transformers.models.gpt2.modeling_gpt2.GPT2LMHeadModel):
|
262 |
+
transpose = False
|
263 |
+
else:
|
264 |
+
transpose = True
|
265 |
+
|
266 |
+
# --- Add GRACE to chosen layers ---
|
267 |
+
edit_module = parent_module(self.model, brackets_to_periods(self.layer))
|
268 |
+
layer_name = self.layer.rsplit(".", 1)[-1]
|
269 |
+
original_layer = getattr(edit_module, layer_name)
|
270 |
+
if type(original_layer) is not GRACEAdapter:
|
271 |
+
setattr(edit_module, layer_name, GRACEAdapter(config, original_layer, transpose=transpose).to(self.device))
|
272 |
+
self.original_layer = copy.deepcopy(original_layer)
|
273 |
+
|
274 |
+
def __call__(self, **kwargs):
|
275 |
+
# if self.config.task == "hallucination":
|
276 |
+
# print(kwargs)
|
277 |
+
# key_id = (kwargs["labels"] == -100).sum() - 1
|
278 |
+
# setattr(eval(f"self.model.{self.layer}"), "key_id", key_id) # Tell GRACE which token to use for its query (default is the last token)
|
279 |
+
return self.model(**kwargs)
|
280 |
+
|
281 |
+
def reset_layer(self):
|
282 |
+
layer_name = self.layer.rsplit(".", 1)[-1]
|
283 |
+
edit_module = parent_module(self.model, brackets_to_periods(self.layer))
|
284 |
+
setattr(edit_module, layer_name, self.original_layer.to(self.device))
|
285 |
+
|
286 |
+
def generate(self, *args, **kwargs):
|
287 |
+
setattr(eval(f"self.model.{self.layer}"), "key_id", -1)
|
288 |
+
return self.model.generate(*args, **kwargs)
|
289 |
+
|
290 |
+
def edit(self, config, tokens):
|
291 |
+
key_id = (tokens["labels"] == -100).sum() - 1
|
292 |
+
setattr(eval(f"self.model.{self.layer}"), "key_id", key_id)
|
293 |
+
|
294 |
+
# --- pass edit label, training mode, and key_id into GRACE ---
|
295 |
+
setattr(eval(f"self.model.{self.layer}"), "training", True)
|
296 |
+
setattr(eval(f"self.model.{self.layer}"), "edit_label", tokens["labels"])
|
297 |
+
|
298 |
+
self.losses = []
|
299 |
+
# --- train GRACE value ---
|
300 |
+
for i in range(config.n_iter):
|
301 |
+
# --- insert iteration into each layer (only initiate keys on iteration 1) ---
|
302 |
+
setattr(eval(f"self.model.{self.layer}"), "iter", i)
|
303 |
+
|
304 |
+
# --- pass tokens through model (including through the GRACE layer) ---
|
305 |
+
outputs = self.model(**tokens)
|
306 |
+
if i == 0:
|
307 |
+
# --- we only need to create an optimizer for the first iteration (but forward pass instantiates the key, so optimzer is passed after first inference) ---
|
308 |
+
optimizer = torch.optim.Adam(self.model.parameters(), config.edit_lr)
|
309 |
+
loss = outputs.loss
|
310 |
+
loss.backward()
|
311 |
+
optimizer.step()
|
312 |
+
optimizer.zero_grad()
|
313 |
+
self.losses.append(loss.detach().cpu().numpy())
|
314 |
+
|
315 |
+
self.loss = loss # Log final loss
|
316 |
+
|
317 |
+
# --- pull out info we want to log from the GRACE layer ---
|
318 |
+
setattr(eval(f"self.model.{self.layer}"), "training", False)
|
319 |
+
chosen_key = getattr(eval(f"self.model.{self.layer}"), "chosen_key")
|
320 |
+
nkeys = len(getattr(eval(f"self.model.{self.layer}"), "keys"))
|
321 |
+
|
322 |
+
self.log_dict["chosen_key"] = chosen_key
|
323 |
+
self.log_dict["nkeys"] = nkeys
|
324 |
+
|
325 |
+
class GRACEAdapter(torch.nn.Module):
|
326 |
+
def __init__(self, config, layer, transpose):
|
327 |
+
super(GRACEAdapter, self).__init__()
|
328 |
+
|
329 |
+
self.layer = layer
|
330 |
+
self.weight = self.layer.weight
|
331 |
+
self.init_epsilon = config.eps
|
332 |
+
self.dist_fn = config.dist_fn
|
333 |
+
self.replacement = config.replacement
|
334 |
+
self.device = layer.weight.device
|
335 |
+
self.config = config
|
336 |
+
self.num_pert = config.num_pert
|
337 |
+
self.key_id = -1
|
338 |
+
self.ensure_replace_token_loc = False
|
339 |
+
|
340 |
+
if transpose:
|
341 |
+
self.key_shape = layer.weight.shape[1]
|
342 |
+
self.value_shape = layer.weight.shape[0]
|
343 |
+
else:
|
344 |
+
self.key_shape = layer.weight.shape[0]
|
345 |
+
self.value_shape = layer.weight.shape[1]
|
346 |
+
self.training = False
|
347 |
+
|
348 |
+
def add_key(self, new_key, new_value):
|
349 |
+
keys = torch.vstack([self.keys, new_key.detach()]) # Add new key to list of keys
|
350 |
+
|
351 |
+
values = torch.nn.Parameter(torch.vstack([self.values, new_value]), requires_grad=True) # Add new value to list of values
|
352 |
+
|
353 |
+
new_epsilon = torch.tensor(self.init_epsilon, device=self.device).view(1)
|
354 |
+
epsilons = torch.vstack([self.epsilons, new_epsilon]) # Add new epsilon to list of epsilons
|
355 |
+
|
356 |
+
key_labels = self.key_labels + [self.edit_label] # Add new key_label to list of key_labels
|
357 |
+
|
358 |
+
return keys, values, epsilons, key_labels
|
359 |
+
|
360 |
+
def init_key_value(self, query, value):
|
361 |
+
key = query.detach()
|
362 |
+
epsilon = torch.tensor(self.init_epsilon, device=self.device, requires_grad=False).view(1)
|
363 |
+
key_label = [self.edit_label]
|
364 |
+
return key, value, epsilon, key_label
|
365 |
+
|
366 |
+
def label_match(self, edit_label, key_label):
|
367 |
+
return edit_label.float().mean() == key_label.float().mean()
|
368 |
+
|
369 |
+
def split_epsilons_in_half(self, nearest_key, smallest_distance):
|
370 |
+
self.epsilons[nearest_key] = (smallest_distance / 2) - 1e-5 # Cut nearest epsilon in half
|
371 |
+
self.epsilons[-1] = smallest_distance / 2 # Cut new epsilon in half
|
372 |
+
|
373 |
+
def forward(self, *args):
|
374 |
+
# Run layer forward and save what it would have returned for this instance
|
375 |
+
layer_out = self.layer(*args)
|
376 |
+
|
377 |
+
### If training, we need to modify the codebook
|
378 |
+
if (not self.training) & ('keys' not in self.__dict__):
|
379 |
+
# If it's not training time and we haven't added any keys yet (this is before doing any editing)
|
380 |
+
# print(self.__dict__)
|
381 |
+
return layer_out
|
382 |
+
else:
|
383 |
+
if not self.training and not self.ensure_replace_token_loc and self.key_id == -1:
|
384 |
+
token_to_edit = args[0].shape[1]-1
|
385 |
+
self.key_id = args[0].shape[1]-1
|
386 |
+
self.ensure_replace_token_loc = True
|
387 |
+
else:
|
388 |
+
token_to_edit = min(self.key_id, args[0].shape[1]-1) # args[0].shape[1] - 1 is sequence length
|
389 |
+
query = args[0][:, token_to_edit, :] # Just use activation for last token
|
390 |
+
if self.config.val_init == "cold":
|
391 |
+
new_value = torch.nn.Parameter(torch.rand(1, self.value_shape, requires_grad=True, device=self.device))
|
392 |
+
elif self.config.val_init == "warm":
|
393 |
+
new_value = torch.nn.Parameter(layer_out[:, token_to_edit, :].detach(), requires_grad=True)
|
394 |
+
|
395 |
+
if 'keys' not in self.__dict__:
|
396 |
+
# If no keys exist, initialize keys, values, epsilons, and key labels
|
397 |
+
self.keys, self.values, self.epsilons, self.key_labels = self.init_key_value(query, new_value)
|
398 |
+
elif self.iter == 0:
|
399 |
+
# Keys exist, so we have decide whether or not to update them (the fact that we've made it to this point means there was an error!)
|
400 |
+
|
401 |
+
# --- search through keys for a match for query ---
|
402 |
+
dists = torch.cdist(self.keys, query, p=2).view(-1, len(query))
|
403 |
+
smallest_distance, nearest_key = dists.min(0)
|
404 |
+
|
405 |
+
if smallest_distance > (self.init_epsilon + self.epsilons[nearest_key]):
|
406 |
+
# If there's no close key, make a new key
|
407 |
+
self.keys, self.values, self.epsilons, self.key_labels = self.add_key(query, new_value)
|
408 |
+
else:
|
409 |
+
# If there is a close key, we need to handle conflicts
|
410 |
+
if not self.label_match(self.edit_label, self.key_labels[nearest_key]):
|
411 |
+
self.keys, self.values, self.epsilons, self.key_labels = self.add_key(query, new_value)
|
412 |
+
self.split_epsilons_in_half(nearest_key, smallest_distance)
|
413 |
+
else:
|
414 |
+
# If the current label is the SAME as the nearest label, just make the nearest epsilon bigger
|
415 |
+
if smallest_distance > self.epsilons[nearest_key]:
|
416 |
+
if self.config.eps_expand== "coverage":
|
417 |
+
self.epsilons[nearest_key] = smallest_distance # Replace nearest epsilon with dist between old key and new key
|
418 |
+
elif self.config.eps_expand == "moving_average":
|
419 |
+
a = 0.5
|
420 |
+
self.keys[nearest_key] = a*self.keys[nearest_key] + (1-a)*query # Move old key to be halfway between
|
421 |
+
self.epsilons[nearest_key] = smallest_distance
|
422 |
+
# self.epsilons[nearest_key] = smallest_distance + self.init_epsilon
|
423 |
+
else:
|
424 |
+
# If not iter 0, we don't need to change keys, we just need to learn the value
|
425 |
+
pass
|
426 |
+
# print(token_to_edit)
|
427 |
+
# compute distance from query to all keys and find the closest keys
|
428 |
+
dists = torch.cdist(self.keys, query, p=2).view(-1, len(query))
|
429 |
+
smallest_dist, self.chosen_key = dists.min(0)
|
430 |
+
smallest_dist = smallest_dist.view(-1, 1)
|
431 |
+
chosen_value = self.values[self.chosen_key]
|
432 |
+
eps = self.epsilons[self.chosen_key].view(-1, 1)
|
433 |
+
|
434 |
+
if (self.config.val_train == "adv") and (self.training):
|
435 |
+
chosen_value = perturb_values(chosen_value, self.num_pert, self.device)
|
436 |
+
|
437 |
+
if self.replacement == "replace_all":
|
438 |
+
layer_out = torch.where((smallest_dist <= eps).view(-1, 1, 1), chosen_value.unsqueeze(1).repeat_interleave(layer_out.shape[1], 1), layer_out)
|
439 |
+
elif self.replacement == "replace_last":
|
440 |
+
layer_out[:, token_to_edit] = torch.where((smallest_dist <= eps), chosen_value, layer_out[:, token_to_edit])
|
441 |
+
elif self.replacement == "replace_prompt":
|
442 |
+
layer_out[:, :token_to_edit] = torch.where((smallest_dist <= eps), chosen_value, layer_out[:, :token_to_edit])
|
443 |
+
else:
|
444 |
+
print("token replacement choice not found")
|
445 |
+
return layer_out
|
easyeditor/models/grace/__init__.py
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
from .grace_main import GraceHyperParams, apply_grace_to_model
|
2 |
+
from .metrics import F1, PPL, Accuracy, is_qa_error, is_acc_error
|
easyeditor/models/grace/__pycache__/GRACE.cpython-39.pyc
ADDED
Binary file (6.67 kB). View file
|
|
easyeditor/models/grace/__pycache__/__init__.cpython-39.pyc
ADDED
Binary file (350 Bytes). View file
|
|
easyeditor/models/grace/__pycache__/grace_hparams.cpython-39.pyc
ADDED
Binary file (1.5 kB). View file
|
|
easyeditor/models/grace/__pycache__/grace_main.cpython-39.pyc
ADDED
Binary file (1.23 kB). View file
|
|
easyeditor/models/grace/__pycache__/metrics.cpython-39.pyc
ADDED
Binary file (2.07 kB). View file
|
|
easyeditor/models/grace/__pycache__/utils.cpython-39.pyc
ADDED
Binary file (3.54 kB). View file
|
|
easyeditor/models/grace/grace_hparams.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
from typing import List
|
3 |
+
from ...util.hparams import HyperParams
|
4 |
+
import yaml
|
5 |
+
|
6 |
+
|
7 |
+
@dataclass
|
8 |
+
class GraceHyperParams(HyperParams):
|
9 |
+
# Experiments
|
10 |
+
|
11 |
+
edit_lr: int
|
12 |
+
n_iter: int
|
13 |
+
# Method
|
14 |
+
eps: float
|
15 |
+
dist_fn: str
|
16 |
+
val_init: str
|
17 |
+
val_train: str
|
18 |
+
val_reg: str
|
19 |
+
reg: str
|
20 |
+
replacement: str
|
21 |
+
eps_expand: str
|
22 |
+
num_pert: str
|
23 |
+
dropout: float
|
24 |
+
|
25 |
+
# Module templates
|
26 |
+
inner_params: List[str]
|
27 |
+
device: int
|
28 |
+
alg_name: str
|
29 |
+
model_name: str
|
30 |
+
|
31 |
+
# Defaults
|
32 |
+
batch_size: int = 128
|
33 |
+
max_length: int = 30
|
34 |
+
model_parallel: bool = False
|
35 |
+
|
36 |
+
@classmethod
|
37 |
+
def from_hparams(cls, hparams_name_or_path: str):
|
38 |
+
if '.yaml' not in hparams_name_or_path:
|
39 |
+
hparams_name_or_path = hparams_name_or_path + '.yaml'
|
40 |
+
|
41 |
+
with open(hparams_name_or_path, "r") as stream:
|
42 |
+
config = yaml.safe_load(stream)
|
43 |
+
config = super().construct_float_from_scientific_notation(config)
|
44 |
+
|
45 |
+
assert (config and config['alg_name'] == 'GRACE') or print(
|
46 |
+
f'GraceHyperParams can not load from {hparams_name_or_path}, '
|
47 |
+
f'alg_name is {config["alg_name"]} ')
|
48 |
+
return cls(**config)
|
easyeditor/models/grace/grace_main.py
ADDED
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict, List, Tuple
|
2 |
+
import torch
|
3 |
+
from copy import deepcopy
|
4 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
5 |
+
from .GRACE import GRACE
|
6 |
+
from .grace_hparams import GraceHyperParams
|
7 |
+
from .utils import tokenize
|
8 |
+
from ...util import nethook
|
9 |
+
import gradio as gr
|
10 |
+
|
11 |
+
|
12 |
+
def apply_grace_to_model(
|
13 |
+
model: AutoModelForCausalLM,
|
14 |
+
tok: AutoTokenizer,
|
15 |
+
requests: List[Dict],
|
16 |
+
hparams: GraceHyperParams,
|
17 |
+
num_steps: int,
|
18 |
+
replacement: str,
|
19 |
+
copy=False,
|
20 |
+
return_orig_weights=False,
|
21 |
+
keep_original_weight=False,
|
22 |
+
**kwargs: Any,
|
23 |
+
) -> Tuple[AutoModelForCausalLM, Dict[str, Any]]:
|
24 |
+
request = requests
|
25 |
+
if copy:
|
26 |
+
model = deepcopy(model)
|
27 |
+
weights_copy = {}
|
28 |
+
device = torch.device('cpu')
|
29 |
+
hparams.n_iter = num_steps
|
30 |
+
hparams.replacement = replacement
|
31 |
+
editor = GRACE(model=model, config=hparams, device=device)
|
32 |
+
|
33 |
+
tokens = tokenize(request, tokenizer=tok, device=device)
|
34 |
+
editor.edit(config=hparams, tokens=tokens)
|
35 |
+
|
36 |
+
editor.to('cpu')
|
37 |
+
gr.Info("Completed editing via GRACE!")
|
38 |
+
return editor
|
39 |
+
|
40 |
+
|
easyeditor/models/grace/metrics.py
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
from .utils import *
|
4 |
+
|
5 |
+
def is_acc_error(model, tokens):
|
6 |
+
# Check whether or not the model's prediction for a batch element is correct
|
7 |
+
labels = tokens["labels"]
|
8 |
+
logits = model(**tokens).logits
|
9 |
+
probs = torch.softmax(logits, -1).squeeze()
|
10 |
+
argmaxs = torch.argmax(probs, dim=-1).squeeze()
|
11 |
+
return labels != argmaxs
|
12 |
+
|
13 |
+
def Accuracy(model, tokens):
|
14 |
+
labels = tokens["labels"]
|
15 |
+
new_tokens = {f"{k}" : v for k, v in tokens.items() if k != "labels"}
|
16 |
+
logits = model(**new_tokens).logits
|
17 |
+
probs = torch.softmax(logits, -1).squeeze()
|
18 |
+
argmaxs = torch.argmax(probs, dim=-1).squeeze()
|
19 |
+
return (labels == argmaxs).float().mean()
|
20 |
+
|
21 |
+
def is_qa_error(model, tokens):
|
22 |
+
preds = model.generate(tokens["input_ids"], max_length=20).squeeze() # Run model to get its predictions
|
23 |
+
labels = tokens["labels"]#[tokens["labels"] != -100]
|
24 |
+
|
25 |
+
if (len(preds) != len(labels)) or ((preds == labels).sum() != len(preds)):
|
26 |
+
return True
|
27 |
+
else:
|
28 |
+
return False
|
29 |
+
|
30 |
+
def PPL(model, batch):
|
31 |
+
input_ids = batch["input_ids"][:, :1024]#.to(device)
|
32 |
+
if "labels" not in batch:
|
33 |
+
target_ids = batch["input_ids"][:, :1024].clone()
|
34 |
+
else:
|
35 |
+
target_ids = batch["labels"][:, :1024].clone()
|
36 |
+
|
37 |
+
with torch.no_grad():
|
38 |
+
outputs = model(input_ids=input_ids, labels=target_ids)
|
39 |
+
nll = outputs.loss
|
40 |
+
|
41 |
+
ppl = torch.exp(nll)#.clip(0, 100)
|
42 |
+
return ppl
|
43 |
+
|
44 |
+
def F1(model, batch):
|
45 |
+
try:
|
46 |
+
preds = model.generate(batch["input_ids"], max_length=20).squeeze()
|
47 |
+
if len(preds) > 1:
|
48 |
+
preds = preds[preds != model.tokenizer.pad_token_id]
|
49 |
+
gold_toks = batch["labels"][batch["labels"] != -100].cpu().squeeze() # -100 might be nonsense
|
50 |
+
num_same = len(np.intersect1d(preds.cpu().squeeze(), gold_toks))
|
51 |
+
if (num_same == 0) or (len(preds.squeeze()) == 0):
|
52 |
+
return 0
|
53 |
+
precision = num_same / len(preds.squeeze())
|
54 |
+
recall = 1.0 * num_same / len(gold_toks)
|
55 |
+
f1 = (2 * precision * recall) / (precision + recall)
|
56 |
+
return f1
|
57 |
+
except:
|
58 |
+
# Every once in a while, the model just returns the stop token
|
59 |
+
return 0
|
easyeditor/models/grace/utils.py
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import transformers
|
2 |
+
import torch
|
3 |
+
import os
|
4 |
+
import numpy as np
|
5 |
+
import datetime
|
6 |
+
import struct
|
7 |
+
from torch.nn.utils.rnn import pad_sequence
|
8 |
+
import torch.nn.functional as F
|
9 |
+
|
10 |
+
def get_inner_params(named_parameters, inner_names):
|
11 |
+
param_dict = dict(named_parameters)
|
12 |
+
return [(n, param_dict[n]) for n in inner_names]
|
13 |
+
|
14 |
+
def param_subset(named_parameters, inner_names):
|
15 |
+
param_dict = dict(named_parameters)
|
16 |
+
return [param_dict[n] for n in inner_names]
|
17 |
+
|
18 |
+
def parent_module(model, pname):
|
19 |
+
components = pname.split('.')
