John Hewitt
commited on
Commit
•
5b1efcc
1
Parent(s):
49ee142
model upload
Browse files- README.md +5 -2
- config.json +36 -0
- configuration_backpack_gpt2.py +42 -0
- modeling_backpack_gpt2.py +226 -0
- pytorch_model.bin +3 -0
README.md
CHANGED
@@ -1,3 +1,6 @@
|
|
1 |
---
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
pipeline_tag: text-generation
|
3 |
+
tags:
|
4 |
+
- text-generation-inference
|
5 |
+
library_name: transformers
|
6 |
+
---
|
config.json
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"BackpackGPT2LMHeadModel"
|
4 |
+
],
|
5 |
+
"auto_map": {
|
6 |
+
"AutoConfig": "configuration_backpack_gpt2.BackpackGPT2Config",
|
7 |
+
"AutoModelForCausalLM": "modeling_backpack_gpt2.BackpackGPT2LMHeadModel"
|
8 |
+
},
|
9 |
+
"activation_function": "gelu_new",
|
10 |
+
"attn_pdrop": 0.1,
|
11 |
+
"bos_token_id": 50256,
|
12 |
+
"embd_pdrop": 0.1,
|
13 |
+
"eos_token_id": 50256,
|
14 |
+
"initializer_range": 0.02,
|
15 |
+
"layer_norm_epsilon": 1e-05,
|
16 |
+
"model_type": "gpt2",
|
17 |
+
"n_embd": 768,
|
18 |
+
"n_head": 12,
|
19 |
+
"n_inner": null,
|
20 |
+
"n_layer": 12,
|
21 |
+
"n_positions": 512,
|
22 |
+
"num_senses": 16,
|
23 |
+
"reorder_and_upcast_attn": false,
|
24 |
+
"resid_pdrop": 0.1,
|
25 |
+
"scale_attn_by_inverse_layer_idx": true,
|
26 |
+
"scale_attn_weights": true,
|
27 |
+
"sense_intermediate_scale": 4,
|
28 |
+
"summary_activation": null,
|
29 |
+
"summary_first_dropout": 0.1,
|
30 |
+
"summary_proj_to_labels": true,
|
31 |
+
"summary_type": "cls_index",
|
32 |
+
"summary_use_proj": true,
|
33 |
+
"transformers_version": "4.29.0.dev0",
|
34 |
+
"use_cache": true,
|
35 |
+
"vocab_size": 50264
|
36 |
+
}
|
configuration_backpack_gpt2.py
ADDED
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers.models.gpt2.configuration_gpt2 import GPT2Config
|
2 |
+
|
3 |
+
class BackpackGPT2Config(GPT2Config):
|
4 |
+
"""
|
5 |
+
This is the configuration class to store the configuration of a [`GPT2Model`] or a [`TFGPT2Model`]. It is used to
|
6 |
+
instantiate a Backpack GPT-2 model according to the specified arguments, defining the model architecture.
|
7 |
+
|
8 |
+
Configuration objects inherit from [`GPT2Config`] and can be used to control the model outputs. Read the
|
9 |
+
documentation from [`GPT2Config`] for more information.
|
10 |
+
|
11 |
+
Args:
|
12 |
+
num_senses (`int`, *optional*, defaults to 16):
|
13 |
+
The number of sense vectors to define for each word.
|
14 |
+
sense_intermediate_scale (`int`, *optional*, defaults ot 4):
|
15 |
+
The hidden dimensionality of the sense vector network.