|
20 |
+
parent = model
|
21 |
+
|
22 |
+
for component in components[:-1]:
|
23 |
+
if hasattr(parent, component):
|
24 |
+
parent = getattr(parent, component)
|
25 |
+
elif component.isdigit():
|
26 |
+
parent = parent[int(component)]
|
27 |
+
else:
|
28 |
+
raise RuntimeError(f"Couldn't find child module {component}")
|
29 |
+
|
30 |
+
if not hasattr(parent, components[-1]):
|
31 |
+
raise RuntimeError(f"Couldn't find child module {components[-1]}")
|
32 |
+
|
33 |
+
return parent
|
34 |
+
|
35 |
+
def uuid(digits=4):
|
36 |
+
if not hasattr(uuid, "uuid_value"):
|
37 |
+
uuid.uuid_value = struct.unpack('I', os.urandom(4))[0] % int(10**digits)
|
38 |
+
|
39 |
+
return uuid.uuid_value
|
40 |
+
|
41 |
+
def ckpt_dir():
|
42 |
+
"""returns the directory in which to store model checkpoints"""
|
43 |
+
path = "./ckpts/"
|
44 |
+
if not os.path.exists(path):
|
45 |
+
os.makedirs(path)
|
46 |
+
return path
|
47 |
+
|
48 |
+
def brackets_to_periods(name):
|
49 |
+
return name.replace("[", ".").replace("]", "")
|
50 |
+
|
51 |
+
def get_params(model):
|
52 |
+
return model.state_dict()
|
53 |
+
|
54 |
+
def get_shape(p, model):
|
55 |
+
# We need to flip the shapes since OpenAI gpt2 uses convs instead of linear
|
56 |
+
return p.shape if isinstance(model, transformers.GPT2LMHeadModel) else (p.shape[1], p.shape[0])
|
57 |
+
|
58 |
+
def get_logits(x):
|
59 |
+
return x.logits if hasattr(x, "logits") else x
|
60 |
+
|
61 |
+
def tokenize(batch, tokenizer, device, test=False):
|
62 |
+
prompt, label = batch["prompt"], batch["target_new"]
|
63 |
+
if not isinstance(prompt, list):
|
64 |
+
prompt=[prompt]
|
65 |
+
if not isinstance(label, list):
|
66 |
+
label=[label]
|
67 |
+
mask_token = -100 # ignore_index of CrossEntropyLoss
|
68 |
+
if test or not label:
|
69 |
+
tokens = tokenizer(list(prompt), return_tensors="pt", padding=True, truncation=True)
|
70 |
+
tokens["labels"] = tokens["input_ids"].clone()
|
71 |
+
tokens["labels"][tokens["input_ids"] == tokenizer.pad_token_id] = mask_token
|
72 |
+
|
73 |
+
else:
|
74 |
+
full_prompt = [f"{p} {l}" for p, l in zip(prompt, label)]
|
75 |
+
prompt_ids = tokenizer(list(prompt), return_tensors="pt", padding=True, truncation=True)["input_ids"]
|
76 |
+
num_prompt_toks = [int((i != tokenizer.pad_token_id).sum()) for i in prompt_ids]
|
77 |
+
tokens = tokenizer(full_prompt, return_tensors="pt", padding=True, truncation=True)
|
78 |
+
tokens["labels"] = tokens["input_ids"].clone()
|
79 |
+
for i in range(len(prompt)):
|
80 |
+
tokens["labels"][i][:num_prompt_toks[i]] = mask_token
|
81 |
+
|
82 |
+
tokens["labels"][tokens["input_ids"] == tokenizer.pad_token_id] = mask_token
|
83 |
+
|
84 |
+
tokens = {f"{k1}" : v1.to(device) for k1, v1 in tokens.items()}
|
85 |
+
return tokens
|
86 |
+
|
easyeditor/util/__init__.py
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
from .logit_lens import LogitLens
|
2 |
+
from .hparams import *
|
easyeditor/util/__pycache__/__init__.cpython-39.pyc
ADDED
Binary file (216 Bytes). View file
|
|
easyeditor/util/__pycache__/hparams.cpython-39.pyc
ADDED
Binary file (1.21 kB). View file
|
|
easyeditor/util/__pycache__/logit_lens.cpython-39.pyc
ADDED
Binary file (3.36 kB). View file
|
|
easyeditor/util/__pycache__/nethook.cpython-39.pyc
ADDED
Binary file (13.2 kB). View file
|
|
easyeditor/util/alg_dict.py
ADDED
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from ..models.rome import ROMEHyperParams, apply_rome_to_model
|
2 |
+
from ..models.memit import MEMITHyperParams, apply_memit_to_model
|
3 |
+
from ..models.kn import KNHyperParams, apply_kn_to_model
|
4 |
+
from ..models.mend import MENDHyperParams, MendRewriteExecutor, MendMultimodalRewriteExecutor
|
5 |
+
from ..models.ft import FTHyperParams, apply_ft_to_model
|
6 |
+
from ..models.serac import SERACHparams, SeracRewriteExecutor, SeracMultimodalRewriteExecutor
|
7 |
+
from ..dataset import ZsreDataset, CounterFactDataset, CaptionDataset, VQADataset
|
8 |
+
from ..models.ike import IKEHyperParams, apply_ike_to_model, apply_ike_to_multimodal_model
|
9 |
+
from ..models.ft_api import FTApiHyperParams, apply_ft_api_to_model
|
10 |
+
from ..models.lora import LoRAHyperParams, apply_lora_to_model
|
11 |
+
from ..models.grace import GraceHyperParams, apply_grace_to_model
|
12 |
+
from ..models.pmet import PMETHyperParams, apply_pmet_to_model
|
13 |
+
from ..models.melo import MELOHyperParams, apply_melo_to_model
|
14 |
+
|
15 |
+
ALG_DICT = {
|
16 |
+
'ROME': apply_rome_to_model,
|
17 |
+
'MEMIT': apply_memit_to_model,
|
18 |
+
"FT": apply_ft_to_model,
|
19 |
+
'KN': apply_kn_to_model,
|
20 |
+
'MEND': MendRewriteExecutor().apply_to_model,
|
21 |
+
'SERAC': SeracRewriteExecutor().apply_to_model,
|
22 |
+
'IKE': apply_ike_to_model,
|
23 |
+
'FT-Api': apply_ft_api_to_model,
|
24 |
+
'LoRA': apply_lora_to_model,
|
25 |
+
'GRACE': apply_grace_to_model,
|
26 |
+
'PMET': apply_pmet_to_model,
|
27 |
+
'MELO': apply_melo_to_model
|
28 |
+
}
|
29 |
+
|
30 |
+
ALG_MULTIMODAL_DICT = {
|
31 |
+
'MEND': MendMultimodalRewriteExecutor().apply_to_model,
|
32 |
+
'SERAC': SeracMultimodalRewriteExecutor().apply_to_model,
|
33 |
+
'SERAC_MULTI': SeracMultimodalRewriteExecutor().apply_to_model,
|
34 |
+
'IKE': apply_ike_to_multimodal_model,
|
35 |
+
}
|
36 |
+
|
37 |
+
DS_DICT = {
|
38 |
+
"cf": CounterFactDataset,
|
39 |
+
"zsre": ZsreDataset,
|
40 |
+
}
|
41 |
+
|
42 |
+
MULTIMODAL_DS_DICT = {
|
43 |
+
"caption": CaptionDataset,
|
44 |
+
"vqa": VQADataset,
|
45 |
+
}
|
easyeditor/util/alg_train_dict.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from ..trainer import MEND
|
2 |
+
from ..trainer import SERAC, SERAC_MULTI
|
3 |
+
|
4 |
+
|
5 |
+
ALG_TRAIN_DICT = {
|
6 |
+
'MEND': MEND,
|
7 |
+
'SERAC': SERAC,
|
8 |
+
'SERAC_MULTI': SERAC_MULTI,
|
9 |
+
}
|
easyeditor/util/generate.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import unicodedata
|
2 |
+
from typing import List, Optional
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
6 |
+
|
7 |
+
from .logit_lens import LogitLens
|
8 |
+
|
9 |
+
|
10 |
+
def generate_interactive(
|
11 |
+
model: AutoModelForCausalLM,
|
12 |
+
tok: AutoTokenizer,
|
13 |
+
top_k: int = 5,
|
14 |
+
max_out_len: int = 200,
|
15 |
+
compare_against: Optional[AutoModelForCausalLM] = None,
|
16 |
+
use_logit_lens: bool = False,
|
17 |
+
layer_module_tmp: str = "transformer.h.{}",
|
18 |
+
ln_f_module: str = "transformer.ln_f",
|
19 |
+
lm_head_module: str = "lm_head",
|
20 |
+
):
|
21 |
+
"""
|
22 |
+
Puts generation in a loop. Allows users to repeatedly provide inputs
|
23 |
+
with which text is generated.
|
24 |
+
"""
|
25 |
+
|
26 |
+
if use_logit_lens:
|
27 |
+
llens_gen = LogitLens(
|
28 |
+
model,
|
29 |
+
tok,
|
30 |
+
layer_module_tmp,
|
31 |
+
ln_f_module,
|
32 |
+
lm_head_module,
|
33 |
+
disabled=not use_logit_lens,
|
34 |
+
)
|
35 |
+
if compare_against:
|
36 |
+
llens_vanilla = LogitLens(
|
37 |
+
compare_against,
|
38 |
+
tok,
|
39 |
+
layer_module_tmp,
|
40 |
+
ln_f_module,
|
41 |
+
lm_head_module,
|
42 |
+
disabled=not use_logit_lens,
|
43 |
+
)
|
44 |
+
|
45 |
+
while True:
|
46 |
+
prompt = input("Enter a prompt: ").strip(" \r\t\n")
|
47 |
+
|
48 |
+
print(
|
49 |
+
f"Argument Model: "
|
50 |
+
f"{generate_fast(model, tok, [prompt], n_gen_per_prompt=1, top_k=top_k, max_out_len=max_out_len)}"
|
51 |
+
)
|
52 |
+
if compare_against:
|
53 |
+
print(
|
54 |
+
f"Baseline Model: "
|
55 |
+
f"{generate_fast(compare_against, tok, [prompt], n_gen_per_prompt=1, top_k=top_k, max_out_len=max_out_len)}"
|
56 |
+
)
|
57 |
+
|
58 |
+
if use_logit_lens:
|
59 |
+
inp_prompt = tok([prompt], padding=True, return_tensors="pt").to(
|
60 |
+
next(model.parameters()).device
|
61 |
+
)
|
62 |
+
|
63 |
+
with llens_gen:
|
64 |
+
model(**inp_prompt)
|
65 |
+
print("\n--- Argument Model Logit Lens ---")
|
66 |
+
llens_gen.pprint()
|
67 |
+
|
68 |
+
if compare_against:
|
69 |
+
with llens_vanilla:
|
70 |
+
compare_against(**inp_prompt)
|
71 |
+
print("--- Baseline Model Logit Lens ---")
|
72 |
+
llens_vanilla.pprint()
|
73 |
+
|
74 |
+
print()
|
75 |
+
|
76 |
+
|
77 |
+
def generate_fast(
|
78 |
+
model: AutoModelForCausalLM,
|
79 |
+
tok: AutoTokenizer,
|
80 |
+
prompts: List[str],
|
81 |
+
n_gen_per_prompt: int = 1,
|
82 |
+
top_k: int = 5,
|
83 |
+
max_out_len: int = 200,
|
84 |
+
vanilla_generation=False,
|
85 |
+
):
|
86 |
+
"""
|
87 |
+
Fast, parallelized auto-regressive text generation with top-k sampling.
|
88 |
+
Our custom implementation.
|
89 |
+
"""
|
90 |
+
|
91 |
+
# Unroll prompts and tokenize
|
92 |
+
inp = [prompt for prompt in prompts for _ in range(n_gen_per_prompt)]
|
93 |
+
inp_tok = tok(inp, padding=True, return_tensors="pt").to(
|
94 |
+
next(model.parameters()).device
|
95 |
+
)
|
96 |
+
input_ids, attention_mask = inp_tok["input_ids"], inp_tok["attention_mask"]
|
97 |
+
if vanilla_generation:
|
98 |
+
gen_txt = model.generate(
|
99 |
+
input_ids=input_ids,
|
100 |
+
attention_mask=attention_mask,
|
101 |
+
max_new_tokens=max_out_len
|
102 |
+
)
|
103 |
+
txt = [tok.decode(x, skip_special_tokens=True) for x in gen_txt.detach().cpu().numpy().tolist()]
|
104 |
+
txt = [
|
105 |
+
unicodedata.normalize("NFKD", x)
|
106 |
+
.replace("\n\n", " ")
|
107 |
+
.replace("<|endoftext|>", "")
|
108 |
+
for x in txt
|
109 |
+
]
|
110 |
+
return txt
|
111 |
+
batch_size = input_ids.size(0)
|
112 |
+
|
113 |
+
# Setup storage of fast generation with attention caches.
|
114 |
+
# `cur_context` is used to define the range of inputs that are not yet
|
115 |
+
# stored in `past_key_values`. At each step, we are generating the
|
116 |
+
# next token for the index at `cur_context.stop + 1`.
|
117 |
+
past_key_values, cur_context = None, slice(0, attention_mask.sum(1).min().item())
|
118 |
+
|
119 |
+
with torch.no_grad():
|
120 |
+
while input_ids.size(1) < max_out_len: # while not exceeding max output length
|
121 |
+
model_out = model(
|
122 |
+
input_ids=input_ids[:, cur_context],
|
123 |
+
attention_mask=None if 'llama'or'baichuan' in model.name_or_path.lower() else attention_mask[:, cur_context],
|
124 |
+
past_key_values=past_key_values,
|
125 |
+
use_cache=True,
|
126 |
+
)
|
127 |
+
logits, past_key_values = model_out.logits, model_out.past_key_values
|
128 |
+
softmax_out = torch.nn.functional.softmax(logits[:, -1, :], dim=1)
|
129 |
+
|
130 |
+
# Top-k sampling
|
131 |
+
tk = torch.topk(softmax_out, top_k, dim=1).indices
|
132 |
+
softmax_out_top_k = torch.gather(softmax_out, 1, tk)
|
133 |
+
softmax_out_top_k = softmax_out_top_k / softmax_out_top_k.sum(1)[:, None]
|
134 |
+
new_tok_indices = torch.multinomial(softmax_out_top_k, 1)
|
135 |
+
new_toks = torch.gather(tk, 1, new_tok_indices)
|
136 |
+
|
137 |
+
# If we're currently generating the continuation for the last token in `input_ids`,
|
138 |
+
# create a new index so we can insert the new token
|
139 |
+
if cur_context.stop == input_ids.size(1):
|
140 |
+
attention_mask = torch.cat(
|
141 |
+
[attention_mask, attention_mask.new_zeros(batch_size, 1)], dim=1
|
142 |
+
)
|
143 |
+
input_ids = torch.cat(
|
144 |
+
[
|
145 |
+
input_ids,
|
146 |
+
input_ids.new_ones(batch_size, 1) * tok.pad_token_id,
|
147 |
+
],
|
148 |
+
dim=1,
|
149 |
+
)
|
150 |
+
|
151 |
+
last_non_masked = attention_mask.sum(1) - 1
|
152 |
+
for i in range(batch_size):
|
153 |
+
new_idx = last_non_masked[i] + 1
|
154 |
+
if last_non_masked[i].item() + 1 != cur_context.stop:
|
155 |
+
continue
|
156 |
+
|
157 |
+
# Stop generating if we've already maxed out for this prompt
|
158 |
+
if new_idx < max_out_len:
|
159 |
+
input_ids[i][new_idx] = new_toks[i]
|
160 |
+
attention_mask[i][new_idx] = 1
|
161 |
+
|
162 |
+
cur_context = slice(cur_context.stop, cur_context.stop + 1)
|
163 |
+
txt = [tok.decode(x, skip_special_tokens=True) for x in input_ids.detach().cpu().numpy().tolist()]
|
164 |
+
txt = [
|
165 |
+
unicodedata.normalize("NFKD", x)
|
166 |
+
.replace("\n\n", " ")
|
167 |
+
.replace("<|endoftext|>", "")
|
168 |
+
for x in txt
|
169 |
+
]
|
170 |
+
|
171 |
+
return txt
|
easyeditor/util/globals.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
|
3 |
+
import logging
|
4 |
+
import os
|
5 |
+
|
6 |
+
import yaml
|
7 |
+
|
8 |
+
|
9 |
+
def get_handler(path, log_name):
|
10 |
+
log_file_path = os.path.join(path, log_name)
|
11 |
+
try:
|
12 |
+
if not os.path.exists(path):
|
13 |
+
print("We are creating the logger files")
|
14 |
+
os.makedirs(path)
|
15 |
+
except:
|
16 |
+
pass
|
17 |
+
file_handler = logging.FileHandler(log_file_path)
|
18 |
+
file_handler.setLevel(logging.DEBUG)
|
19 |
+
file_handler.setFormatter(logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s'))
|
20 |
+
|
21 |
+
stream_handler = logging.StreamHandler()
|
22 |
+
stream_handler.setLevel(logging.DEBUG)
|
23 |
+
stream_handler.setFormatter(logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s'))
|
24 |
+
return file_handler, stream_handler
|
25 |
+
|
26 |
+
|
27 |
+
# def get_run_dir(dir_name):
|
28 |
+
#
|
29 |
+
# alg_dir = RESULTS_DIR / dir_name
|
30 |
+
# if alg_dir.exists():
|
31 |
+
# id_list = [
|
32 |
+
# int(str(x).split("_")[-1])
|
33 |
+
# for x in alg_dir.iterdir()
|
34 |
+
# if str(x).split("_")[-1].isnumeric()
|
35 |
+
# ]
|
36 |
+
# run_id = 0 if not id_list else max(id_list) + 1
|
37 |
+
# else:
|
38 |
+
# run_id = 0
|
39 |
+
# run_dir = RESULTS_DIR / dir_name / f"run_{str(run_id).zfill(3)}"
|
40 |
+
# run_dir.mkdir(parents=True, exist_ok=True)
|
41 |
+
# print(f"Results will be stored at {run_dir}")
|
42 |
+
#
|
43 |
+
# return run_dir
|
easyeditor/util/hparams.py
ADDED
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
from dataclasses import dataclass
|
3 |
+
from dataclasses import asdict
|
4 |
+
|
5 |
+
|
6 |
+
@dataclass
|
7 |
+
class HyperParams:
|
8 |
+
"""
|
9 |
+
Simple wrapper to store hyperparameters for Python-based rewriting methods.
|
10 |
+
"""
|
11 |
+
|
12 |
+
@classmethod
|
13 |
+
def from_json(cls, fpath):
|
14 |
+
with open(fpath, "r") as f:
|
15 |
+
data = json.load(f)
|
16 |
+
|
17 |
+
return cls(**data)
|
18 |
+
|
19 |
+
def construct_float_from_scientific_notation(config: dict):
|
20 |
+
for key, value in config.items():
|
21 |
+
if isinstance(value, str):
|
22 |
+
try:
|
23 |
+
# Convert scalar to float if it is in scientific notation format
|
24 |
+
config[key] = float(value)
|
25 |
+
except:
|
26 |
+
pass
|
27 |
+
return config
|
28 |
+
|
29 |
+
def to_dict(config) -> dict:
|
30 |
+
dict = asdict(config)
|
31 |
+
return dict
|
32 |
+
|
33 |
+
|
34 |
+
|
35 |
+
# @classmethod
|
36 |
+
# def from_hparams(cls, hparams_name_or_path: str):
|
37 |
+
#
|
38 |
+
# if '.yaml' not in hparams_name_or_path:
|
39 |
+
# hparams_name_or_path = hparams_name_or_path + '.yaml'
|
40 |
+
# config = compose(hparams_name_or_path)
|
41 |
+
#
|
42 |
+
# assert config.alg_name in ALG_DICT.keys() or print(f'Editing Alg name {config.alg_name} not supported yet.')
|
43 |
+
#
|
44 |
+
# params_class, apply_algo = ALG_DICT[config.alg_name]
|
45 |
+
#
|
46 |
+
# return params_class(**config)
|
easyeditor/util/logit_lens.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from collections import defaultdict
|
2 |
+
from typing import Dict, Optional
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
6 |
+
|
7 |
+
from . import nethook
|
8 |
+
|
9 |
+
|
10 |
+
class LogitLens:
|
11 |
+
"""
|
12 |
+
Applies the LM head at the output of each hidden layer, then analyzes the
|
13 |
+
resultant token probability distribution.
|
14 |
+
|
15 |
+
Only works when hooking outputs of *one* individual generation.
|
16 |
+
|
17 |
+
Inspiration: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens
|
18 |
+
|
19 |
+
Warning: when running multiple times (e.g. generation), will return
|
20 |
+
outputs _only_ for the last processing step.