|
16 |
+
|
17 |
+
Example:
|
18 |
+
|
19 |
+
```python
|
20 |
+
>>> from transformers import BackpackGPT2Config, BackpackGPT2Model
|
21 |
+
|
22 |
+
>>> # Initializing a GPT2 configuration
|
23 |
+
>>> configuration = BackpackGPT2Config()
|
24 |
+
|
25 |
+
>>> # Initializing a model (with random weights) from the configuration
|
26 |
+
>>> model = BackpackGPT2Model(configuration)
|
27 |
+
|
28 |
+
>>> # Accessing the model configuration
|
29 |
+
>>> configuration = model.config
|
30 |
+
"""
|
31 |
+
|
32 |
+
def __init__(self,
|
33 |
+
vocab_size=50264,
|
34 |
+
num_senses=16,
|
35 |
+
sense_intermediate_scale=4,
|
36 |
+
n_positions=512,
|
37 |
+
scale_attn_by_inverse_layer_idx=True,
|
38 |
+
**kwargs,
|
39 |
+
):
|
40 |
+
self.num_senses = num_senses
|
41 |
+
self.sense_intermediate_scale = sense_intermediate_scale
|
42 |
+
super().__init__(vocab_size=vocab_size, n_positions=n_positions, scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx, **kwargs)
|
modeling_backpack_gpt2.py
ADDED
@@ -0,0 +1,226 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
from typing import Optional, Tuple
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.utils.checkpoint
|
7 |
+
from torch import nn
|
8 |
+
|
9 |
+
from transformers.activations import ACT2FN
|
10 |
+
from transformers.pytorch_utils import Conv1D
|
11 |
+
from transformers.utils import (
|
12 |
+
ModelOutput,
|
13 |
+
logging,
|
14 |
+
)
|
15 |
+
from transformers.models.gpt2.modeling_gpt2 import GPT2Model, GPT2PreTrainedModel
|
16 |
+
from .configuration_backpack_gpt2 import BackpackGPT2Config
|
17 |
+
|
18 |
+
logger = logging.get_logger(__name__)
|
19 |
+
|
20 |
+
|
21 |
+
### Backpack-Specific
|
22 |
+
class BackpackGPT2PreTrainedModel(GPT2PreTrainedModel):
|
23 |
+
"""
|
24 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
25 |
+
models.
|
26 |
+
"""
|
27 |
+
_keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias"]
|
28 |
+
|
29 |
+
config_class = BackpackGPT2Config
|
30 |
+
base_model_prefix = "backpack"
|
31 |
+
is_parallelizable = True
|
32 |
+
supports_gradient_checkpointing = False
|
33 |
+
_no_split_modules = ["GPT2Block", "BackpackNoMixBlock"]
|
34 |
+
|
35 |
+
def __init__(self, *inputs, **kwargs):
|
36 |
+
super().__init__(*inputs, **kwargs)
|
37 |
+
|
38 |
+
class BackpackMLP(nn.Module):
|
39 |
+
|
40 |
+
def __init__(self, embed_dim, intermediate_dim, out_dim, config):
|
41 |
+
super().__init__()
|
42 |
+
self.c_fc = Conv1D(intermediate_dim, embed_dim)
|
43 |
+
self.c_proj = Conv1D(out_dim, intermediate_dim)
|
44 |
+
self.act = ACT2FN[config.activation_function]
|
45 |
+
self.dropout = nn.Dropout(config.resid_pdrop)
|
46 |
+
|
47 |
+
def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
|
48 |
+
hidden_states = self.c_fc(hidden_states)
|
49 |
+
hidden_states = self.act(hidden_states)
|
50 |
+
hidden_states = self.c_proj(hidden_states)
|
51 |
+
hidden_states = self.dropout(hidden_states)
|
52 |
+
return hidden_states
|
53 |
+
|
54 |
+
class BackpackNoMixBlock(nn.Module):
|
55 |
+
|
56 |
+
def __init__(self, config):
|
57 |
+
super().__init__()
|
58 |
+
self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
59 |
+
self.ln_2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
60 |
+
self.mlp = BackpackMLP(config.n_embd, config.n_embd*4, config.n_embd, config)
|
61 |
+
self.resid_dropout1 = nn.Dropout(config.resid_pdrop)
|
62 |
+
self.resid_dropout2 = nn.Dropout(config.resid_pdrop)
|
63 |
+
|
64 |
+
def forward(self, hidden_states, residual):
|
65 |
+