|
21 |
+
"""
|
22 |
+
|
23 |
+
def __init__(
|
24 |
+
self,
|
25 |
+
model: AutoModelForCausalLM,
|
26 |
+
tok: AutoTokenizer,
|
27 |
+
layer_module_tmp: str,
|
28 |
+
ln_f_module: str,
|
29 |
+
lm_head_module: str,
|
30 |
+
disabled: bool = False,
|
31 |
+
):
|
32 |
+
self.disabled = disabled
|
33 |
+
self.model, self.tok = model, tok
|
34 |
+
self.n_layers = self.model.config.n_layer
|
35 |
+
|
36 |
+
self.lm_head, self.ln_f = (
|
37 |
+
nethook.get_module(model, lm_head_module),
|
38 |
+
nethook.get_module(model, ln_f_module),
|
39 |
+
)
|
40 |
+
|
41 |
+
self.output: Optional[Dict] = None
|
42 |
+
self.td: Optional[nethook.TraceDict] = None
|
43 |
+
self.trace_layers = [
|
44 |
+
layer_module_tmp.format(layer) for layer in range(self.n_layers)
|
45 |
+
]
|
46 |
+
|
47 |
+
def __enter__(self):
|
48 |
+
if not self.disabled:
|
49 |
+
self.td = nethook.TraceDict(
|
50 |
+
self.model,
|
51 |
+
self.trace_layers,
|
52 |
+
retain_input=False,
|
53 |
+
retain_output=True,
|
54 |
+
)
|
55 |
+
self.td.__enter__()
|
56 |
+
|
57 |
+
def __exit__(self, *args):
|
58 |
+
if self.disabled:
|
59 |
+
return
|
60 |
+
self.td.__exit__(*args)
|
61 |
+
|
62 |
+
self.output = {layer: [] for layer in range(self.n_layers)}
|
63 |
+
|
64 |
+
with torch.no_grad():
|
65 |
+
for layer, (_, t) in enumerate(self.td.items()):
|
66 |
+
cur_out = t.output[0]
|
67 |
+
assert (
|
68 |
+
cur_out.size(0) == 1
|
69 |
+
), "Make sure you're only running LogitLens on single generations only."
|
70 |
+
|
71 |
+
self.output[layer] = torch.softmax(
|
72 |
+
self.lm_head(self.ln_f(cur_out[:, -1, :])), dim=1
|
73 |
+
)
|
74 |
+
|
75 |
+
return self.output
|
76 |
+
|
77 |
+
def pprint(self, k=5):
|
78 |
+
to_print = defaultdict(list)
|
79 |
+
|
80 |
+
for layer, pred in self.output.items():
|
81 |
+
rets = torch.topk(pred[0], k)
|
82 |
+
for i in range(k):
|
83 |
+
to_print[layer].append(
|
84 |
+
(
|
85 |
+
self.tok.decode(rets[1][i]),
|
86 |
+
round(rets[0][i].item() * 1e2) / 1e2,
|
87 |
+
)
|
88 |
+
)
|
89 |
+
|
90 |
+
print(
|
91 |
+
"\n".join(
|
92 |
+
[
|
93 |
+
f"{layer}: {[(el[0], round(el[1] * 1e2)) for el in to_print[layer]]}"
|
94 |
+
for layer in range(self.n_layers)
|
95 |
+
]
|
96 |
+
)
|
97 |
+
)
|
easyeditor/util/nethook.py
ADDED
@@ -0,0 +1,451 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Utilities for instrumenting a torch model.
|
3 |
+
|
4 |
+
Trace will hook one layer at a time.
|
5 |
+
TraceDict will hook multiple layers at once.
|
6 |
+
subsequence slices intervals from Sequential modules.
|
7 |
+
get_module, replace_module, get_parameter resolve dotted names.
|
8 |
+
set_requires_grad recursively sets requires_grad in module parameters.
|
9 |
+
"""
|
10 |
+
|
11 |
+
import contextlib
|
12 |
+
import copy
|
13 |
+
import inspect
|
14 |
+
from collections import OrderedDict
|
15 |
+
|
16 |
+
import torch
|
17 |
+
|
18 |
+
|
19 |
+
class Trace(contextlib.AbstractContextManager):
|
20 |
+
"""
|
21 |
+
To retain the output of the named layer during the computation of
|
22 |
+
the given network:
|
23 |
+
|
24 |
+
with Trace(net, 'layer.name') as ret:
|
25 |
+
_ = net(inp)
|
26 |
+
representation = ret.output
|
27 |
+
|
28 |
+
A layer module can be passed directly without a layer name, and
|
29 |
+
its output will be retained. By default, a direct reference to
|
30 |
+
the output object is returned, but options can control this:
|
31 |
+
|
32 |
+
clone=True - retains a copy of the output, which can be
|
33 |
+
useful if you want to see the output before it might
|
34 |
+
be modified by the network in-place later.
|
35 |
+
detach=True - retains a detached reference or copy. (By
|
36 |
+
default the value would be left attached to the graph.)
|
37 |
+
retain_grad=True - request gradient to be retained on the
|
38 |
+
output. After backward(), ret.output.grad is populated.
|
39 |
+
|
40 |
+
retain_input=True - also retains the input.
|
41 |
+
retain_output=False - can disable retaining the output.
|
42 |
+
edit_output=fn - calls the function to modify the output
|
43 |
+
of the layer before passing it the rest of the model.
|
44 |
+
fn can optionally accept (output, layer) arguments
|
45 |
+
for the original output and the layer name.
|
46 |
+
stop=True - throws a StopForward exception after the layer
|
47 |
+
is run, which allows running just a portion of a model.
|
48 |
+
"""
|
49 |
+
|
50 |
+
def __init__(
|
51 |
+
self,
|
52 |
+
module,
|
53 |
+
layer=None,
|
54 |
+
retain_output=True,
|
55 |
+
retain_input=False,
|
56 |
+
clone=False,
|
57 |
+
detach=False,
|
58 |
+
retain_grad=False,
|
59 |
+
edit_output=None,
|
60 |
+
stop=False,
|
61 |
+
):
|
62 |
+
"""
|
63 |
+
Method to replace a forward method with a closure that
|
64 |
+
intercepts the call, and tracks the hook so that it can be reverted.
|
65 |
+
"""
|
66 |
+
retainer = self
|
67 |
+
self.layer = layer
|
68 |
+
if layer is not None:
|
69 |
+
module = get_module(module, layer)
|
70 |
+
|
71 |
+
def retain_hook(m, inputs, output):
|
72 |
+
if retain_input:
|
73 |
+
retainer.input = recursive_copy(
|
74 |
+
inputs[0] if len(inputs) == 1 else inputs,
|
75 |
+
clone=clone,
|
76 |
+
detach=detach,
|
77 |
+
retain_grad=False,
|
78 |
+
) # retain_grad applies to output only.
|
79 |
+
if edit_output:
|
80 |
+
output = invoke_with_optional_args(
|
81 |
+
edit_output, output=output, layer=self.layer
|
82 |
+
)
|
83 |
+
if retain_output:
|
84 |
+
retainer.output = recursive_copy(
|
85 |
+
output, clone=clone, detach=detach, retain_grad=retain_grad
|
86 |
+
)
|
87 |
+
# When retain_grad is set, also insert a trivial
|
88 |
+
# copy operation. That allows in-place operations
|
89 |
+
# to follow without error.
|
90 |
+
if retain_grad:
|
91 |
+
output = recursive_copy(retainer.output, clone=True, detach=False)
|
92 |
+
if stop:
|
93 |
+
raise StopForward()
|
94 |
+
return output
|
95 |
+
|
96 |
+
self.registered_hook = module.register_forward_hook(retain_hook)
|
97 |
+
self.stop = stop
|
98 |
+
|
99 |
+
def __enter__(self):
|
100 |
+
return self
|
101 |
+
|
102 |
+
def __exit__(self, type, value, traceback):
|
103 |
+
self.close()
|
104 |
+
if self.stop and issubclass(type, StopForward):
|
105 |
+
return True
|
106 |
+
|
107 |
+
def close(self):
|
108 |
+
self.registered_hook.remove()
|
109 |
+
|
110 |
+
|
111 |
+
class TraceDict(OrderedDict, contextlib.AbstractContextManager):
|
112 |
+
"""
|
113 |
+
To retain the output of multiple named layers during the computation
|
114 |
+
of the given network:
|
115 |
+
|
116 |
+
with TraceDict(net, ['layer1.name1', 'layer2.name2']) as ret:
|
117 |
+
_ = net(inp)
|
118 |
+
representation = ret['layer1.name1'].output
|
119 |
+
|
120 |
+
If edit_output is provided, it should be a function that takes
|
121 |
+
two arguments: output, and the layer name; and then it returns the
|
122 |
+
modified output.
|
123 |
+
|
124 |
+
Other arguments are the same as Trace. If stop is True, then the
|
125 |
+
execution of the network will be stopped after the last layer
|
126 |
+
listed (even if it would not have been the last to be executed).
|
127 |
+
"""
|
128 |
+
|
129 |
+
def __init__(
|
130 |
+
self,
|
131 |
+
module,
|
132 |
+
layers=None,
|
133 |
+
retain_output=True,
|
134 |
+
retain_input=False,
|
135 |
+
clone=False,
|
136 |
+
detach=False,
|
137 |
+
retain_grad=False,
|
138 |
+
edit_output=None,
|
139 |
+
stop=False,
|
140 |
+
):
|
141 |
+
self.stop = stop
|
142 |
+
|
143 |
+
def flag_last_unseen(it):
|
144 |
+
try:
|
145 |
+
it = iter(it)
|
146 |
+
prev = next(it)
|
147 |
+
seen = set([prev])
|
148 |
+
except StopIteration:
|
149 |
+
return
|
150 |
+
for item in it:
|
151 |
+
if item not in seen:
|
152 |
+
yield False, prev
|
153 |
+
seen.add(item)
|
154 |
+
prev = item
|
155 |
+
yield True, prev
|
156 |
+
|
157 |
+
for is_last, layer in flag_last_unseen(layers):
|
158 |
+
self[layer] = Trace(
|
159 |
+
module=module,
|
160 |
+
layer=layer,
|
161 |
+
retain_output=retain_output,
|
162 |
+
retain_input=retain_input,
|
163 |
+
clone=clone,
|
164 |
+
detach=detach,
|
165 |
+
retain_grad=retain_grad,
|
166 |
+
edit_output=edit_output,
|
167 |
+
stop=stop and is_last,
|
168 |
+
)
|
169 |
+
|
170 |
+
def __enter__(self):
|
171 |
+
return self
|
172 |
+
|
173 |
+
def __exit__(self, type, value, traceback):
|
174 |
+
self.close()
|
175 |
+
if self.stop and issubclass(type, StopForward):
|
176 |
+
return True
|
177 |
+
|
178 |
+
def close(self):
|
179 |
+
for layer, trace in reversed(self.items()):
|
180 |
+
trace.close()
|
181 |
+
|
182 |
+
|
183 |
+
class StopForward(Exception):
|
184 |
+
"""
|
185 |
+
If the only output needed from running a network is the retained
|
186 |
+
submodule then Trace(submodule, stop=True) will stop execution
|
187 |
+
immediately after the retained submodule by raising the StopForward()
|
188 |
+
exception. When Trace is used as context manager, it catches that
|
189 |
+
exception and can be used as follows:
|
190 |
+
|
191 |
+
with Trace(net, layername, stop=True) as tr:
|
192 |
+
net(inp) # Only runs the network up to layername
|
193 |
+
print(tr.output)
|
194 |
+
"""
|
195 |
+
|
196 |
+
pass
|
197 |
+
|
198 |
+
|
199 |
+
def recursive_copy(x, clone=None, detach=None, retain_grad=None):
|
200 |
+
"""
|
201 |
+
Copies a reference to a tensor, or an object that contains tensors,
|
202 |
+
optionally detaching and cloning the tensor(s). If retain_grad is
|
203 |
+
true, the original tensors are marked to have grads retained.
|
204 |
+
"""
|
205 |
+
if not clone and not detach and not retain_grad:
|
206 |
+
return x
|
207 |
+
if isinstance(x, torch.Tensor):
|
208 |
+
if retain_grad:
|
209 |
+
if not x.requires_grad:
|
210 |
+
x.requires_grad = True
|
211 |
+
x.retain_grad()
|
212 |
+
elif detach:
|
213 |
+
x = x.detach()
|
214 |
+
if clone:
|
215 |
+
x = x.clone()
|
216 |
+
return x
|
217 |
+
# Only dicts, lists, and tuples (and subclasses) can be copied.
|
218 |
+
if isinstance(x, dict):
|
219 |
+
return type(x)({k: recursive_copy(v) for k, v in x.items()})
|
220 |
+
elif isinstance(x, (list, tuple)):
|
221 |
+
return type(x)([recursive_copy(v) for v in x])
|
222 |
+
else:
|
223 |
+
assert False, f"Unknown type {type(x)} cannot be broken into tensors."
|
224 |
+
|
225 |
+
|
226 |
+
def subsequence(
|
227 |
+
sequential,
|
228 |
+
first_layer=None,
|
229 |
+
last_layer=None,
|
230 |
+
after_layer=None,
|
231 |
+
upto_layer=None,
|
232 |
+
single_layer=None,
|
233 |
+
share_weights=False,
|
234 |
+
):
|
235 |
+
"""
|
236 |
+
Creates a subsequence of a pytorch Sequential model, copying over
|
237 |
+
modules together with parameters for the subsequence. Only
|
238 |
+
modules from first_layer to last_layer (inclusive) are included,
|
239 |
+
or modules between after_layer and upto_layer (exclusive).
|
240 |
+
Handles descent into dotted layer names as long as all references
|
241 |
+
are within nested Sequential models.
|
242 |
+
|
243 |
+
If share_weights is True, then references the original modules
|
244 |
+
and their parameters without copying them. Otherwise, by default,
|
245 |
+
makes a separate brand-new copy.
|
246 |
+
"""
|
247 |
+
assert (single_layer is None) or (
|
248 |
+
first_layer is last_layer is after_layer is upto_layer is None
|
249 |
+
)
|
250 |
+
if single_layer is not None:
|
251 |
+
first_layer = single_layer
|
252 |
+
last_layer = single_layer
|
253 |
+
first, last, after, upto = [
|
254 |
+
None if d is None else d.split(".")
|
255 |
+
for d in [first_layer, last_layer, after_layer, upto_layer]
|
256 |
+
]
|
257 |
+
return hierarchical_subsequence(
|
258 |
+
sequential,
|
259 |
+
first=first,
|
260 |
+
last=last,
|
261 |
+
after=after,
|
262 |
+
upto=upto,
|
263 |
+
share_weights=share_weights,
|
264 |
+
)
|
265 |
+
|
266 |
+
|
267 |
+
def hierarchical_subsequence(
|
268 |
+
sequential, first, last, after, upto, share_weights=False, depth=0
|
269 |
+
):
|
270 |
+
"""
|
271 |
+
Recursive helper for subsequence() to support descent into dotted
|
272 |
+
layer names. In this helper, first, last, after, and upto are
|
273 |
+
arrays of names resulting from splitting on dots. Can only
|
274 |
+
descend into nested Sequentials.
|
275 |
+
"""
|
276 |
+
assert (last is None) or (upto is None)
|
277 |
+
assert (first is None) or (after is None)
|
278 |
+
if first is last is after is upto is None:
|
279 |
+
return sequential if share_weights else copy.deepcopy(sequential)
|
280 |
+
assert isinstance(sequential, torch.nn.Sequential), (
|
281 |
+
".".join((first or last or after or upto)[:depth] or "arg") + " not Sequential"
|
282 |
+
)
|
283 |
+
including_children = (first is None) and (after is None)
|
284 |
+
included_children = OrderedDict()
|
285 |
+
# A = current level short name of A.
|
286 |
+
# AN = full name for recursive descent if not innermost.
|
287 |
+
(F, FN), (L, LN), (A, AN), (U, UN) = [
|
288 |
+
(d[depth], (None if len(d) == depth + 1 else d))
|
289 |
+
if d is not None
|
290 |
+
else (None, None)
|
291 |
+
for d in [first, last, after, upto]
|
292 |
+
]
|
293 |
+
for name, layer in sequential._modules.items():
|
294 |
+
if name == F:
|
295 |
+
first = None
|
296 |
+
including_children = True
|
297 |
+
if name == A and AN is not None: # just like F if not a leaf.
|
298 |
+
after = None
|
299 |
+
including_children = True
|
300 |
+
if name == U and UN is None:
|
301 |
+
upto = None
|
302 |
+
including_children = False
|
303 |
+
if including_children:
|
304 |
+
# AR = full name for recursive descent if name matches.
|
305 |
+
FR, LR, AR, UR = [
|
306 |
+
n if n is None or n[depth] == name else None for n in [FN, LN, AN, UN]
|
307 |
+
]
|
308 |
+
chosen = hierarchical_subsequence(
|
309 |
+
layer,
|
310 |
+
first=FR,
|
311 |
+
last=LR,
|
312 |
+
after=AR,
|
313 |
+
upto=UR,
|
314 |
+
share_weights=share_weights,
|
315 |
+
depth=depth + 1,
|
316 |
+
)
|
317 |
+
if chosen is not None:
|
318 |
+
included_children[name] = chosen
|
319 |
+
if name == L:
|
320 |
+
last = None
|
321 |
+
including_children = False
|
322 |
+
if name == U and UN is not None: # just like L if not a leaf.
|
323 |
+
upto = None
|
324 |
+
including_children = False
|
325 |
+
if name == A and AN is None:
|
326 |
+
after = None
|
327 |
+
including_children = True
|
328 |
+
for name in [first, last, after, upto]:
|
329 |
+
if name is not None:
|
330 |
+
raise ValueError("Layer %s not found" % ".".join(name))
|
331 |
+
# Omit empty subsequences except at the outermost level,
|
332 |
+
# where we should not return None.
|
333 |
+
if not len(included_children) and depth > 0:
|
334 |
+
return None
|
335 |
+
result = torch.nn.Sequential(included_children)
|
336 |
+
result.training = sequential.training
|
337 |
+
return result
|
338 |
+
|
339 |
+
|
340 |
+
def set_requires_grad(requires_grad, *models):
|
341 |
+
"""
|
342 |
+
Sets requires_grad true or false for all parameters within the
|
343 |
+
models passed.
|
344 |
+
"""
|
345 |
+
for model in models:
|
346 |
+
if isinstance(model, torch.nn.Module):
|
347 |
+
for param in model.parameters():
|
348 |
+
param.requires_grad = requires_grad
|
349 |
+
elif isinstance(model, (torch.nn.Parameter, torch.Tensor)):
|
350 |
+
model.requires_grad = requires_grad
|
351 |
+
else:
|
352 |
+
assert False, "unknown type %r" % type(model)
|
353 |
+
|
354 |
+
|
355 |
+
def get_module(model, name):
|
356 |
+
"""
|
357 |
+
Finds the named module within the given model.
|
358 |
+
"""
|
359 |
+
for n, m in model.named_modules():
|
360 |
+
if n == name:
|
361 |
+
return m
|
362 |
+
raise LookupError(name)
|
363 |
+
|
364 |
+
|
365 |
+
def get_parameter(model, name):
|
366 |
+
"""
|
367 |
+
Finds the named parameter within the given model.
|
368 |
+
"""
|
369 |
+
for n, p in model.named_parameters():
|
370 |
+
if n == name:
|
371 |
+
return p
|
372 |
+
raise LookupError(name)
|
373 |
+
|
374 |
+
|
375 |
+
def replace_module(model, name, new_module):
|
376 |
+
"""
|
377 |
+
Replaces the named module within the given model.
|
378 |
+
"""
|
379 |
+
if "." in name:
|
380 |
+
parent_name, attr_name = name.rsplit(".", 1)
|
381 |
+
model = get_module(model, parent_name)
|
382 |
+
# original_module = getattr(model, attr_name)
|
383 |
+
setattr(model, attr_name, new_module)
|
384 |
+
|
385 |
+
|
386 |
+
def invoke_with_optional_args(fn, *args, **kwargs):
|
387 |
+
"""
|
388 |
+
Invokes a function with only the arguments that it
|
389 |
+
is written to accept, giving priority to arguments
|
390 |
+
that match by-name, using the following rules.
|
391 |
+
(1) arguments with matching names are passed by name.
|
392 |
+
(2) remaining non-name-matched args are passed by order.
|
393 |
+
(3) extra caller arguments that the function cannot
|
394 |
+
accept are not passed.
|
395 |
+
(4) extra required function arguments that the caller
|
396 |
+
cannot provide cause a TypeError to be raised.
|
397 |
+
Ordinary python calling conventions are helpful for
|
398 |
+
supporting a function that might be revised to accept
|
399 |
+
extra arguments in a newer version, without requiring the
|
400 |
+
caller to pass those new arguments. This function helps
|
401 |
+
support function callers that might be revised to supply
|
402 |
+
extra arguments, without requiring the callee to accept
|
403 |
+
those new arguments.
|
404 |
+
"""
|
405 |
+
argspec = inspect.getfullargspec(fn)
|
406 |
+
pass_args = []
|
407 |
+
used_kw = set()
|
408 |
+
unmatched_pos = []
|
409 |
+
used_pos = 0
|
410 |
+
defaulted_pos = len(argspec.args) - (
|
411 |
+
0 if not argspec.defaults else len(argspec.defaults)
|
412 |
+
)
|
413 |
+
# Pass positional args that match name first, then by position.
|
414 |
+
for i, n in enumerate(argspec.args):
|
415 |
+
if n in kwargs:
|
416 |
+
pass_args.append(kwargs[n])
|
417 |
+
used_kw.add(n)
|
418 |
+
elif used_pos < len(args):
|
419 |
+
pass_args.append(args[used_pos])
|
420 |
+
used_pos += 1
|
421 |
+
else:
|
422 |
+
unmatched_pos.append(len(pass_args))
|
423 |
+
pass_args.append(
|
424 |
+
None if i < defaulted_pos else argspec.defaults[i - defaulted_pos]
|
425 |
+
)
|
426 |
+
# Fill unmatched positional args with unmatched keyword args in order.
|
427 |
+
if len(unmatched_pos):
|
428 |
+
for k, v in kwargs.items():
|
429 |
+
if k in used_kw or k in argspec.kwonlyargs:
|
430 |
+
continue
|
431 |
+
pass_args[unmatched_pos[0]] = v
|
432 |
+
used_kw.add(k)
|
433 |
+
unmatched_pos = unmatched_pos[1:]
|
434 |
+
if len(unmatched_pos) == 0:
|
435 |
+
break
|
436 |
+
else:
|
437 |
+
if unmatched_pos[0] < defaulted_pos:
|
438 |
+
unpassed = ", ".join(
|
439 |
+
argspec.args[u] for u in unmatched_pos if u < defaulted_pos
|
440 |
+
)
|
441 |
+
raise TypeError(f"{fn.__name__}() cannot be passed {unpassed}.")
|
442 |
+
# Pass remaining kw args if they can be accepted.
|
443 |
+
pass_kw = {
|
444 |
+
k: v
|
445 |
+
for k, v in kwargs.items()
|
446 |
+
if k not in used_kw and (k in argspec.kwonlyargs or argspec.varargs is not None)
|
447 |
+
}
|
448 |
+
# Pass remaining positional args if they can be accepted.
|
449 |
+
if argspec.varargs is not None:
|
450 |
+
pass_args += list(args[used_pos:])
|
451 |
+
return fn(*pass_args, **pass_kw)
|
easyeditor/util/perplexity.py
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
3 |
+
|
4 |
+
|
5 |
+
def perplexity(
|
6 |
+
model: AutoModelForCausalLM,
|
7 |
+
tok: AutoTokenizer,
|
8 |
+
text: str,
|
9 |
+
max_input_length: int = None,
|
10 |
+
):
|
11 |
+
"""
|
12 |
+
Computes perplexity of a piece of text, measured on a reference model.
|
13 |
+
Text is truncated to max_input_length tokens.
|
14 |
+
"""
|
15 |
+
|
16 |
+
inputs = tok(
|
17 |
+
[text], return_tensors="pt", max_length=max_input_length, truncation=True
|
18 |
+
).to("cuda")
|
19 |
+
|
20 |
+
logits = torch.nn.functional.log_softmax(model(**inputs).logits, dim=2)
|
21 |
+
log_probs = torch.gather(logits[:, :-1, :], 2, inputs["input_ids"][:, 1:, None])[0]
|
22 |
+
|
23 |
+
# Perplexity = exp(-1/N * log P(x_1, ..., x_n))
|
24 |
+
return torch.exp(-1 / inputs["input_ids"].size(1) * log_probs.sum()).item()
|
easyeditor/util/runningstats.py
ADDED
@@ -0,0 +1,1883 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
To use a runningstats object,
|
3 |
+
|
4 |
+
1. Create the the desired stat object, e.g., `m = Mean()`
|
5 |
+
2. Feed it batches via the add method, e.g., `m.add(batch)`
|
6 |
+
3. Repeat step 2 any number of times.
|
7 |
+
4. Read out the statistic of interest, e.g., `m.mean()`
|
8 |
+
|
9 |
+
Built-in runningstats objects include:
|
10 |
+
|
11 |
+
Mean - produces mean().
|
12 |
+
Variance - mean() and variance() and stdev().
|
13 |
+
Covariance - mean(), covariance(), correlation(), variance(), stdev().
|
14 |
+
SecondMoment - moment() is the non-mean-centered covariance, E[x x^T].
|
15 |
+
Quantile - quantile(), min(), max(), median(), mean(), variance(), stdev().
|
16 |
+
TopK - topk() returns (values, indexes).
|
17 |
+
Bincount - bincount() histograms nonnegative integer data.
|
18 |
+
IoU - intersection(), union(), iou() tally binary co-occurrences.
|
19 |
+
History - history() returns concatenation of data.
|
20 |
+
CrossCovariance - covariance between two signals, without self-covariance.
|
21 |
+
CrossIoU - iou between two signals, without self-IoU.
|
22 |
+
CombinedStat - aggregates any set of stats.
|
23 |
+
|
24 |
+
Add more running stats by subclassing the Stat class.
|
25 |
+
|
26 |
+
These statistics are vectorized along dim>=1, so stat.add()
|
27 |
+
should supply a two-dimensional input where the zeroth
|
28 |
+
dimension is the batch/sampling dimension and the first
|
29 |
+
dimension is the feature dimension.
|
30 |
+
|
31 |
+
The data type and device used matches the data passed to add();
|
32 |
+
for example, for higher-precision covariances, convert to double
|
33 |
+
before calling add().
|
34 |
+
|
35 |
+
It is common to want to compute and remember a statistic sampled
|
36 |
+
over a Dataset, computed in batches, possibly caching the computed
|
37 |
+
statistic in a file. The tally(stat, dataset, cache) handles
|
38 |
+
this pattern. It takes a statistic, a dataset, and a cache filename
|
39 |
+
and sets up a data loader that can be run (or not, if cached) to
|
40 |
+
compute the statistic, adopting the convention that cached stats are
|
41 |
+
saved to and loaded from numpy npz files.
|
42 |
+
"""
|
43 |
+
|
44 |
+
import math
|
45 |
+
import os
|
46 |
+
import random
|
47 |
+
import struct
|
48 |
+
|
49 |
+
import numpy
|
50 |
+
import torch
|
51 |
+
from torch.utils.data.sampler import Sampler
|
52 |
+
|
53 |
+
|
54 |
+
def tally(stat, dataset, cache=None, quiet=False, **kwargs):
|
55 |
+
"""
|
56 |
+
To use tally, write code like the following.
|
57 |
+
|
58 |
+
stat = Mean()
|
59 |
+
ds = MyDataset()
|
60 |
+
for batch in tally(stat, ds, cache='mymean.npz', batch_size=50):
|
61 |
+
stat.add(batch)
|
62 |
+
mean = stat.mean()
|
63 |
+
|
64 |
+
The first argument should be the Stat being computed. After the
|
65 |
+
loader is exhausted, tally will bring this stat to the cpu and
|
66 |
+
cache it (if a cache is specified).
|
67 |
+
|
68 |
+
The dataset can be a torch Dataset or a plain Tensor, or it can
|
69 |
+
be a callable that returns one of those.
|
70 |
+
|
71 |
+
Details on caching via the cache= argument:
|
72 |
+
|
73 |
+
If the given filename cannot be loaded, tally will leave the
|
74 |
+
statistic object empty and set up a DataLoader object so that
|
75 |
+
the loop can be run. After the last iteration of the loop, the
|
76 |
+
completed statistic will be moved to the cpu device and also
|
77 |
+
saved in the cache file.
|
78 |
+
|
79 |
+
If the cached statistic can be loaded from the given file, tally
|
80 |
+
will not set up the data loader and instead will return a fully
|
81 |
+
loaded statistic object (on the cpu device) and an empty list as
|
82 |
+
the loader.
|
83 |
+
|
84 |
+
The `with cache_load_enabled(False):` context manager can
|
85 |
+
be used to disable loading from the cache.
|
86 |
+
|
87 |
+
If needed, a DataLoader will be created to wrap the dataset:
|
88 |
+
|
89 |
+
Keyword arguments of tally are passed to the DataLoader,
|
90 |
+
so batch_size, num_workers, pin_memory, etc. can be specified.
|
91 |
+
|
92 |
+
Subsampling is supported via sample_size= and random_sample=:
|
93 |
+
|
94 |
+
If sample_size=N is specified, rather than loading the whole
|
95 |
+
dataset, only the first N items are sampled. If additionally
|
96 |
+
random_sample=S is specified, the pseudorandom seed S will be
|
97 |
+
used to select a fixed psedorandom sample of size N to sample.
|
98 |
+
"""
|
99 |
+
assert isinstance(stat, Stat)
|
100 |
+
args = {}
|
101 |
+
for k in ["sample_size"]:
|
102 |
+
if k in kwargs:
|
103 |
+
args[k] = kwargs[k]
|
104 |
+
cached_state = load_cached_state(cache, args, quiet=quiet)
|
105 |
+
if cached_state is not None:
|
106 |
+
stat.load_state_dict(cached_state)
|
107 |
+
|
108 |
+
def empty_loader():
|
109 |
+
return
|
110 |
+
yield
|
111 |
+
|
112 |
+
return empty_loader()
|
113 |
+
loader = make_loader(dataset, **kwargs)
|
114 |
+
|
115 |
+
def wrapped_loader():
|
116 |
+
yield from loader
|
117 |
+
stat.to_(device="cpu")
|
118 |
+
if cache is not None:
|
119 |
+
save_cached_state(cache, stat, args)
|
120 |
+
|
121 |
+
return wrapped_loader()
|
122 |
+
|
123 |
+
|
124 |
+
class cache_load_enabled:
|
125 |
+
"""
|
126 |
+
When used as a context manager, cache_load_enabled(False) will prevent
|
127 |
+
tally from loading cached statsitics, forcing them to be recomputed.
|
128 |
+
"""
|
129 |
+
|
130 |
+
def __init__(self, enabled=True):
|
131 |
+
self.prev = False
|
132 |
+
self.enabled = enabled
|
133 |
+
|
134 |
+
def __enter__(self):
|
135 |
+
global global_load_cache_enabled
|
136 |
+
self.prev = global_load_cache_enabled
|
137 |
+
global_load_cache_enabled = self.enabled
|
138 |
+
|
139 |
+
def __exit__(self, exc_type, exc_value, traceback):
|
140 |
+
global global_load_cache_enabled
|
141 |
+
global_load_cache_enabled = self.prev
|
142 |
+
|
143 |
+
|
144 |
+
class Stat:
|
145 |
+
"""
|
146 |
+
Abstract base class for a running pytorch statistic.
|
147 |
+
"""
|
148 |
+
|
149 |
+
def __init__(self, state):
|
150 |
+
"""
|
151 |
+
By convention, all Stat subclasses can be initialized by passing
|
152 |
+
state=; and then they will initialize by calling load_state_dict.
|
153 |
+
"""
|
154 |
+
self.load_state_dict(resolve_state_dict(state))
|
155 |
+
|
156 |
+
def add(self, x, *args, **kwargs):
|
157 |
+
"""
|
158 |
+
Observes a batch of samples to be incorporated into the statistic.
|
159 |
+
Dimension 0 should be the batch dimension, and dimension 1 should
|
160 |
+
be the feature dimension of the pytorch tensor x.
|
161 |
+
"""
|
162 |
+
pass
|
163 |
+
|
164 |
+
def load_state_dict(self, d):
|
165 |
+
"""
|
166 |
+
Loads this Stat from a dictionary of numpy arrays as saved
|
167 |
+
by state_dict.
|
168 |
+
"""
|
169 |
+
pass
|
170 |
+
|
171 |
+
def state_dict(self):
|
172 |
+
"""
|
173 |
+
Saves this Stat as a dictionary of numpy arrays that can be
|
174 |
+
stored in an npz or reloaded later using load_state_dict.
|
175 |
+
"""
|
176 |
+
return {}
|
177 |
+
|
178 |
+
def save(self, filename):
|
179 |
+
"""
|
180 |
+
Saves this stat as an npz file containing the state_dict.
|
181 |
+
"""
|
182 |
+
save_cached_state(filename, self, {})
|
183 |
+
|
184 |
+
def load(self, filename):
|
185 |
+
"""
|
186 |
+
Loads this stat from an npz file containing a saved state_dict.
|
187 |
+
"""
|
188 |
+
self.load_state_dict(load_cached_state(filename, {}, quiet=True, throw=True))
|
189 |
+
|
190 |
+
def to_(self, device):
|
191 |
+
"""
|
192 |
+
Moves this Stat to the given device.
|
193 |
+
"""
|
194 |
+
pass
|
195 |
+
|
196 |
+
def cpu_(self):
|
197 |
+
"""
|
198 |
+
Moves this Stat to the cpu device.
|
199 |
+
"""
|
200 |
+
self.to_("cpu")
|
201 |
+
|
202 |
+
def cuda_(self):
|
203 |
+
"""
|
204 |
+
Moves this Stat to the default cuda device.
|
205 |
+
"""
|
206 |
+
self.to_("cuda")
|
207 |
+
|
208 |
+
def _normalize_add_shape(self, x, attr="data_shape"):
|
209 |
+
"""
|
210 |
+
Flattens input data to 2d.
|
211 |
+
"""
|
212 |
+
if not torch.is_tensor(x):
|
213 |
+
x = torch.tensor(x)
|
214 |
+
if len(x.shape) < 1:
|
215 |
+
x = x.view(-1)
|
216 |
+
data_shape = getattr(self, attr, None)
|
217 |
+
if data_shape is None:
|
218 |
+
data_shape = x.shape[1:]
|
219 |
+
setattr(self, attr, data_shape)
|
220 |
+
else:
|
221 |
+
assert x.shape[1:] == data_shape
|
222 |
+
return x.view(x.shape[0], int(numpy.prod(data_shape)))
|
223 |
+
|
224 |
+
def _restore_result_shape(self, x, attr="data_shape"):
|
225 |
+
"""
|
226 |
+
Restores output data to input data shape.
|
227 |
+
"""
|
228 |
+
data_shape = getattr(self, attr, None)
|
229 |
+
if data_shape is None:
|
230 |
+
return x
|
231 |
+
return x.view(data_shape * len(x.shape))
|
232 |
+
|
233 |
+
|
234 |
+
class Mean(Stat):
|
235 |
+
"""
|
236 |
+
Running mean.
|
237 |
+
"""
|
238 |
+
|
239 |
+
def __init__(self, state=None):
|
240 |
+
if state is not None:
|
241 |
+
return super().__init__(state)
|
242 |
+
self.count = 0
|
243 |
+
self.batchcount = 0
|
244 |
+
self._mean = None
|
245 |
+
self.data_shape = None
|
246 |
+
|
247 |
+
def add(self, a):
|
248 |
+
a = self._normalize_add_shape(a)
|
249 |
+
if len(a) == 0:
|
250 |
+
return
|
251 |
+
batch_count = a.shape[0]
|
252 |
+
batch_mean = a.sum(0) / batch_count
|
253 |
+
self.batchcount += 1
|
254 |
+
# Initial batch.
|
255 |
+
if self._mean is None:
|
256 |
+
self.count = batch_count
|
257 |
+
self._mean = batch_mean
|
258 |
+
return
|
259 |
+
# Update a batch using Chan-style update for numerical stability.
|
260 |
+
self.count += batch_count
|
261 |
+
new_frac = float(batch_count) / self.count
|
262 |
+
# Update the mean according to the batch deviation from the old mean.
|
263 |
+
delta = batch_mean.sub_(self._mean).mul_(new_frac)
|
264 |
+
self._mean.add_(delta)
|
265 |
+
|
266 |
+
def size(self):
|
267 |
+
return self.count
|
268 |
+
|
269 |
+
def mean(self):
|
270 |
+
return self._restore_result_shape(self._mean)
|
271 |
+
|
272 |
+
def to_(self, device):
|
273 |
+
if self._mean is not None:
|
274 |
+
self._mean = self._mean.to(device)
|
275 |
+
|
276 |
+
def load_state_dict(self, state):
|
277 |
+
self.count = state["count"]
|
278 |
+
self.batchcount = state["batchcount"]
|
279 |
+
self._mean = torch.from_numpy(state["mean"])
|
280 |
+
self.data_shape = (
|
281 |
+
None if state["data_shape"] is None else tuple(state["data_shape"])
|
282 |
+
)
|
283 |
+
|
284 |
+
def state_dict(self):
|
285 |
+
return dict(
|
286 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
287 |
+
count=self.count,
|
288 |
+
data_shape=self.data_shape and tuple(self.data_shape),
|
289 |
+
batchcount=self.batchcount,
|
290 |
+
mean=self._mean.cpu().numpy(),
|
291 |
+
)
|
292 |
+
|
293 |
+
|
294 |
+
class NormMean(Mean):
|
295 |
+
"""
|
296 |
+
Running average of the norm of input vectors
|
297 |
+
"""
|
298 |
+
|
299 |
+
def __init__(self, state=None):
|
300 |
+
super().__init__(state)
|
301 |
+
|
302 |
+
def add(self, a):
|
303 |
+
super().add(a.norm(dim=-1))
|
304 |
+
|
305 |
+
|
306 |
+
class Variance(Stat):
|
307 |
+
"""
|
308 |
+
Running computation of mean and variance. Use this when you just need
|
309 |
+
basic stats without covariance.
|
310 |
+
"""
|
311 |
+
|
312 |
+
def __init__(self, state=None):
|
313 |
+
if state is not None:
|
314 |
+
return super().__init__(state)
|
315 |
+
self.count = 0
|
316 |
+
self.batchcount = 0
|
317 |
+
self._mean = None
|
318 |
+
self.v_cmom2 = None
|
319 |
+
self.data_shape = None
|
320 |
+
|
321 |
+
def add(self, a):
|
322 |
+
a = self._normalize_add_shape(a)
|
323 |
+
if len(a) == 0:
|
324 |
+
return
|
325 |
+
batch_count = a.shape[0]
|
326 |
+
batch_mean = a.sum(0) / batch_count
|
327 |
+
centered = a - batch_mean
|
328 |
+
self.batchcount += 1
|
329 |
+
# Initial batch.
|
330 |
+
if self._mean is None:
|
331 |
+
self.count = batch_count
|
332 |
+
self._mean = batch_mean
|
333 |
+
self.v_cmom2 = centered.pow(2).sum(0)
|
334 |
+
return
|
335 |
+
# Update a batch using Chan-style update for numerical stability.
|
336 |
+
oldcount = self.count
|
337 |
+
self.count += batch_count
|
338 |
+
new_frac = float(batch_count) / self.count
|
339 |
+
# Update the mean according to the batch deviation from the old mean.
|
340 |
+
delta = batch_mean.sub_(self._mean).mul_(new_frac)
|
341 |
+
self._mean.add_(delta)
|
342 |
+
# Update the variance using the batch deviation
|
343 |
+
self.v_cmom2.add_(centered.pow(2).sum(0))
|
344 |
+
self.v_cmom2.add_(delta.pow_(2).mul_(new_frac * oldcount))
|
345 |
+
|
346 |
+
def size(self):
|
347 |
+
return self.count
|
348 |
+
|
349 |
+
def mean(self):
|
350 |
+
return self._restore_result_shape(self._mean)
|
351 |
+
|
352 |
+
def variance(self, unbiased=True):
|
353 |
+
return self._restore_result_shape(
|
354 |
+
self.v_cmom2 / (self.count - (1 if unbiased else 0))
|
355 |
+
)
|
356 |
+
|
357 |
+
def stdev(self, unbiased=True):
|
358 |
+
return self.variance(unbiased=unbiased).sqrt()
|
359 |
+
|
360 |
+
def to_(self, device):
|
361 |
+
if self._mean is not None:
|
362 |
+
self._mean = self._mean.to(device)
|
363 |
+
if self.v_cmom2 is not None:
|
364 |
+
self.v_cmom2 = self.v_cmom2.to(device)
|
365 |
+
|
366 |
+
def load_state_dict(self, state):
|
367 |
+
self.count = state["count"]
|
368 |
+
self.batchcount = state["batchcount"]
|
369 |
+
self._mean = torch.from_numpy(state["mean"])
|
370 |
+
self.v_cmom2 = torch.from_numpy(state["cmom2"])
|
371 |
+
self.data_shape = (
|
372 |
+
None if state["data_shape"] is None else tuple(state["data_shape"])
|
373 |
+
)
|
374 |
+
|
375 |
+
def state_dict(self):
|
376 |
+
return dict(
|
377 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
378 |
+
count=self.count,
|
379 |
+
data_shape=self.data_shape and tuple(self.data_shape),
|
380 |
+
batchcount=self.batchcount,
|
381 |
+
mean=self._mean.cpu().numpy(),
|
382 |
+
cmom2=self.v_cmom2.cpu().numpy(),
|
383 |
+
)
|
384 |
+
|
385 |
+
|
386 |
+
class Covariance(Stat):
|
387 |
+
"""
|
388 |
+
Running computation. Use this when the entire covariance matrix is needed,
|
389 |
+
and when the whole covariance matrix fits in the GPU.
|
390 |
+
|
391 |
+
Chan-style numerically stable update of mean and full covariance matrix.