residual = self.resid_dropout1(hidden_states) + residual
|
66 |
+
hidden_states = self.ln_1(residual)
|
67 |
+
mlp_out = self.mlp(hidden_states)
|
68 |
+
residual = self.resid_dropout2(mlp_out) + residual
|
69 |
+
hidden_states = self.ln_2(residual)
|
70 |
+
return hidden_states
|
71 |
+
|
72 |
+
|
73 |
+
class BackpackSenseNetwork(nn.Module):
|
74 |
+
def __init__(self, config, num_senses, device=None, dtype=None):
|
75 |
+
super().__init__()
|
76 |
+
self.num_senses = num_senses
|
77 |
+
#self.embeddings = embeddings
|
78 |
+
self.n_embd = config.n_embd
|
79 |
+
|
80 |
+
self.dropout = nn.Dropout(config.embd_pdrop)
|
81 |
+
self.block = BackpackNoMixBlock(config)
|
82 |
+
self.ln = nn.LayerNorm(self.n_embd, eps=config.layer_norm_epsilon)
|
83 |
+
self.final_mlp = BackpackMLP(
|
84 |
+
embed_dim=config.n_embd,
|
85 |
+
intermediate_dim=config.sense_intermediate_scale*config.n_embd,
|
86 |
+
out_dim=config.n_embd*config.num_senses,
|
87 |
+
config=config,
|
88 |
+
)
|
89 |
+
|
90 |
+
def forward(self, input_embeds):
|
91 |
+
residual = self.dropout(input_embeds)
|
92 |
+
hidden_states = self.ln(residual)
|
93 |
+
hidden_states = self.block(hidden_states, residual)
|
94 |
+
senses = self.final_mlp(hidden_states)
|
95 |
+
bs, s, nvd = senses.shape
|
96 |
+
return senses.reshape(bs, s, self.num_senses, self.n_embd).transpose(1,2) # (bs, nv, s, d)
|
97 |
+
|
98 |
+
class BackpackWeightNetwork(nn.Module):
|
99 |
+
|
100 |
+
def __init__(self, num_senses, embed_dim):
|
101 |
+
super().__init__()
|
102 |
+
self.n_embd = embed_dim
|
103 |
+
self.num_senses = num_senses
|
104 |
+
self.c_attn = nn.Linear(embed_dim, 2*embed_dim)
|
105 |
+
self.softmax_scale = None
|
106 |
+
|
107 |
+
def forward(self, encoded):
|
108 |
+
b, s, d = encoded.shape
|
109 |
+
encoded = self.c_attn(encoded) # (b, s, 2*d)
|
110 |
+
encoded = encoded.reshape(b, s, 2, self.num_senses, d // self.num_senses) #(b, s, 2, nv, d//nv)
|
111 |
+
batch_size, seqlen = encoded.shape[0], encoded.shape[1]
|
112 |
+
|
113 |
+
# compute scores & mask
|
114 |
+
q, k = encoded.unbind(dim=2)
|
115 |
+
softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
|
116 |
+
scores = torch.einsum('bthd,bshd->bhts', q, k * softmax_scale)
|
117 |
+
causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
|
118 |
+
scores = scores + causal_mask.to(dtype=scores.dtype)
|
119 |
+
|
120 |
+
return torch.softmax(scores, dim=-1, dtype=q.dtype)
|
121 |
+
|
122 |
+
|
123 |
+
@dataclass
|
124 |
+
class BackpackGPT2BaseModelOutput(ModelOutput):
|
125 |
+
hidden_states: torch.FloatTensor = None
|
126 |
+
contextualization: torch.FloatTensor = None
|
127 |
+
|
128 |
+
class BackpackGPT2Model(BackpackGPT2PreTrainedModel):
|
129 |
+
_keys_to_ignore_on_load_missing = [r".*attn.masked_bias", r".*attn.bias"]
|
130 |
+
|
131 |
+
def __init__(self, config):
|
132 |
+
super().__init__(config)
|
133 |
+
|
134 |
+
self.embed_dim = config.n_embd
|
135 |
+
|
136 |
+
self.num_senses = config.num_senses
|
137 |
+
self.gpt2_model = GPT2Model(config)
|
138 |
+
self.sense_network = BackpackSenseNetwork(config, self.num_senses, self.gpt2_model.wte)
|
139 |
+
self.word_embeddings = self.gpt2_model.wte
|
140 |
+
self.position_embeddings = self.gpt2_model.wpe
|
141 |
+
self.sense_weight_net = BackpackWeightNetwork(self.num_senses, self.embed_dim)
|
142 |
+
# Model parallel
|
143 |
+
self.model_parallel = False
|
144 |
+
self.device_map = None
|
145 |
+
self.gradient_checkpointing = False
|
146 |
+
|
147 |
+
def get_num_senses(self):
|
148 |
+
return self.num_senses