|
392 |
+
Chan, Golub. LeVeque. 1983. http://www.jstor.org/stable/2683386
|
393 |
+
"""
|
394 |
+
|
395 |
+
def __init__(self, state=None):
|
396 |
+
if state is not None:
|
397 |
+
return super().__init__(state)
|
398 |
+
self.count = 0
|
399 |
+
self._mean = None
|
400 |
+
self.cmom2 = None
|
401 |
+
self.data_shape = None
|
402 |
+
|
403 |
+
def add(self, a):
|
404 |
+
a = self._normalize_add_shape(a)
|
405 |
+
if len(a) == 0:
|
406 |
+
return
|
407 |
+
batch_count = a.shape[0]
|
408 |
+
# Initial batch.
|
409 |
+
if self._mean is None:
|
410 |
+
self.count = batch_count
|
411 |
+
self._mean = a.sum(0) / batch_count
|
412 |
+
centered = a - self._mean
|
413 |
+
self.cmom2 = centered.t().mm(centered)
|
414 |
+
return
|
415 |
+
# Update a batch using Chan-style update for numerical stability.
|
416 |
+
self.count += batch_count
|
417 |
+
# Update the mean according to the batch deviation from the old mean.
|
418 |
+
delta = a - self._mean
|
419 |
+
self._mean.add_(delta.sum(0) / self.count)
|
420 |
+
delta2 = a - self._mean
|
421 |
+
# Update the variance using the batch deviation
|
422 |
+
self.cmom2.addmm_(mat1=delta.t(), mat2=delta2)
|
423 |
+
|
424 |
+
def to_(self, device):
|
425 |
+
if self._mean is not None:
|
426 |
+
self._mean = self._mean.to(device)
|
427 |
+
if self.cmom2 is not None:
|
428 |
+
self.cmom2 = self.cmom2.to(device)
|
429 |
+
|
430 |
+
def mean(self):
|
431 |
+
return self._restore_result_shape(self._mean)
|
432 |
+
|
433 |
+
def covariance(self, unbiased=True):
|
434 |
+
return self._restore_result_shape(
|
435 |
+
self.cmom2 / (self.count - (1 if unbiased else 0))
|
436 |
+
)
|
437 |
+
|
438 |
+
def correlation(self, unbiased=True):
|
439 |
+
cov = self.cmom2 / (self.count - (1 if unbiased else 0))
|
440 |
+
rstdev = cov.diag().sqrt().reciprocal()
|
441 |
+
return self._restore_result_shape(rstdev[:, None] * cov * rstdev[None, :])
|
442 |
+
|
443 |
+
def variance(self, unbiased=True):
|
444 |
+
return self._restore_result_shape(
|
445 |
+
self.cmom2.diag() / (self.count - (1 if unbiased else 0))
|
446 |
+
)
|
447 |
+
|
448 |
+
def stdev(self, unbiased=True):
|
449 |
+
return self.variance(unbiased=unbiased).sqrt()
|
450 |
+
|
451 |
+
def state_dict(self):
|
452 |
+
return dict(
|
453 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
454 |
+
count=self.count,
|
455 |
+
data_shape=self.data_shape and tuple(self.data_shape),
|
456 |
+
mean=self._mean.cpu().numpy(),
|
457 |
+
cmom2=self.cmom2.cpu().numpy(),
|
458 |
+
)
|
459 |
+
|
460 |
+
def load_state_dict(self, state):
|
461 |
+
self.count = state["count"]
|
462 |
+
self._mean = torch.from_numpy(state["mean"])
|
463 |
+
self.cmom2 = torch.from_numpy(state["cmom2"])
|
464 |
+
self.data_shape = (
|
465 |
+
None if state["data_shape"] is None else tuple(state["data_shape"])
|
466 |
+
)
|
467 |
+
|
468 |
+
|
469 |
+
class SecondMoment(Stat):
|
470 |
+
"""
|
471 |
+
Running computation. Use this when the entire non-centered 2nd-moment
|
472 |
+
'covariance-like' matrix is needed, and when the whole matrix fits
|
473 |
+
in the GPU.
|
474 |
+
"""
|
475 |
+
|
476 |
+
def __init__(self, split_batch=True, state=None):
|
477 |
+
if state is not None:
|
478 |
+
return super().__init__(state)
|
479 |
+
self.count = 0
|
480 |
+
self.mom2 = None
|
481 |
+
self.split_batch = split_batch
|
482 |
+
|
483 |
+
def add(self, a):
|
484 |
+
a = self._normalize_add_shape(a)
|
485 |
+
if len(a) == 0:
|
486 |
+
return
|
487 |
+
# Initial batch reveals the shape of the data.
|
488 |
+
if self.count == 0:
|
489 |
+
self.mom2 = a.new(a.shape[1], a.shape[1]).zero_()
|
490 |
+
batch_count = a.shape[0]
|
491 |
+
# Update the covariance using the batch deviation
|
492 |
+
self.count += batch_count
|
493 |
+
self.mom2 += a.t().mm(a)
|
494 |
+
|
495 |
+
def to_(self, device):
|
496 |
+
if self.mom2 is not None:
|
497 |
+
self.mom2 = self.mom2.to(device)
|
498 |
+
|
499 |
+
def moment(self):
|
500 |
+
return self.mom2 / self.count
|
501 |
+
|
502 |
+
def state_dict(self):
|
503 |
+
return dict(
|
504 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
505 |
+
count=self.count,
|
506 |
+
mom2=self.mom2.cpu().numpy(),
|
507 |
+
)
|
508 |
+
|
509 |
+
def load_state_dict(self, state):
|
510 |
+
self.count = int(state["count"])
|
511 |
+
self.mom2 = torch.from_numpy(state["mom2"])
|
512 |
+
|
513 |
+
|
514 |
+
class Bincount(Stat):
|
515 |
+
"""
|
516 |
+
Running bincount. The counted array should be an integer type with
|
517 |
+
non-negative integers.
|
518 |
+
"""
|
519 |
+
|
520 |
+
def __init__(self, state=None):
|
521 |
+
if state is not None:
|
522 |
+
return super().__init__(state)
|
523 |
+
self.count = 0
|
524 |
+
self._bincount = None
|
525 |
+
|
526 |
+
def add(self, a, size=None):
|
527 |
+
a = a.view(-1)
|
528 |
+
bincount = a.bincount()
|
529 |
+
if self._bincount is None:
|
530 |
+
self._bincount = bincount
|
531 |
+
elif len(self._bincount) < len(bincount):
|
532 |
+
bincount[: len(self._bincount)] += self._bincount
|
533 |
+
self._bincount = bincount
|
534 |
+
else:
|
535 |
+
self._bincount[: len(bincount)] += bincount
|
536 |
+
if size is None:
|
537 |
+
self.count += len(a)
|
538 |
+
else:
|
539 |
+
self.count += size
|
540 |
+
|
541 |
+
def to_(self, device):
|
542 |
+
self._bincount = self._bincount.to(device)
|
543 |
+
|
544 |
+
def size(self):
|
545 |
+
return self.count
|
546 |
+
|
547 |
+
def bincount(self):
|
548 |
+
return self._bincount
|
549 |
+
|
550 |
+
def state_dict(self):
|
551 |
+
return dict(
|
552 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
553 |
+
count=self.count,
|
554 |
+
bincount=self._bincount.cpu().numpy(),
|
555 |
+
)
|
556 |
+
|
557 |
+
def load_state_dict(self, dic):
|
558 |
+
self.count = int(dic["count"])
|
559 |
+
self._bincount = torch.from_numpy(dic["bincount"])
|
560 |
+
|
561 |
+
|
562 |
+
class CrossCovariance(Stat):
|
563 |
+
"""
|
564 |
+
Covariance. Use this when an off-diagonal block of the covariance
|
565 |
+
matrix is needed (e.g., when the whole covariance matrix does
|
566 |
+
not fit in the GPU, this could use a quarter of the memory).
|
567 |
+
|
568 |
+
Chan-style numerically stable update of mean and full covariance matrix.
|
569 |
+
Chan, Golub. LeVeque. 1983. http://www.jstor.org/stable/2683386
|
570 |
+
"""
|
571 |
+
|
572 |
+
def __init__(self, split_batch=True, state=None):
|
573 |
+
if state is not None:
|
574 |
+
return super().__init__(state)
|
575 |
+
self.count = 0
|
576 |
+
self._mean = None
|
577 |
+
self.cmom2 = None
|
578 |
+
self.v_cmom2 = None
|
579 |
+
self.split_batch = split_batch
|
580 |
+
|
581 |
+
def add(self, a, b):
|
582 |
+
if len(a.shape) == 1:
|
583 |
+
a = a[None, :]
|
584 |
+
b = b[None, :]
|
585 |
+
assert a.shape[0] == b.shape[0]
|
586 |
+
if len(a.shape) > 2:
|
587 |
+
a, b = [
|
588 |
+
d.view(d.shape[0], d.shape[1], -1)
|
589 |
+
.permute(0, 2, 1)
|
590 |
+
.reshape(-1, d.shape[1])
|
591 |
+
for d in [a, b]
|
592 |
+
]
|
593 |
+
batch_count = a.shape[0]
|
594 |
+
# Initial batch.
|
595 |
+
if self._mean is None:
|
596 |
+
self.count = batch_count
|
597 |
+
self._mean = [d.sum(0) / batch_count for d in [a, b]]
|
598 |
+
centered = [d - bm for d, bm in zip([a, b], self._mean)]
|
599 |
+
self.v_cmom2 = [c.pow(2).sum(0) for c in centered]
|
600 |
+
self.cmom2 = centered[0].t().mm(centered[1])
|
601 |
+
return
|
602 |
+
# Update a batch using Chan-style update for numerical stability.
|
603 |
+
self.count += batch_count
|
604 |
+
# Update the mean according to the batch deviation from the old mean.
|
605 |
+
delta = [(d - bm) for d, bm in zip([a, b], self._mean)]
|
606 |
+
for m, d in zip(self._mean, delta):
|
607 |
+
m.add_(d.sum(0) / self.count)
|
608 |
+
delta2 = [(d - bm) for d, bm in zip([a, b], self._mean)]
|
609 |
+
# Update the cross-covariance using the batch deviation
|
610 |
+
self.cmom2.addmm_(mat1=delta[0].t(), mat2=delta2[1])
|
611 |
+
# Update the variance using the batch deviation
|
612 |
+
for vc2, d, d2 in zip(self.v_cmom2, delta, delta2):
|
613 |
+
vc2.add_((d * d2).sum(0))
|
614 |
+
|
615 |
+
def mean(self):
|
616 |
+
return self._mean
|
617 |
+
|
618 |
+
def variance(self, unbiased=True):
|
619 |
+
return [vc2 / (self.count - (1 if unbiased else 0)) for vc2 in self.v_cmom2]
|
620 |
+
|
621 |
+
def stdev(self, unbiased=True):
|
622 |
+
return [v.sqrt() for v in self.variance(unbiased=unbiased)]
|
623 |
+
|
624 |
+
def covariance(self, unbiased=True):
|
625 |
+
return self.cmom2 / (self.count - (1 if unbiased else 0))
|
626 |
+
|
627 |
+
def correlation(self):
|
628 |
+
covariance = self.covariance(unbiased=False)
|
629 |
+
rstdev = [s.reciprocal() for s in self.stdev(unbiased=False)]
|
630 |
+
cor = rstdev[0][:, None] * covariance * rstdev[1][None, :]
|
631 |
+
# Remove NaNs
|
632 |
+
cor[torch.isnan(cor)] = 0
|
633 |
+
return cor
|
634 |
+
|
635 |
+
def to_(self, device):
|
636 |
+
self._mean = [m.to(device) for m in self._mean]
|
637 |
+
self.v_cmom2 = [vcs.to(device) for vcs in self.v_cmom2]
|
638 |
+
self.cmom2 = self.cmom2.to(device)
|
639 |
+
|
640 |
+
def state_dict(self):
|
641 |
+
return dict(
|
642 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
643 |
+
count=self.count,
|
644 |
+
mean_a=self._mean[0].cpu().numpy(),
|
645 |
+
mean_b=self._mean[1].cpu().numpy(),
|
646 |
+
cmom2_a=self.v_cmom2[0].cpu().numpy(),
|
647 |
+
cmom2_b=self.v_cmom2[1].cpu().numpy(),
|
648 |
+
cmom2=self.cmom2.cpu().numpy(),
|
649 |
+
)
|
650 |
+
|
651 |
+
def load_state_dict(self, state):
|
652 |
+
self.count = int(state["count"])
|
653 |
+
self._mean = [torch.from_numpy(state[f"mean_{k}"]) for k in "ab"]
|
654 |
+
self.v_cmom2 = [torch.from_numpy(state[f"cmom2_{k}"]) for k in "ab"]
|
655 |
+
self.cmom2 = torch.from_numpy(state["cmom2"])
|
656 |
+
|
657 |
+
|
658 |
+
def _float_from_bool(a):
|
659 |
+
"""
|
660 |
+
Since pytorch only supports matrix multiplication on float,
|
661 |
+
IoU computations are done using floating point types.
|
662 |
+
|
663 |
+
This function binarizes the input (positive to True and
|
664 |
+
nonpositive to False), and converts from bool to float.
|
665 |
+
If the data is already a floating-point type, it leaves
|
666 |
+
it keeps the same type; otherwise it uses float.
|
667 |
+
"""
|
668 |
+
if a.dtype == torch.bool:
|
669 |
+
return a.float()
|
670 |
+
if a.dtype.is_floating_point:
|
671 |
+
return a.sign().clamp_(0)
|
672 |
+
return (a > 0).float()
|
673 |
+
|
674 |
+
|
675 |
+
class IoU(Stat):
|
676 |
+
"""
|
677 |
+
Running computation of intersections and unions of all features.
|
678 |
+
"""
|
679 |
+
|
680 |
+
def __init__(self, state=None):
|
681 |
+
if state is not None:
|
682 |
+
return super().__init__(state)
|
683 |
+
self.count = 0
|
684 |
+
self._intersection = None
|
685 |
+
|
686 |
+
def add(self, a):
|
687 |
+
assert len(a.shape) == 2
|
688 |
+
a = _float_from_bool(a)
|
689 |
+
if self._intersection is None:
|
690 |
+
self._intersection = torch.mm(a.t(), a)
|
691 |
+
else:
|
692 |
+
self._intersection.addmm_(a.t(), a)
|
693 |
+
self.count += len(a)
|
694 |
+
|
695 |
+
def size(self):
|
696 |
+
return self.count
|
697 |
+
|
698 |
+
def intersection(self):
|
699 |
+
return self._intersection
|
700 |
+
|
701 |
+
def union(self):
|
702 |
+
total = self._intersection.diagonal(0)
|
703 |
+
return total[:, None] + total[None, :] - self._intersection
|
704 |
+
|
705 |
+
def iou(self):
|
706 |
+
return self.intersection() / (self.union() + 1e-20)
|
707 |
+
|
708 |
+
def to_(self, _device):
|
709 |
+
self._intersection = self._intersection.to(_device)
|
710 |
+
|
711 |
+
def state_dict(self):
|
712 |
+
return dict(
|
713 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
714 |
+
count=self.count,
|
715 |
+
intersection=self._intersection.cpu().numpy(),
|
716 |
+
)
|
717 |
+
|
718 |
+
def load_state_dict(self, state):
|
719 |
+
self.count = int(state["count"])
|
720 |
+
self._intersection = torch.tensor(state["intersection"])
|
721 |
+
|
722 |
+
|
723 |
+
class CrossIoU(Stat):
|
724 |
+
"""
|
725 |
+
Running computation of intersections and unions of two binary vectors.
|
726 |
+
"""
|
727 |
+
|
728 |
+
def __init__(self, state=None):
|
729 |
+
if state is not None:
|
730 |
+
return super().__init__(state)
|
731 |
+
self.count = 0
|
732 |
+
self._intersection = None
|
733 |
+
self.total_a = None
|
734 |
+
self.total_b = None
|
735 |
+
|
736 |
+
def add(self, a, b):
|
737 |
+
assert len(a.shape) == 2 and len(b.shape) == 2
|
738 |
+
assert len(a) == len(b), f"{len(a)} vs {len(b)}"
|
739 |
+
a = _float_from_bool(a) # CUDA only supports mm on float...
|
740 |
+
b = _float_from_bool(b) # otherwise we would use integers.
|
741 |
+
intersection = torch.mm(a.t(), b)
|
742 |
+
asum = a.sum(0)
|
743 |
+
bsum = b.sum(0)
|
744 |
+
if self._intersection is None:
|
745 |
+
self._intersection = intersection
|
746 |
+
self.total_a = asum
|
747 |
+
self.total_b = bsum
|
748 |
+
else:
|
749 |
+
self._intersection += intersection
|
750 |
+
self.total_a += asum
|
751 |
+
self.total_b += bsum
|
752 |
+
self.count += len(a)
|
753 |
+
|
754 |
+
def size(self):
|
755 |
+
return self.count
|
756 |
+
|
757 |
+
def intersection(self):
|
758 |
+
return self._intersection
|
759 |
+
|
760 |
+
def union(self):
|
761 |
+
return self.total_a[:, None] + self.total_b[None, :] - self._intersection
|
762 |
+
|
763 |
+
def iou(self):
|
764 |
+
return self.intersection() / (self.union() + 1e-20)
|
765 |
+
|
766 |
+
def to_(self, _device):
|
767 |
+
self.total_a = self.total_a.to(_device)
|
768 |
+
self.total_b = self.total_b.to(_device)
|
769 |
+
self._intersection = self._intersection.to(_device)
|
770 |
+
|
771 |
+
def state_dict(self):
|
772 |
+
return dict(
|
773 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
774 |
+
count=self.count,
|
775 |
+
total_a=self.total_a.cpu().numpy(),
|
776 |
+
total_b=self.total_b.cpu().numpy(),
|
777 |
+
intersection=self._intersection.cpu().numpy(),
|
778 |
+
)
|
779 |
+
|
780 |
+
def load_state_dict(self, state):
|
781 |
+
self.count = int(state["count"])
|
782 |
+
self.total_a = torch.tensor(state["total_a"])
|
783 |
+
self.total_b = torch.tensor(state["total_b"])
|
784 |
+
self._intersection = torch.tensor(state["intersection"])
|
785 |
+
|
786 |
+
|
787 |
+
class Quantile(Stat):
|
788 |
+
"""
|
789 |
+
Streaming randomized quantile computation for torch.
|
790 |
+
|
791 |
+
Add any amount of data repeatedly via add(data). At any time,
|
792 |
+
quantile estimates be read out using quantile(q).
|
793 |
+
|
794 |
+
Implemented as a sorted sample that retains at least r samples
|
795 |
+
(by default r = 3072); the number of retained samples will grow to
|
796 |
+
a finite ceiling as the data is accumulated. Accuracy scales according
|
797 |
+
to r: the default is to set resolution to be accurate to better than about
|
798 |
+
0.1%, while limiting storage to about 50,000 samples.
|
799 |
+
|
800 |
+
Good for computing quantiles of huge data without using much memory.
|
801 |
+
Works well on arbitrary data with probability near 1.
|
802 |
+
|
803 |
+
Based on the optimal KLL quantile algorithm by Karnin, Lang, and Liberty
|
804 |
+
from FOCS 2016. http://ieee-focs.org/FOCS-2016-Papers/3933a071.pdf
|
805 |
+
"""
|
806 |
+
|
807 |
+
def __init__(self, r=3 * 1024, buffersize=None, seed=None, state=None):
|
808 |
+
if state is not None:
|
809 |
+
return super().__init__(state)
|
810 |
+
self.depth = None
|
811 |
+
self.dtype = None
|
812 |
+
self.device = None
|
813 |
+
resolution = r * 2 # sample array is at least half full before discard
|
814 |
+
self.resolution = resolution
|
815 |
+
# Default buffersize: 128 samples (and smaller than resolution).
|
816 |
+
if buffersize is None:
|
817 |
+
buffersize = min(128, (resolution + 7) // 8)
|
818 |
+
self.buffersize = buffersize
|
819 |
+
self.samplerate = 1.0
|
820 |
+
self.data = None
|
821 |
+
self.firstfree = [0]
|
822 |
+
self.randbits = torch.ByteTensor(resolution)
|
823 |
+
self.currentbit = len(self.randbits) - 1
|
824 |
+
self.extremes = None
|
825 |
+
self.count = 0
|
826 |
+
self.batchcount = 0
|
827 |
+
|
828 |
+
def size(self):
|
829 |
+
return self.count
|
830 |
+
|
831 |
+
def _lazy_init(self, incoming):
|
832 |
+
self.depth = incoming.shape[1]
|
833 |
+
self.dtype = incoming.dtype
|
834 |
+
self.device = incoming.device
|
835 |
+
self.data = [
|
836 |
+
torch.zeros(
|
837 |
+
self.depth, self.resolution, dtype=self.dtype, device=self.device
|
838 |
+
)
|
839 |
+
]
|
840 |
+
self.extremes = torch.zeros(self.depth, 2, dtype=self.dtype, device=self.device)
|
841 |
+
self.extremes[:, 0] = float("inf")
|
842 |
+
self.extremes[:, -1] = -float("inf")
|
843 |
+
|
844 |
+
def to_(self, device):
|
845 |
+
"""Switches internal storage to specified device."""
|
846 |
+
if device != self.device:
|
847 |
+
old_data = self.data
|
848 |
+
old_extremes = self.extremes
|
849 |
+
self.data = [d.to(device) for d in self.data]
|
850 |
+
self.extremes = self.extremes.to(device)
|
851 |
+
self.device = self.extremes.device
|
852 |
+
del old_data
|
853 |
+
del old_extremes
|
854 |
+
|
855 |
+
def add(self, incoming):
|
856 |
+
if self.depth is None:
|
857 |
+
self._lazy_init(incoming)
|
858 |
+
assert len(incoming.shape) == 2
|
859 |
+
assert incoming.shape[1] == self.depth, (incoming.shape[1], self.depth)
|
860 |
+
self.count += incoming.shape[0]
|
861 |
+
self.batchcount += 1
|
862 |
+
# Convert to a flat torch array.
|
863 |
+
if self.samplerate >= 1.0:
|
864 |
+
self._add_every(incoming)
|
865 |
+
return
|
866 |
+
# If we are sampling, then subsample a large chunk at a time.
|
867 |
+
self._scan_extremes(incoming)
|
868 |
+
chunksize = int(math.ceil(self.buffersize / self.samplerate))
|
869 |
+
for index in range(0, len(incoming), chunksize):
|
870 |
+
batch = incoming[index : index + chunksize]
|
871 |
+
sample = sample_portion(batch, self.samplerate)
|
872 |
+
if len(sample):
|
873 |
+
self._add_every(sample)
|
874 |
+
|
875 |
+
def _add_every(self, incoming):
|
876 |
+
supplied = len(incoming)
|
877 |
+
index = 0
|
878 |
+
while index < supplied:
|
879 |
+
ff = self.firstfree[0]
|
880 |
+
available = self.data[0].shape[1] - ff
|
881 |
+
if available == 0:
|
882 |
+
if not self._shift():
|
883 |
+
# If we shifted by subsampling, then subsample.
|
884 |
+
incoming = incoming[index:]
|
885 |
+
if self.samplerate >= 0.5:
|
886 |
+
# First time sampling - the data source is very large.
|
887 |
+
self._scan_extremes(incoming)
|
888 |
+
incoming = sample_portion(incoming, self.samplerate)
|
889 |
+
index = 0
|
890 |
+
supplied = len(incoming)
|
891 |
+
ff = self.firstfree[0]
|
892 |
+
available = self.data[0].shape[1] - ff
|
893 |
+
copycount = min(available, supplied - index)
|
894 |
+
self.data[0][:, ff : ff + copycount] = torch.t(
|
895 |
+
incoming[index : index + copycount, :]
|
896 |
+
)
|
897 |
+
self.firstfree[0] += copycount
|
898 |
+
index += copycount
|
899 |
+
|
900 |
+
def _shift(self):
|
901 |
+
index = 0
|
902 |
+
# If remaining space at the current layer is less than half prev
|
903 |
+
# buffer size (rounding up), then we need to shift it up to ensure
|
904 |
+
# enough space for future shifting.
|
905 |
+
while self.data[index].shape[1] - self.firstfree[index] < (
|
906 |
+
-(-self.data[index - 1].shape[1] // 2) if index else 1
|
907 |
+
):
|
908 |
+
if index + 1 >= len(self.data):
|
909 |
+
return self._expand()
|
910 |
+
data = self.data[index][:, 0 : self.firstfree[index]]
|
911 |
+
data = data.sort()[0]
|
912 |
+
if index == 0 and self.samplerate >= 1.0:
|
913 |
+
self._update_extremes(data[:, 0], data[:, -1])
|
914 |
+
offset = self._randbit()
|
915 |
+
position = self.firstfree[index + 1]
|
916 |
+
subset = data[:, offset::2]
|
917 |
+
self.data[index + 1][:, position : position + subset.shape[1]] = subset
|
918 |
+
self.firstfree[index] = 0
|
919 |
+
self.firstfree[index + 1] += subset.shape[1]
|
920 |
+
index += 1
|
921 |
+
return True
|
922 |
+
|
923 |
+
def _scan_extremes(self, incoming):
|
924 |
+
# When sampling, we need to scan every item still to get extremes
|
925 |
+
self._update_extremes(
|
926 |
+
torch.min(incoming, dim=0)[0], torch.max(incoming, dim=0)[0]
|
927 |
+
)
|
928 |
+
|
929 |
+
def _update_extremes(self, minr, maxr):
|
930 |
+
self.extremes[:, 0] = torch.min(
|
931 |
+
torch.stack([self.extremes[:, 0], minr]), dim=0
|
932 |
+
)[0]
|
933 |
+
self.extremes[:, -1] = torch.max(
|
934 |
+
torch.stack([self.extremes[:, -1], maxr]), dim=0
|
935 |
+
)[0]
|
936 |
+
|
937 |
+
def _randbit(self):
|
938 |
+
self.currentbit += 1
|
939 |
+
if self.currentbit >= len(self.randbits):
|
940 |
+
self.randbits.random_(to=2)
|
941 |
+
self.currentbit = 0
|
942 |
+
return self.randbits[self.currentbit]
|
943 |
+
|
944 |
+
def state_dict(self):
|
945 |
+
state = dict(
|
946 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
947 |
+
resolution=self.resolution,
|
948 |
+
depth=self.depth,
|
949 |
+
buffersize=self.buffersize,
|
950 |
+
samplerate=self.samplerate,
|
951 |
+
sizes=numpy.array([d.shape[1] for d in self.data]),
|
952 |
+
extremes=self.extremes.cpu().detach().numpy(),
|
953 |
+
size=self.count,
|
954 |
+
batchcount=self.batchcount,
|
955 |
+
)
|
956 |
+
for i, (d, f) in enumerate(zip(self.data, self.firstfree)):
|
957 |
+
state[f"data.{i}"] = d.cpu().detach().numpy()[:, :f].T
|
958 |
+
return state
|
959 |
+
|
960 |
+
def load_state_dict(self, state):
|
961 |
+
self.resolution = int(state["resolution"])
|
962 |
+
self.randbits = torch.ByteTensor(self.resolution)
|
963 |
+
self.currentbit = len(self.randbits) - 1
|
964 |
+
self.depth = int(state["depth"])
|
965 |
+
self.buffersize = int(state["buffersize"])
|
966 |
+
self.samplerate = float(state["samplerate"])
|
967 |
+
firstfree = []
|
968 |
+
buffers = []
|
969 |
+
for i, s in enumerate(state["sizes"]):
|
970 |
+
d = state[f"data.{i}"]
|
971 |
+
firstfree.append(d.shape[0])
|
972 |
+
buf = numpy.zeros((d.shape[1], s), dtype=d.dtype)
|
973 |
+
buf[:, : d.shape[0]] = d.T
|
974 |
+
buffers.append(torch.from_numpy(buf))
|
975 |
+
self.firstfree = firstfree
|
976 |
+
self.data = buffers
|
977 |
+
self.extremes = torch.from_numpy((state["extremes"]))
|
978 |
+
self.count = int(state["size"])
|
979 |
+
self.batchcount = int(state.get("batchcount", 0))
|
980 |
+
self.dtype = self.extremes.dtype
|
981 |
+
self.device = self.extremes.device
|
982 |
+
|
983 |
+
def min(self):
|
984 |
+
return self.minmax()[0]
|
985 |
+
|
986 |
+
def max(self):
|
987 |
+
return self.minmax()[-1]
|
988 |
+
|
989 |
+
def minmax(self):
|
990 |
+
if self.firstfree[0]:
|
991 |
+
self._scan_extremes(self.data[0][:, : self.firstfree[0]].t())
|
992 |
+
return self.extremes.clone()
|
993 |
+
|
994 |
+
def median(self):
|
995 |
+
return self.quantiles(0.5)
|
996 |
+
|
997 |
+
def mean(self):
|
998 |
+
return self.integrate(lambda x: x) / self.count
|
999 |
+
|
1000 |
+
def variance(self, unbiased=True):
|
1001 |
+
mean = self.mean()[:, None]
|
1002 |
+
return self.integrate(lambda x: (x - mean).pow(2)) / (
|
1003 |
+
self.count - (1 if unbiased else 0)
|
1004 |
+
)
|
1005 |
+
|
1006 |
+
def stdev(self, unbiased=True):
|
1007 |
+
return self.variance(unbiased=unbiased).sqrt()
|
1008 |
+
|
1009 |
+
def _expand(self):
|
1010 |
+
cap = self._next_capacity()
|
1011 |
+
if cap > 0:
|
1012 |
+
# First, make a new layer of the proper capacity.
|
1013 |
+
self.data.insert(
|
1014 |
+
0, torch.zeros(self.depth, cap, dtype=self.dtype, device=self.device)
|
1015 |
+
)
|
1016 |
+
self.firstfree.insert(0, 0)
|
1017 |
+
else:
|
1018 |
+
# Unless we're so big we are just subsampling.
|
1019 |
+
assert self.firstfree[0] == 0
|
1020 |
+
self.samplerate *= 0.5
|
1021 |
+
for index in range(1, len(self.data)):
|
1022 |
+
# Scan for existing data that needs to be moved down a level.
|
1023 |
+
amount = self.firstfree[index]
|
1024 |
+
if amount == 0:
|
1025 |
+
continue
|
1026 |
+
position = self.firstfree[index - 1]
|
1027 |
+
# Move data down if it would leave enough empty space there
|
1028 |
+
# This is the key invariant: enough empty space to fit half
|
1029 |
+
# of the previous level's buffer size (rounding up)
|
1030 |
+
if self.data[index - 1].shape[1] - (amount + position) >= (
|
1031 |
+
-(-self.data[index - 2].shape[1] // 2) if (index - 1) else 1
|
1032 |
+
):
|
1033 |
+
self.data[index - 1][:, position : position + amount] = self.data[
|
1034 |
+
index
|
1035 |
+
][:, :amount]
|
1036 |
+
self.firstfree[index - 1] += amount
|
1037 |
+
self.firstfree[index] = 0
|
1038 |
+
else:
|
1039 |
+
# Scrunch the data if it would not.
|
1040 |
+
data = self.data[index][:, :amount]
|
1041 |
+
data = data.sort()[0]
|
1042 |
+
if index == 1:
|
1043 |
+
self._update_extremes(data[:, 0], data[:, -1])
|
1044 |
+
offset = self._randbit()
|
1045 |
+
scrunched = data[:, offset::2]
|
1046 |
+
self.data[index][:, : scrunched.shape[1]] = scrunched
|
1047 |
+
self.firstfree[index] = scrunched.shape[1]
|
1048 |
+
return cap > 0
|
1049 |
+
|
1050 |
+
def _next_capacity(self):
|
1051 |
+
cap = int(math.ceil(self.resolution * (0.67 ** len(self.data))))
|
1052 |
+
if cap < 2:
|
1053 |
+
return 0
|
1054 |
+
# Round up to the nearest multiple of 8 for better GPU alignment.
|
1055 |
+
cap = -8 * (-cap // 8)
|
1056 |
+
return max(self.buffersize, cap)
|
1057 |
+
|
1058 |
+
def _weighted_summary(self, sort=True):
|
1059 |
+
if self.firstfree[0]:
|
1060 |
+
self._scan_extremes(self.data[0][:, : self.firstfree[0]].t())
|
1061 |
+
size = sum(self.firstfree)
|
1062 |
+
weights = torch.FloatTensor(size) # Floating point
|
1063 |
+
summary = torch.zeros(self.depth, size, dtype=self.dtype, device=self.device)
|
1064 |
+
index = 0
|
1065 |
+
for level, ff in enumerate(self.firstfree):
|
1066 |
+
if ff == 0:
|
1067 |
+
continue
|
1068 |
+
summary[:, index : index + ff] = self.data[level][:, :ff]
|
1069 |
+
weights[index : index + ff] = 2.0**level
|
1070 |
+
index += ff
|
1071 |
+
assert index == summary.shape[1]
|
1072 |
+
if sort:
|
1073 |
+
summary, order = torch.sort(summary, dim=-1)
|
1074 |
+
weights = weights[order.view(-1).cpu()].view(order.shape)
|
1075 |
+
summary = torch.cat(
|
1076 |
+
[self.extremes[:, :1], summary, self.extremes[:, 1:]], dim=-1
|
1077 |
+
)
|
1078 |
+
weights = torch.cat(
|
1079 |
+
[
|
1080 |
+
torch.zeros(weights.shape[0], 1),
|
1081 |
+
weights,
|
1082 |
+
torch.zeros(weights.shape[0], 1),
|
1083 |
+
],
|
1084 |
+
dim=-1,
|
1085 |
+
)
|
1086 |
+
return (summary, weights)
|
1087 |
+
|
1088 |
+
def quantiles(self, quantiles):
|
1089 |
+
if not hasattr(quantiles, "cpu"):
|
1090 |
+
quantiles = torch.tensor(quantiles)
|
1091 |
+
qshape = quantiles.shape
|
1092 |
+
if self.count == 0:
|
1093 |
+
return torch.full((self.depth,) + qshape, torch.nan)
|
1094 |
+
summary, weights = self._weighted_summary()
|
1095 |
+
cumweights = torch.cumsum(weights, dim=-1) - weights / 2
|
1096 |
+
cumweights /= torch.sum(weights, dim=-1, keepdim=True)
|
1097 |
+
result = torch.zeros(
|
1098 |
+
self.depth, quantiles.numel(), dtype=self.dtype, device=self.device
|
1099 |
+
)
|
1100 |
+
# numpy is needed for interpolation
|
1101 |
+
nq = quantiles.view(-1).cpu().detach().numpy()
|
1102 |
+
ncw = cumweights.cpu().detach().numpy()
|
1103 |
+
nsm = summary.cpu().detach().numpy()
|
1104 |
+
for d in range(self.depth):
|
1105 |
+
result[d] = torch.tensor(
|
1106 |
+
numpy.interp(nq, ncw[d], nsm[d]), dtype=self.dtype, device=self.device
|
1107 |
+
)
|
1108 |
+
return result.view((self.depth,) + qshape)
|
1109 |
+
|
1110 |
+
def integrate(self, fun):
|
1111 |
+
result = []
|
1112 |
+
for level, ff in enumerate(self.firstfree):
|
1113 |
+
if ff == 0:
|
1114 |
+
continue
|
1115 |
+
result.append(
|
1116 |
+
torch.sum(fun(self.data[level][:, :ff]) * (2.0**level), dim=-1)
|
1117 |
+
)
|
1118 |
+
if len(result) == 0:
|
1119 |
+
return None
|
1120 |
+
return torch.stack(result).sum(dim=0) / self.samplerate
|
1121 |
+
|
1122 |
+
def readout(self, count=1001):
|
1123 |
+
return self.quantiles(torch.linspace(0.0, 1.0, count))
|
1124 |
+
|
1125 |
+
def normalize(self, data):
|
1126 |
+
"""
|
1127 |
+
Given input data as taken from the training distirbution,
|
1128 |
+
normalizes every channel to reflect quantile values,
|
1129 |
+
uniformly distributed, within [0, 1].
|
1130 |
+
"""
|
1131 |
+
assert self.count > 0
|
1132 |
+
assert data.shape[0] == self.depth
|
1133 |
+
summary, weights = self._weighted_summary()
|
1134 |
+
cumweights = torch.cumsum(weights, dim=-1) - weights / 2
|
1135 |
+
cumweights /= torch.sum(weights, dim=-1, keepdim=True)
|
1136 |
+
result = torch.zeros_like(data).float()
|
1137 |
+
# numpy is needed for interpolation
|
1138 |
+
ndata = data.cpu().numpy().reshape((data.shape[0], -1))
|
1139 |
+
ncw = cumweights.cpu().numpy()
|
1140 |
+
nsm = summary.cpu().numpy()
|
1141 |
+
for d in range(self.depth):
|
1142 |
+
normed = torch.tensor(
|
1143 |
+
numpy.interp(ndata[d], nsm[d], ncw[d]),
|
1144 |
+
dtype=torch.float,
|
1145 |
+
device=data.device,
|
1146 |
+
).clamp_(0.0, 1.0)
|
1147 |
+
if len(data.shape) > 1:
|
1148 |
+
normed = normed.view(*(data.shape[1:]))
|
1149 |
+
result[d] = normed
|
1150 |
+
return result
|
1151 |
+
|
1152 |
+
|
1153 |
+
def sample_portion(vec, p=0.5):
|
1154 |
+
"""
|
1155 |
+
Subsamples a fraction (given by p) of the given batch. Used by
|
1156 |
+
Quantile when the data gets very very large.
|
1157 |
+
"""
|
1158 |
+
bits = torch.bernoulli(
|
1159 |
+
torch.zeros(vec.shape[0], dtype=torch.uint8, device=vec.device), p
|
1160 |
+
)
|
1161 |
+
return vec[bits]
|
1162 |
+
|
1163 |
+
|
1164 |
+
class TopK:
|
1165 |
+
"""
|
1166 |
+
A class to keep a running tally of the the top k values (and indexes)
|
1167 |
+
of any number of torch feature components. Will work on the GPU if
|
1168 |
+
the data is on the GPU. Tracks largest by default, but tracks smallest
|
1169 |
+
if largest=False is passed.
|
1170 |
+
|
1171 |
+
This version flattens all arrays to avoid crashes.
|
1172 |
+
"""
|
1173 |
+
|
1174 |
+
def __init__(self, k=100, largest=True, state=None):
|
1175 |
+
if state is not None:
|
1176 |
+
return super().__init__(state)
|
1177 |
+
self.k = k
|
1178 |
+
self.count = 0
|
1179 |
+
# This version flattens all data internally to 2-d tensors,
|
1180 |
+
# to avoid crashes with the current pytorch topk implementation.
|
1181 |
+
# The data is puffed back out to arbitrary tensor shapes on output.
|
1182 |
+
self.data_shape = None
|
1183 |
+
self.top_data = None
|
1184 |
+
self.top_index = None
|
1185 |
+
self.next = 0
|
1186 |
+
self.linear_index = 0
|
1187 |
+
self.perm = None
|
1188 |
+
self.largest = largest
|
1189 |
+
|
1190 |
+
def add(self, data, index=None):
|
1191 |
+
"""
|
1192 |
+
Adds a batch of data to be considered for the running top k.
|
1193 |
+
The zeroth dimension enumerates the observations. All other
|
1194 |
+
dimensions enumerate different features.
|
1195 |
+
"""
|
1196 |
+
if self.top_data is None:
|
1197 |
+
# Allocation: allocate a buffer of size 5*k, at least 10, for each.
|
1198 |
+
self.data_shape = data.shape[1:]
|
1199 |
+
feature_size = int(numpy.prod(self.data_shape))
|
1200 |
+
self.top_data = torch.zeros(
|
1201 |
+
feature_size, max(10, self.k * 5), out=data.new()
|
1202 |
+
)
|
1203 |
+
self.top_index = self.top_data.clone().long()
|
1204 |
+
self.linear_index = (
|
1205 |
+
0
|
1206 |
+
if len(data.shape) == 1
|
1207 |
+
else torch.arange(feature_size, out=self.top_index.new()).mul_(
|
1208 |
+
self.top_data.shape[-1]
|
1209 |
+
)[:, None]
|
1210 |
+
)
|
1211 |
+
size = data.shape[0]
|
1212 |
+
sk = min(size, self.k)
|
1213 |
+
if self.top_data.shape[-1] < self.next + sk:
|
1214 |
+
# Compression: if full, keep topk only.
|
1215 |
+
self.top_data[:, : self.k], self.top_index[:, : self.k] = self.topk(
|
1216 |
+
sorted=False, flat=True
|
1217 |
+
)
|
1218 |
+
self.next = self.k
|
1219 |
+
# Pick: copy the top sk of the next batch into the buffer.
|
1220 |
+
# Currently strided topk is slow. So we clone after transpose.
|
1221 |
+
# TODO: remove the clone() if it becomes faster.
|
1222 |
+
cdata = data.reshape(size, numpy.prod(data.shape[1:])).t().clone()
|
1223 |
+
td, ti = cdata.topk(sk, sorted=False, largest=self.largest)
|
1224 |
+
self.top_data[:, self.next : self.next + sk] = td
|
1225 |
+
if index is not None:
|
1226 |
+
ti = index[ti]
|
1227 |
+
else:
|
1228 |
+
ti = ti + self.count
|
1229 |
+
self.top_index[:, self.next : self.next + sk] = ti
|
1230 |
+
self.next += sk
|
1231 |
+
self.count += size
|
1232 |
+
|
1233 |
+
def size(self):
|
1234 |
+
return self.count
|
1235 |
+
|
1236 |
+
def topk(self, sorted=True, flat=False):
|
1237 |
+
"""
|
1238 |
+
Returns top k data items and indexes in each dimension,
|
1239 |
+
with channels in the first dimension and k in the last dimension.