|
149 |
+
|
150 |
+
def get_word_embeddings(self):
|
151 |
+
return self.word_embeddings
|
152 |
+
|
153 |
+
def get_sense_network(self):
|
154 |
+
return self.sense_network
|
155 |
+
|
156 |
+
def forward(self, input_ids, position_ids):
|
157 |
+
# Compute senses
|
158 |
+
sense_input_embeds = self.word_embeddings(input_ids)
|
159 |
+
senses = self.sense_network(sense_input_embeds) # (bs, nv, s, d)
|
160 |
+
|
161 |
+
# Compute contextualization weights
|
162 |
+
contextl_hidden_states = self.gpt2_model(input_ids, position_ids=position_ids).last_hidden_state # (bs, s, d)
|
163 |
+
contextualization = self.sense_weight_net(contextl_hidden_states) # (bs, nv, s, s)
|
164 |
+
|
165 |
+
# Compute resulting outputs
|
166 |
+
hidden_states = torch.sum(contextualization @ senses, dim=1) # (bs, nv, s, d) -> (bs, s, d)
|
167 |
+
return BackpackGPT2BaseModelOutput(
|
168 |
+
hidden_states=hidden_states,
|
169 |
+
contextualization=contextualization,
|
170 |
+
)
|
171 |
+
|
172 |
+
def run_with_custom_contextualization(self, input_ids, contextualization):
|
173 |
+
# Compute senses
|
174 |
+
sense_input_embeds = self.word_embeddings(input_ids)
|
175 |
+
senses = self.sense_network(sense_input_embeds) # (bs, nv, s, d)
|
176 |
+
|
177 |
+
# Compute resulting outputs
|
178 |
+
hidden_states = torch.sum(contextualization @ senses, dim=1) # (bs, nv, s, d) -> (bs, s, d)
|
179 |
+
return BackpackGPT2BaseModelOutput(
|
180 |
+
hidden_states=hidden_states,
|
181 |
+
contextualization=contextualization,
|
182 |
+
)
|
183 |
+
|
184 |
+
@dataclass
|
185 |
+
class BackpackGPT2LMHeadModelOutput(ModelOutput):
|
186 |
+
logits: torch.FloatTensor = None
|
187 |
+
contextualization: torch.FloatTensor = None
|
188 |
+
|
189 |
+
class BackpackGPT2LMHeadModel(BackpackGPT2PreTrainedModel):
|
190 |
+
_keys_to_ignore_on_load_missing = [r".*attn.masked_bias", r".*attn.bias"]
|
191 |
+
|
192 |
+
def __init__(self, config):
|
193 |
+
super().__init__(config)
|
194 |
+
self.backpack = BackpackGPT2Model(config)
|
195 |
+
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
196 |
+
|
197 |
+
# Model parallel
|
198 |
+
self.model_parallel = False
|
199 |
+
self.device_map = None
|
200 |
+
|
201 |
+
self.tie_weights()
|
202 |
+
|
203 |
+
def tie_weights(self):
|
204 |
+
self.lm_head.weight = self.backpack.word_embeddings.weight # also tied with the underlying underlying transf
|
205 |
+
|
206 |
+
def get_lm_head(self):
|
207 |
+
return self.lm_head
|
208 |
+
|
209 |
+
def forward(self, input_ids, position_ids=None):
|
210 |
+
outputs = self.backpack(input_ids, position_ids=position_ids)
|
211 |
+
hidden_states, contextualization = outputs.hidden_states, outputs.contextualization
|
212 |
+
lm_logits = self.lm_head(hidden_states) # (bs, s, V)
|
213 |
+
return BackpackGPT2LMHeadModelOutput(
|
214 |
+
logits=lm_logits,
|
215 |
+
contextualization=contextualization,
|
216 |
+
)
|
217 |
+
|
218 |
+
def run_with_custom_contextualization(self, input_ids, contextualization):
|
219 |
+
outputs = self.backpack.run_with_custom_contextualization(input_ids, contextualization)
|
220 |
+
hidden_states, contextualization = outputs.hidden_states, outputs.contextualization
|
221 |
+
lm_logits = self.lm_head(hidden_states)
|
222 |
+
return BackpackGPT2LMHeadModelOutput(
|
223 |
+
logits=lm_logits,
|
224 |
+
contextualization=contextualization,
|
225 |
+
)
|
226 |
+
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9c0db4ac7b9af81ea53a1278a708f8fedf02f98c5ef2b70f6453b2110471f27f
|
3 |
+
size 683550781
|