|
1240 |
+
"""
|
1241 |
+
k = min(self.k, self.next)
|
1242 |
+
# bti are top indexes relative to buffer array.
|
1243 |
+
td, bti = self.top_data[:, : self.next].topk(
|
1244 |
+
k, sorted=sorted, largest=self.largest
|
1245 |
+
)
|
1246 |
+
# we want to report top indexes globally, which is ti.
|
1247 |
+
ti = self.top_index.view(-1)[(bti + self.linear_index).view(-1)].view(
|
1248 |
+
*bti.shape
|
1249 |
+
)
|
1250 |
+
if flat:
|
1251 |
+
return td, ti
|
1252 |
+
else:
|
1253 |
+
return (
|
1254 |
+
td.view(*(self.data_shape + (-1,))),
|
1255 |
+
ti.view(*(self.data_shape + (-1,))),
|
1256 |
+
)
|
1257 |
+
|
1258 |
+
def to_(self, device):
|
1259 |
+
if self.top_data is not None:
|
1260 |
+
self.top_data = self.top_data.to(device)
|
1261 |
+
if self.top_index is not None:
|
1262 |
+
self.top_index = self.top_index.to(device)
|
1263 |
+
if isinstance(self.linear_index, torch.Tensor):
|
1264 |
+
self.linear_index = self.linear_index.to(device)
|
1265 |
+
|
1266 |
+
def state_dict(self):
|
1267 |
+
return dict(
|
1268 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
1269 |
+
k=self.k,
|
1270 |
+
count=self.count,
|
1271 |
+
largest=self.largest,
|
1272 |
+
data_shape=self.data_shape and tuple(self.data_shape),
|
1273 |
+
top_data=self.top_data.cpu().detach().numpy(),
|
1274 |
+
top_index=self.top_index.cpu().detach().numpy(),
|
1275 |
+
next=self.next,
|
1276 |
+
linear_index=(
|
1277 |
+
self.linear_index.cpu().numpy()
|
1278 |
+
if isinstance(self.linear_index, torch.Tensor)
|
1279 |
+
else self.linear_index
|
1280 |
+
),
|
1281 |
+
perm=self.perm,
|
1282 |
+
)
|
1283 |
+
|
1284 |
+
def load_state_dict(self, state):
|
1285 |
+
self.k = int(state["k"])
|
1286 |
+
self.count = int(state["count"])
|
1287 |
+
self.largest = bool(state.get("largest", True))
|
1288 |
+
self.data_shape = (
|
1289 |
+
None if state["data_shape"] is None else tuple(state["data_shape"])
|
1290 |
+
)
|
1291 |
+
self.top_data = torch.from_numpy(state["top_data"])
|
1292 |
+
self.top_index = torch.from_numpy(state["top_index"])
|
1293 |
+
self.next = int(state["next"])
|
1294 |
+
self.linear_index = (
|
1295 |
+
torch.from_numpy(state["linear_index"])
|
1296 |
+
if len(state["linear_index"].shape) > 0
|
1297 |
+
else int(state["linear_index"])
|
1298 |
+
)
|
1299 |
+
|
1300 |
+
|
1301 |
+
class History(Stat):
|
1302 |
+
"""
|
1303 |
+
Accumulates the concatenation of all the added data.
|
1304 |
+
"""
|
1305 |
+
|
1306 |
+
def __init__(self, data=None, state=None):
|
1307 |
+
if state is not None:
|
1308 |
+
return super().__init__(state)
|
1309 |
+
self._data = data
|
1310 |
+
self._added = []
|
1311 |
+
|
1312 |
+
def _cat_added(self):
|
1313 |
+
if len(self._added):
|
1314 |
+
self._data = torch.cat(
|
1315 |
+
([self._data] if self._data is not None else []) + self._added
|
1316 |
+
)
|
1317 |
+
self._added = []
|
1318 |
+
|
1319 |
+
def add(self, d):
|
1320 |
+
self._added.append(d)
|
1321 |
+
if len(self._added) > 100:
|
1322 |
+
self._cat_added()
|
1323 |
+
|
1324 |
+
def history(self):
|
1325 |
+
self._cat_added()
|
1326 |
+
return self._data
|
1327 |
+
|
1328 |
+
def load_state_dict(self, state):
|
1329 |
+
data = state["data"]
|
1330 |
+
self._data = None if data is None else torch.from_numpy(data)
|
1331 |
+
self._added = []
|
1332 |
+
|
1333 |
+
def state_dict(self):
|
1334 |
+
self._cat_added()
|
1335 |
+
return dict(
|
1336 |
+
constructor=self.__module__ + "." + self.__class__.__name__ + "()",
|
1337 |
+
data=None if self._data is None else self._data.cpu().numpy(),
|
1338 |
+
)
|
1339 |
+
|
1340 |
+
def to_(self, device):
|
1341 |
+
"""Switches internal storage to specified device."""
|
1342 |
+
self._cat_added()
|
1343 |
+
if self._data is not None:
|
1344 |
+
self._data = self._data.to(device)
|
1345 |
+
|
1346 |
+
|
1347 |
+
class CombinedStat(Stat):
|
1348 |
+
"""
|
1349 |
+
A Stat that bundles together multiple Stat objects.
|
1350 |
+
Convenient for loading and saving a state_dict made up of a
|
1351 |
+
hierarchy of stats, and for use with the tally() function.
|
1352 |
+
Example:
|
1353 |
+
|
1354 |
+
cs = CombinedStat(m=Mean(), q=Quantile())
|
1355 |
+
for [b] in tally(cs, MyDataSet(), cache=fn, batch_size=100):
|
1356 |
+
cs.add(b)
|
1357 |
+
print(cs.m.mean())
|
1358 |
+
print(cs.q.median())
|
1359 |
+
"""
|
1360 |
+
|
1361 |
+
def __init__(self, state=None, **kwargs):
|
1362 |
+
self._objs = kwargs
|
1363 |
+
if state is not None:
|
1364 |
+
return super().__init__(state)
|
1365 |
+
|
1366 |
+
def __getattr__(self, k):
|
1367 |
+
if k in self._objs:
|
1368 |
+
return self._objs[k]
|
1369 |
+
raise AttributeError()
|
1370 |
+
|
1371 |
+
def add(self, d, *args, **kwargs):
|
1372 |
+
for obj in self._objs.values():
|
1373 |
+
obj.add(d, *args, **kwargs)
|
1374 |
+
|
1375 |
+
def load_state_dict(self, state):
|
1376 |
+
for prefix, obj in self._objs.items():
|
1377 |
+
obj.load_state_dict(pull_key_prefix(prefix, state))
|
1378 |
+
|
1379 |
+
def state_dict(self):
|
1380 |
+
result = {}
|
1381 |
+
for prefix, obj in self._objs.items():
|
1382 |
+
result.update(push_key_prefix(prefix, obj.state_dict()))
|
1383 |
+
return result
|
1384 |
+
|
1385 |
+
def to_(self, device):
|
1386 |
+
"""Switches internal storage to specified device."""
|
1387 |
+
for v in self._objs.values():
|
1388 |
+
v.to_(device)
|
1389 |
+
|
1390 |
+
|
1391 |
+
def push_key_prefix(prefix, d):
|
1392 |
+
"""
|
1393 |
+
Returns a dict with the same values as d, but where each key
|
1394 |
+
adds the prefix, followed by a dot.
|
1395 |
+
"""
|
1396 |
+
return {prefix + "." + k: v for k, v in d.items()}
|
1397 |
+
|
1398 |
+
|
1399 |
+
def pull_key_prefix(prefix, d):
|
1400 |
+
"""
|
1401 |
+
Returns a filtered dict of all the items of d that start with
|
1402 |
+
the given key prefix, plus a dot, with that prefix removed.
|
1403 |
+
"""
|
1404 |
+
pd = prefix + "."
|
1405 |
+
lpd = len(pd)
|
1406 |
+
return {k[lpd:]: v for k, v in d.items() if k.startswith(pd)}
|
1407 |
+
|
1408 |
+
|
1409 |
+
# We wish to be able to save None (null) values in numpy npz files,
|
1410 |
+
# yet do so without setting the unsecure 'allow_pickle' flag. To do
|
1411 |
+
# that, we will encode null as a special kind of IEEE 754 NaN value.
|
1412 |
+
# Inspired by https://github.com/zuiderkwast/nanbox/blob/master/nanbox.h
|
1413 |
+
# we follow the same Nanboxing scheme used in JavaScriptCore
|
1414 |
+
# (search for JSCJSValue.h#L435), which encodes null values in NaN
|
1415 |
+
# as the NaN value with hex pattern 0xfff8000000000002.
|
1416 |
+
|
1417 |
+
null_numpy_value = numpy.array(
|
1418 |
+
struct.unpack(">d", struct.pack(">Q", 0xFFF8000000000002))[0], dtype=numpy.float64
|
1419 |
+
)
|
1420 |
+
|
1421 |
+
|
1422 |
+
def is_null_numpy_value(v):
|
1423 |
+
"""
|
1424 |
+
True if v is a 64-bit float numpy scalar NaN matching null_numpy_value.
|
1425 |
+
"""
|
1426 |
+
return (
|
1427 |
+
isinstance(v, numpy.ndarray)
|
1428 |
+
and numpy.ndim(v) == 0
|
1429 |
+
and v.dtype == numpy.float64
|
1430 |
+
and numpy.isnan(v)
|
1431 |
+
and 0xFFF8000000000002 == struct.unpack(">Q", struct.pack(">d", v))[0]
|
1432 |
+
)
|
1433 |
+
|
1434 |
+
|
1435 |
+
def box_numpy_null(d):
|
1436 |
+
"""
|
1437 |
+
Replaces None with null_numpy_value, leaving non-None values unchanged.
|
1438 |
+
Recursively descends into a dictionary replacing None values.
|
1439 |
+
"""
|
1440 |
+
try:
|
1441 |
+
return {k: box_numpy_null(v) for k, v in d.items()}
|
1442 |
+
except Exception:
|
1443 |
+
return null_numpy_value if d is None else d
|
1444 |
+
|
1445 |
+
|
1446 |
+
def unbox_numpy_null(d):
|
1447 |
+
"""
|
1448 |
+
Reverses box_numpy_null, replacing null_numpy_value with None.
|
1449 |
+
Recursively descends into a dictionary replacing None values.
|
1450 |
+
"""
|
1451 |
+
try:
|
1452 |
+
return {k: unbox_numpy_null(v) for k, v in d.items()}
|
1453 |
+
except Exception:
|
1454 |
+
return None if is_null_numpy_value(d) else d
|
1455 |
+
|
1456 |
+
|
1457 |
+
def resolve_state_dict(s):
|
1458 |
+
"""
|
1459 |
+
Resolves a state, which can be a filename or a dict-like object.
|
1460 |
+
"""
|
1461 |
+
if isinstance(s, str):
|
1462 |
+
return unbox_numpy_null(numpy.load(s))
|
1463 |
+
return s
|
1464 |
+
|
1465 |
+
|
1466 |
+
global_load_cache_enabled = True
|
1467 |
+
|
1468 |
+
|
1469 |
+
def load_cached_state(cachefile, args, quiet=False, throw=False):
|
1470 |
+
"""
|
1471 |
+
Resolves a state, which can be a filename or a dict-like object.
|
1472 |
+
"""
|
1473 |
+
if not global_load_cache_enabled or cachefile is None:
|
1474 |
+
return None
|
1475 |
+
try:
|
1476 |
+
if isinstance(cachefile, dict):
|
1477 |
+
dat = cachefile
|
1478 |
+
cachefile = "state" # for printed messages
|
1479 |
+
else:
|
1480 |
+
dat = unbox_numpy_null(numpy.load(cachefile))
|
1481 |
+
for a, v in args.items():
|
1482 |
+
if a not in dat or dat[a] != v:
|
1483 |
+
if not quiet:
|
1484 |
+
print("%s %s changed from %s to %s" % (cachefile, a, dat[a], v))
|
1485 |
+
return None
|
1486 |
+
except (FileNotFoundError, ValueError) as e:
|
1487 |
+
if throw:
|
1488 |
+
raise e
|
1489 |
+
return None
|
1490 |
+
else:
|
1491 |
+
if not quiet:
|
1492 |
+
print("Loading cached %s" % cachefile)
|
1493 |
+
return dat
|
1494 |
+
|
1495 |
+
|
1496 |
+
def save_cached_state(cachefile, obj, args):
|
1497 |
+
"""
|
1498 |
+
Saves the state_dict of the given object in a dict or npz file.
|
1499 |
+
"""
|
1500 |
+
if cachefile is None:
|
1501 |
+
return
|
1502 |
+
dat = obj.state_dict()
|
1503 |
+
for a, v in args.items():
|
1504 |
+
if a in dat:
|
1505 |
+
assert dat[a] == v
|
1506 |
+
dat[a] = v
|
1507 |
+
if isinstance(cachefile, dict):
|
1508 |
+
cachefile.clear()
|
1509 |
+
cachefile.update(dat)
|
1510 |
+
else:
|
1511 |
+
os.makedirs(os.path.dirname(cachefile), exist_ok=True)
|
1512 |
+
numpy.savez(cachefile, **box_numpy_null(dat))
|
1513 |
+
|
1514 |
+
|
1515 |
+
class FixedSubsetSampler(Sampler):
|
1516 |
+
"""Represents a fixed sequence of data set indices.
|
1517 |
+
Subsets can be created by specifying a subset of output indexes.
|
1518 |
+
"""
|
1519 |
+
|
1520 |
+
def __init__(self, samples):
|
1521 |
+
self.samples = samples
|
1522 |
+
|
1523 |
+
def __iter__(self):
|
1524 |
+
return iter(self.samples)
|
1525 |
+
|
1526 |
+
def __len__(self):
|
1527 |
+
return len(self.samples)
|
1528 |
+
|
1529 |
+
def __getitem__(self, key):
|
1530 |
+
return self.samples[key]
|
1531 |
+
|
1532 |
+
def subset(self, new_subset):
|
1533 |
+
return FixedSubsetSampler(self.dereference(new_subset))
|
1534 |
+
|
1535 |
+
def dereference(self, indices):
|
1536 |
+
"""
|
1537 |
+
Translate output sample indices (small numbers indexing the sample)
|
1538 |
+
to input sample indices (larger number indexing the original full set)
|
1539 |
+
"""
|
1540 |
+
return [self.samples[i] for i in indices]
|
1541 |
+
|
1542 |
+
|
1543 |
+
class FixedRandomSubsetSampler(FixedSubsetSampler):
|
1544 |
+
"""Samples a fixed number of samples from the dataset, deterministically.
|
1545 |
+
Arguments:
|
1546 |
+
data_source,
|
1547 |
+
sample_size,
|
1548 |
+
seed (optional)
|
1549 |
+
"""
|
1550 |
+
|
1551 |
+
def __init__(self, data_source, start=None, end=None, seed=1):
|
1552 |
+
rng = random.Random(seed)
|
1553 |
+
shuffled = list(range(len(data_source)))
|
1554 |
+
rng.shuffle(shuffled)
|
1555 |
+
self.data_source = data_source
|
1556 |
+
super(FixedRandomSubsetSampler, self).__init__(shuffled[start:end])
|
1557 |
+
|
1558 |
+
def class_subset(self, class_filter):
|
1559 |
+
"""
|
1560 |
+
Returns only the subset matching the given rule.
|
1561 |
+
"""
|
1562 |
+
if isinstance(class_filter, int):
|
1563 |
+
|
1564 |
+
def rule(d):
|
1565 |
+
return d[1] == class_filter
|
1566 |
+
|
1567 |
+
else:
|
1568 |
+
rule = class_filter
|
1569 |
+
return self.subset(
|
1570 |
+
[i for i, j in enumerate(self.samples) if rule(self.data_source[j])]
|
1571 |
+
)
|
1572 |
+
|
1573 |
+
|
1574 |
+
def make_loader(
|
1575 |
+
dataset, sample_size=None, batch_size=1, sampler=None, random_sample=None, **kwargs
|
1576 |
+
):
|
1577 |
+
"""Utility for creating a dataloader on fixed sample subset."""
|
1578 |
+
import typing
|
1579 |
+
|
1580 |
+
if isinstance(dataset, typing.Callable):
|
1581 |
+
# To support deferred dataset loading, support passing a factory
|
1582 |
+
# that creates the dataset when called.
|
1583 |
+
dataset = dataset()
|
1584 |
+
if isinstance(dataset, torch.Tensor):
|
1585 |
+
# The dataset can be a simple tensor.
|
1586 |
+
dataset = torch.utils.data.TensorDataset(dataset)
|
1587 |
+
if sample_size is not None:
|
1588 |
+
assert sampler is None, "sampler cannot be specified with sample_size"
|
1589 |
+
if sample_size > len(dataset):
|
1590 |
+
print(
|
1591 |
+
"Warning: sample size %d > dataset size %d"
|
1592 |
+
% (sample_size, len(dataset))
|
1593 |
+
)
|
1594 |
+
sample_size = len(dataset)
|
1595 |
+
if random_sample is None:
|
1596 |
+
sampler = FixedSubsetSampler(list(range(sample_size)))
|
1597 |
+
else:
|
1598 |
+
sampler = FixedRandomSubsetSampler(
|
1599 |
+
dataset, seed=random_sample, end=sample_size
|
1600 |
+
)
|
1601 |
+
return torch.utils.data.DataLoader(
|
1602 |
+
dataset, sampler=sampler, batch_size=batch_size, **kwargs
|
1603 |
+
)
|
1604 |
+
|
1605 |
+
|
1606 |
+
# Unit Tests
|
1607 |
+
def _unit_test():
|
1608 |
+
import warnings
|
1609 |
+
|
1610 |
+
warnings.filterwarnings("error")
|
1611 |
+
import argparse
|
1612 |
+
import random
|
1613 |
+
import shutil
|
1614 |
+
import tempfile
|
1615 |
+
import time
|
1616 |
+
|
1617 |
+
parser = argparse.ArgumentParser(description="Test things out")
|
1618 |
+
parser.add_argument("--mode", default="cpu", help="cpu or cuda")
|
1619 |
+
parser.add_argument("--test_size", type=int, default=1000000)
|
1620 |
+
args = parser.parse_args()
|
1621 |
+
testdir = tempfile.mkdtemp()
|
1622 |
+
batch_size = random.randint(500, 1500)
|
1623 |
+
|
1624 |
+
# Test NaNboxing.
|
1625 |
+
assert numpy.isnan(null_numpy_value)
|
1626 |
+
assert is_null_numpy_value(null_numpy_value)
|
1627 |
+
assert not is_null_numpy_value(numpy.nan)
|
1628 |
+
|
1629 |
+
# Test Covariance
|
1630 |
+
goal = torch.tensor(numpy.random.RandomState(1).standard_normal(10 * 10)).view(
|
1631 |
+
10, 10
|
1632 |
+
)
|
1633 |
+
data = (
|
1634 |
+
torch.tensor(numpy.random.RandomState(2).standard_normal(args.test_size * 10))
|
1635 |
+
.view(args.test_size, 10)
|
1636 |
+
.mm(goal)
|
1637 |
+
)
|
1638 |
+
data += torch.randn(1, 10) * 999
|
1639 |
+
dcov = data.t().cov()
|
1640 |
+
dcorr = data.t().corrcoef()
|
1641 |
+
rcov = Covariance()
|
1642 |
+
rcov.add(data) # All one batch
|
1643 |
+
assert (rcov.covariance() - dcov).abs().max() < 1e-16
|
1644 |
+
cs = CombinedStat(cov=Covariance(), xcov=CrossCovariance())
|
1645 |
+
ds = torch.utils.data.TensorDataset(data)
|
1646 |
+
for [a] in tally(cs, ds, batch_size=9876):
|
1647 |
+
cs.cov.add(a)
|
1648 |
+
cs.xcov.add(a[:, :3], a[:, 3:])
|
1649 |
+
assert (data.mean(0) - cs.cov.mean()).abs().max() < 1e-12
|
1650 |
+
assert (dcov - cs.cov.covariance()).abs().max() < 2e-12
|
1651 |
+
assert (dcov[:3, 3:] - cs.xcov.covariance()).abs().max() < 1e-12
|
1652 |
+
assert (dcov.diagonal() - torch.cat(cs.xcov.variance())).abs().max() < 1e-12
|
1653 |
+
assert (dcorr - cs.cov.correlation()).abs().max() < 2e-12
|
1654 |
+
|
1655 |
+
# Test CrossCovariance and CrossIoU
|
1656 |
+
fn = f"{testdir}/cross_cache.npz"
|
1657 |
+
ds = torch.utils.data.TensorDataset(
|
1658 |
+
(
|
1659 |
+
torch.arange(args.test_size)[:, None] % torch.arange(1, 6)[None, :] == 0
|
1660 |
+
).double(),
|
1661 |
+
(
|
1662 |
+
torch.arange(args.test_size)[:, None] % torch.arange(5, 8)[None, :] == 0
|
1663 |
+
).double(),
|
1664 |
+
)
|
1665 |
+
c = CombinedStat(c=CrossCovariance(), iou=CrossIoU())
|
1666 |
+
riou = IoU()
|
1667 |
+
count = 0
|
1668 |
+
for [a, b] in tally(c, ds, cache=fn, batch_size=100):
|
1669 |
+
count += 1
|
1670 |
+
c.add(a, b)
|
1671 |
+
riou.add(torch.cat([a, b], dim=1))
|
1672 |
+
assert count == -(-args.test_size // 100)
|
1673 |
+
cor = c.c.correlation()
|
1674 |
+
iou = c.iou.iou()
|
1675 |
+
assert cor.shape == iou.shape == (5, 3)
|
1676 |
+
assert iou[4, 0] == 1.0
|
1677 |
+
assert abs(iou[0, 2] + (-args.test_size // 7 / float(args.test_size))) < 1e-6
|
1678 |
+
assert abs(cor[4, 0] - 1.0) < 1e-2
|
1679 |
+
assert abs(cor[0, 2] - 0.0) < 1e-6
|
1680 |
+
assert all((riou.iou()[:5, -3:] == iou).view(-1))
|
1681 |
+
assert all(riou.iou().diagonal(0) == 1)
|
1682 |
+
c = CombinedStat(c=CrossCovariance(), iou=CrossIoU())
|
1683 |
+
count = 0
|
1684 |
+
for [a, b] in tally(c, ds, cache=fn, batch_size=10):
|
1685 |
+
count += 1
|
1686 |
+
c.add(a, b)
|
1687 |
+
assert count == 0
|
1688 |
+
assert all((c.c.correlation() == cor).view(-1))
|
1689 |
+
assert all((c.iou.iou() == iou).view(-1))
|
1690 |
+
|
1691 |
+
# Test Concatantaion, Mean, Bincount and tally.
|
1692 |
+
fn = f"{testdir}/series_cache.npz"
|
1693 |
+
count = 0
|
1694 |
+
ds = torch.utils.data.TensorDataset(torch.arange(args.test_size))
|
1695 |
+
c = CombinedStat(s=History(), m=Mean(), b=Bincount())
|
1696 |
+
for [b] in tally(c, ds, cache=fn, batch_size=batch_size):
|
1697 |
+
count += 1
|
1698 |
+
c.add(b)
|
1699 |
+
assert count == -(-args.test_size // batch_size)
|
1700 |
+
assert len(c.s.history()) == args.test_size
|
1701 |
+
assert c.s.history()[-1] == args.test_size - 1
|
1702 |
+
assert all(c.s.history() == ds.tensors[0])
|
1703 |
+
assert all(c.b.bincount() == torch.ones(args.test_size))
|
1704 |
+
assert c.m.mean() == float(args.test_size - 1) / 2.0
|
1705 |
+
c2 = CombinedStat(s=History(), m=Mean(), b=Bincount())
|
1706 |
+
batches = tally(c2, ds, cache=fn)
|
1707 |
+
assert len(c2.s.history()) == args.test_size
|
1708 |
+
assert all(c2.s.history() == c.s.history())
|
1709 |
+
assert all(c2.b.bincount() == torch.ones(args.test_size))
|
1710 |
+
assert c2.m.mean() == c.m.mean()
|
1711 |
+
count = 0
|
1712 |
+
for b in batches:
|
1713 |
+
count += 1
|
1714 |
+
assert count == 0 # Shouldn't do anything when it's cached
|
1715 |
+
|
1716 |
+
# An adverarial case: we keep finding more numbers in the middle
|
1717 |
+
# as the stream goes on.
|
1718 |
+
amount = args.test_size
|
1719 |
+
quantiles = 1000
|
1720 |
+
data = numpy.arange(float(amount))
|
1721 |
+
data[1::2] = data[-1::-2] + (len(data) - 1)
|
1722 |
+
data /= 2
|
1723 |
+
depth = 50
|
1724 |
+
alldata = data[:, None] + (numpy.arange(depth) * amount)[None, :]
|
1725 |
+
actual_sum = torch.FloatTensor(numpy.sum(alldata * alldata, axis=0))
|
1726 |
+
amt = amount // depth
|
1727 |
+
for r in range(depth):
|
1728 |
+
numpy.random.shuffle(alldata[r * amt : r * amt + amt, r])
|
1729 |
+
if args.mode == "cuda":
|
1730 |
+
alldata = torch.cuda.FloatTensor(alldata)
|
1731 |
+
device = torch.device("cuda")
|
1732 |
+
else:
|
1733 |
+
alldata = torch.FloatTensor(alldata)
|
1734 |
+
device = None
|
1735 |
+
starttime = time.time()
|
1736 |
+
cs = CombinedStat(
|
1737 |
+
qc=Quantile(),
|
1738 |
+
m=Mean(),
|
1739 |
+
v=Variance(),
|
1740 |
+
c=Covariance(),
|
1741 |
+
s=SecondMoment(),
|
1742 |
+
t=TopK(),
|
1743 |
+
i=IoU(),
|
1744 |
+
)
|
1745 |
+
# Feed data in little batches
|
1746 |
+
i = 0
|
1747 |
+
while i < len(alldata):
|
1748 |
+
batch_size = numpy.random.randint(1000)
|
1749 |
+
cs.add(alldata[i : i + batch_size])
|
1750 |
+
i += batch_size
|
1751 |
+
# Test state dict
|
1752 |
+
saved = cs.state_dict()
|
1753 |
+
# numpy.savez(f'{testdir}/saved.npz', **box_numpy_null(saved))
|
1754 |
+
# saved = unbox_numpy_null(numpy.load(f'{testdir}/saved.npz'))
|
1755 |
+
cs.save(f"{testdir}/saved.npz")
|
1756 |
+
loaded = unbox_numpy_null(numpy.load(f"{testdir}/saved.npz"))
|
1757 |
+
assert set(loaded.keys()) == set(saved.keys())
|
1758 |
+
|
1759 |
+
# Restore using state=saved in constructor.
|
1760 |
+
cs2 = CombinedStat(
|
1761 |
+
qc=Quantile(),
|
1762 |
+
m=Mean(),
|
1763 |
+
v=Variance(),
|
1764 |
+
c=Covariance(),
|
1765 |
+
s=SecondMoment(),
|
1766 |
+
t=TopK(),
|
1767 |
+
i=IoU(),
|
1768 |
+
state=saved,
|
1769 |
+
)
|
1770 |
+
# saved = unbox_numpy_null(numpy.load(f'{testdir}/saved.npz'))
|
1771 |
+
assert not cs2.qc.device.type == "cuda"
|
1772 |
+
cs2.to_(device)
|
1773 |
+
# alldata = alldata.cpu()
|
1774 |
+
cs2.add(alldata)
|
1775 |
+
actual_sum *= 2
|
1776 |
+
# print(abs(alldata.mean(0) - cs2.m.mean()) / alldata.mean())
|
1777 |
+
assert all(abs(alldata.mean(0) - cs2.m.mean()) / alldata.mean() < 1e-5)
|
1778 |
+
assert all(abs(alldata.mean(0) - cs2.v.mean()) / alldata.mean() < 1e-5)
|
1779 |
+
assert all(abs(alldata.mean(0) - cs2.c.mean()) / alldata.mean() < 1e-5)
|
1780 |
+
# print(abs(alldata.var(0) - cs2.v.variance()) / alldata.var(0))
|
1781 |
+
assert all(abs(alldata.var(0) - cs2.v.variance()) / alldata.var(0) < 1e-3)
|
1782 |
+
assert all(abs(alldata.var(0) - cs2.c.variance()) / alldata.var(0) < 1e-2)
|
1783 |
+
# print(abs(alldata.std(0) - cs2.v.stdev()) / alldata.std(0))
|
1784 |
+
assert all(abs(alldata.std(0) - cs2.v.stdev()) / alldata.std(0) < 1e-4)
|
1785 |
+
# print(abs(alldata.std(0) - cs2.c.stdev()) / alldata.std(0))
|
1786 |
+
assert all(abs(alldata.std(0) - cs2.c.stdev()) / alldata.std(0) < 2e-3)
|
1787 |
+
moment = (alldata.t() @ alldata) / len(alldata)
|
1788 |
+
# print(abs(moment - cs2.s.moment()) / moment.abs())
|
1789 |
+
assert all((abs(moment - cs2.s.moment()) / moment.abs()).view(-1) < 1e-2)
|
1790 |
+
assert all(alldata.max(dim=0)[0] == cs2.t.topk()[0][:, 0])
|
1791 |
+
assert cs2.i.iou()[0, 0] == 1
|
1792 |
+
assert all((cs2.i.iou()[1:, 1:] == 1).view(-1))
|
1793 |
+
assert all(cs2.i.iou()[1:, 0] < 1)
|
1794 |
+
assert all(cs2.i.iou()[1:, 0] == cs2.i.iou()[0, 1:])
|
1795 |
+
|
1796 |
+
# Restore using cs.load() method.
|
1797 |
+
cs = CombinedStat(
|
1798 |
+
qc=Quantile(),
|
1799 |
+
m=Mean(),
|
1800 |
+
v=Variance(),
|
1801 |
+
c=Covariance(),
|
1802 |
+
s=SecondMoment(),
|
1803 |
+
t=TopK(),
|
1804 |
+
i=IoU(),
|
1805 |
+
)
|
1806 |
+
cs.load(f"{testdir}/saved.npz")
|
1807 |
+
assert not cs.qc.device.type == "cuda"
|
1808 |
+
cs.to_(device)
|
1809 |
+
cs.add(alldata)
|
1810 |
+
# actual_sum *= 2
|
1811 |
+
# print(abs(alldata.mean(0) - cs.m.mean()) / alldata.mean())
|
1812 |
+
assert all(abs(alldata.mean(0) - cs.m.mean()) / alldata.mean() < 1e-5)
|
1813 |
+
assert all(abs(alldata.mean(0) - cs.v.mean()) / alldata.mean() < 1e-5)
|
1814 |
+
assert all(abs(alldata.mean(0) - cs.c.mean()) / alldata.mean() < 1e-5)
|
1815 |
+
# print(abs(alldata.var(0) - cs.v.variance()) / alldata.var(0))
|
1816 |
+
assert all(abs(alldata.var(0) - cs.v.variance()) / alldata.var(0) < 1e-3)
|
1817 |
+
assert all(abs(alldata.var(0) - cs.c.variance()) / alldata.var(0) < 1e-2)
|
1818 |
+
# print(abs(alldata.std(0) - cs.v.stdev()) / alldata.std(0))
|
1819 |
+
assert all(abs(alldata.std(0) - cs.v.stdev()) / alldata.std(0) < 1e-4)
|
1820 |
+
# print(abs(alldata.std(0) - cs.c.stdev()) / alldata.std(0))
|
1821 |
+
assert all(abs(alldata.std(0) - cs.c.stdev()) / alldata.std(0) < 2e-3)
|
1822 |
+
moment = (alldata.t() @ alldata) / len(alldata)
|
1823 |
+
# print(abs(moment - cs.s.moment()) / moment.abs())
|
1824 |
+
assert all((abs(moment - cs.s.moment()) / moment.abs()).view(-1) < 1e-2)
|
1825 |
+
assert all(alldata.max(dim=0)[0] == cs.t.topk()[0][:, 0])
|
1826 |
+
assert cs.i.iou()[0, 0] == 1
|
1827 |
+
assert all((cs.i.iou()[1:, 1:] == 1).view(-1))
|
1828 |
+
assert all(cs.i.iou()[1:, 0] < 1)
|
1829 |
+
assert all(cs.i.iou()[1:, 0] == cs.i.iou()[0, 1:])
|
1830 |
+
|
1831 |
+
# Randomized quantile test
|
1832 |
+
qc = cs.qc
|
1833 |
+
ro = qc.readout(1001).cpu()
|
1834 |
+
endtime = time.time()
|
1835 |
+
gt = (
|
1836 |
+
torch.linspace(0, amount, quantiles + 1)[None, :]
|
1837 |
+
+ (torch.arange(qc.depth, dtype=torch.float) * amount)[:, None]
|
1838 |
+
)
|
1839 |
+
maxreldev = torch.max(torch.abs(ro - gt) / amount) * quantiles
|
1840 |
+
print("Randomized quantile test results:")
|
1841 |
+
print("Maximum relative deviation among %d perentiles: %f" % (quantiles, maxreldev))
|
1842 |
+
minerr = torch.max(
|
1843 |
+
torch.abs(
|
1844 |
+
qc.minmax().cpu()[:, 0] - torch.arange(qc.depth, dtype=torch.float) * amount
|
1845 |
+
)
|
1846 |
+
)
|
1847 |
+
maxerr = torch.max(
|
1848 |
+
torch.abs(
|
1849 |
+
(qc.minmax().cpu()[:, -1] + 1)
|
1850 |
+
- (torch.arange(qc.depth, dtype=torch.float) + 1) * amount
|
1851 |
+
)
|
1852 |
+
)
|
1853 |
+
print("Minmax error %f, %f" % (minerr, maxerr))
|
1854 |
+
interr = torch.max(
|
1855 |
+
torch.abs(qc.integrate(lambda x: x * x).cpu() - actual_sum) / actual_sum
|
1856 |
+
)
|
1857 |
+
print("Integral error: %f" % interr)
|
1858 |
+
medianerr = torch.max(
|
1859 |
+
torch.abs(qc.median() - alldata.median(0)[0]) / alldata.median(0)[0]
|
1860 |
+
).cpu()
|
1861 |
+
print("Median error: %f" % medianerr)
|
1862 |
+
meanerr = torch.max(torch.abs(qc.mean() - alldata.mean(0)) / alldata.mean(0)).cpu()
|
1863 |
+
print("Mean error: %f" % meanerr)
|
1864 |
+
varerr = torch.max(torch.abs(qc.variance() - alldata.var(0)) / alldata.var(0)).cpu()
|
1865 |
+
print("Variance error: %f" % varerr)
|
1866 |
+
counterr = (
|
1867 |
+
(qc.integrate(lambda x: torch.ones(x.shape[-1]).cpu()) - qc.size())
|
1868 |
+
/ (0.0 + qc.size())
|
1869 |
+
).item()
|
1870 |
+
print("Count error: %f" % counterr)
|
1871 |
+
print("Time %f" % (endtime - starttime))
|
1872 |
+
# Algorithm is randomized, so some of these will fail with low probability.
|
1873 |
+
assert maxreldev < 1.0
|
1874 |
+
assert minerr == 0.0
|
1875 |
+
assert maxerr == 0.0
|
1876 |
+
assert interr < 0.01
|
1877 |
+
assert abs(counterr) < 0.001
|
1878 |
+
shutil.rmtree(testdir, ignore_errors=True)
|
1879 |
+
print("OK")
|
1880 |
+
|
1881 |
+
|
1882 |
+
if __name__ == "__main__":
|
1883 |
+
_unit_test()
|
hparams/GRACE/README.md
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
alg_name: "GRACE"
|
2 |
+
model_name: "./hugging_cache/gpt-j-6B"
|
3 |
+
device: 0
|
4 |
+
|
5 |
+
inner_params:
|
6 |
+
- transformer.h[25].mlp.fc_out.weight
|
7 |
+
|
8 |
+
edit_lr: 1.0
|
9 |
+
n_iter: 200
|
10 |
+
eps: 1.0
|
11 |
+
dist_fn: euc # euc, mmd, cos
|
12 |
+
val_init: cold # cold, warm
|
13 |
+
val_train: sgd # sgd, pert
|
14 |
+
val_reg: None # early
|
15 |
+
reg: early_stop # early_stop
|
16 |
+
replacement: replace_last # replace_last, replace_all, replace_prompt
|
17 |
+
eps_expand: coverage # , moving_avg, decay
|
18 |
+
num_pert: 8 # only matters when using perturbation training
|
19 |
+
dropout: 0.0
|
hparams/GRACE/gpt2.yaml
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
alg_name: "GRACE"
|
2 |
+
model_name: "./hugging_cache/gpt2"
|
3 |
+
device: cpu
|
4 |
+
|
5 |
+
inner_params:
|
6 |
+
- transformer.h[8].mlp.c_fc.weight
|
7 |
+
|
8 |
+
edit_lr: 1.0
|
9 |
+
n_iter: 30
|
10 |
+
eps: 1.0
|
11 |
+
dist_fn: euc # euc, mmd, cos
|
12 |
+
val_init: cold # cold, warm
|
13 |
+
val_train: sgd # sgd, pert
|
14 |
+
val_reg: None # early
|
15 |
+
reg: early_stop # early_stop
|
16 |
+
replacement: replace_last # replace_last, replace_all, replace_prompt
|
17 |
+
eps_expand: coverage # , moving_avg, decay
|
18 |
+
num_pert: 8 # only matters when using perturbation training
|
19 |
+
dropout: 0.0
|
hparams/config.yaml
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
save_dir: models/
|
2 |
+
log_dir: logs/
|
3 |
+
|
4 |
+
defaults:
|
5 |
+
alg_name: KN # Editing Method
|
6 |
+
hparams_name: KN/t5-3b # Edited Model Config Path
|
requirements.txt
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
datasets==1.18.3
|
2 |
+
einops==0.4.0
|
3 |
+
gpustat==1.1
|
4 |
+
hydra-core==1.1.1
|
5 |
+
higher==0.2.1
|
6 |
+
importlib-metadata==6.3.0
|
7 |
+
matplotlib==3.5.1
|
8 |
+
nltk==3.6.5
|
9 |
+
numpy==1.22.1
|
10 |
+
omegaconf==2.1.1
|
11 |
+
pandas==1.4.0
|
12 |
+
PyYAML==6.0
|
13 |
+
scikit-learn==1.0.2
|
14 |
+
scipy==1.7.3
|
15 |
+
sentence-transformers==2.2.2
|
16 |
+
tokenizers==0.13.3
|
17 |
+
torch==2.0.1
|
18 |
+
tqdm==4.62.3
|
19 |
+
transformers==4.30.1
|
20 |
+
openai==0.27.9
|
21 |
+
peft==0.5.0
|
22 |
+
timm==0.9.7
|
23 |
+
iopath==0.1.10
|
24 |
+
opencv-python==4.8.0.76
|
25 |
+
gradio
|
utils.py
ADDED
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModel
|
2 |
+
from transformers import GPT2TokenizerFast, GPT2Tokenizer
|
3 |
+
from easyeditor import apply_grace_to_model, GraceHyperParams,nethook
|
4 |
+
import torch
|
5 |
+
import gradio as gr
|
6 |
+
|
7 |
+
|
8 |
+
|
9 |
+
def edit(prompt, target_new, num_steps, replacement):
|
10 |
+
request={"prompt":prompt,"target_new":target_new}
|
11 |
+
hparams = GraceHyperParams.from_hparams("./hparams/GRACE/gpt2.yaml")
|
12 |
+
|
13 |
+
model = AutoModelForCausalLM.from_pretrained("./models/gpt2", device_map='cpu')
|
14 |
+
tok = GPT2Tokenizer.from_pretrained("./models/gpt2")
|
15 |
+
tok.pad_token_id = tok.eos_token_id
|
16 |
+
global edit_model
|
17 |
+
edit_model = apply_grace_to_model(model,tok,request,hparams, num_steps, replacement)
|
18 |
+
return prompt
|
19 |
+
|
20 |
+
def generate(input_text, target_new=None):
|
21 |
+
tok = GPT2Tokenizer.from_pretrained("./models/gpt2")
|
22 |
+
hparams = GraceHyperParams.from_hparams("./hparams/GRACE/gpt2.yaml")
|
23 |
+
tok.pad_token_id = tok.eos_token_id
|
24 |
+
|
25 |
+
global edit_model
|
26 |
+
|
27 |
+
if target_new is None:
|
28 |
+
max_new_tokens = 25
|
29 |
+
else:
|
30 |
+
max_new_tokens = len(tok.encode(target_new))
|
31 |
+
prompt_len = len(input_text)
|
32 |
+
input_ids = tok.encode(input_text, return_tensors='pt').to('cpu')
|
33 |
+
edit_output = edit_model.generate(input_ids, max_new_tokens=max_new_tokens, pad_token_id=tok.eos_token_id)
|
34 |
+
edit_reply = tok.decode(edit_output[0], skip_special_tokens=True)
|
35 |
+
torch.cuda.empty_cache()
|
36 |
+
|
37 |
+
ori_model = AutoModelForCausalLM.from_pretrained("./models/gpt2").to('cpu')
|
38 |
+
ori_output = ori_model.generate(input_ids, max_new_tokens=max_new_tokens, pad_token_id=tok.eos_token_id)
|
39 |
+
ori_reply = tok.decode(ori_output[0], skip_special_tokens=True)
|
40 |
+
ori_reply = [(_, 'output') if i > prompt_len else (_, None) for i, _ in enumerate(ori_reply)]
|
41 |
+
edit_reply = [(_, 'output') if i > prompt_len else (_, None) for i, _ in enumerate(edit_reply)]
|
42 |
+
return ori_reply, edit_reply
|