Spaces:
Runtime error
Runtime error
wsntxxn
commited on
Commit
•
6065472
1
Parent(s):
35ff792
Add AudioCaps checkpoint
Browse files- app.py +97 -0
- checkpoints/audiocaps/ckpt.pth +3 -0
- checkpoints/audiocaps/config.yaml +30 -0
- models/__init__.py +92 -0
- models/base.py +504 -0
- models/cnn_encoder.py +810 -0
- models/eff_latent_encoder.py +347 -0
- models/kd_wrapper.py +226 -0
- models/transformer_decoder.py +214 -0
- models/transformer_model.py +264 -0
- requirements.txt +2 -0
- text_tokenizer.py +107 -0
- utils/model_util.py +186 -0
- utils/train_util.py +117 -0
app.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
import argparse
|
3 |
+
from functools import partial
|
4 |
+
import gradio as gr
|
5 |
+
import torch
|
6 |
+
from torchaudio.functional import resample
|
7 |
+
|
8 |
+
import utils.train_util as train_util
|
9 |
+
|
10 |
+
|
11 |
+
def load_model(cfg,
|
12 |
+
ckpt_path,
|
13 |
+
device):
|
14 |
+
model = train_util.init_model_from_config(cfg["model"])
|
15 |
+
ckpt = torch.load(ckpt_path, "cpu")
|
16 |
+
train_util.load_pretrained_model(model, ckpt)
|
17 |
+
model.eval()
|
18 |
+
model = model.to(device)
|
19 |
+
tokenizer = train_util.init_obj_from_dict(cfg["tokenizer"])
|
20 |
+
if not tokenizer.loaded:
|
21 |
+
tokenizer.load_state_dict(ckpt["tokenizer"])
|
22 |
+
model.set_index(tokenizer.bos, tokenizer.eos, tokenizer.pad)
|
23 |
+
return model, tokenizer
|
24 |
+
|
25 |
+
|
26 |
+
def infer(file, device, model, tokenizer, target_sr):
|
27 |
+
sr, wav = file
|
28 |
+
wav = torch.as_tensor(wav)
|
29 |
+
if wav.dtype == torch.short:
|
30 |
+
wav = wav / 2 ** 15
|
31 |
+
elif wav.dtype == torch.int:
|
32 |
+
wav = wav / 2 ** 31
|
33 |
+
if wav.ndim > 1:
|
34 |
+
wav = wav.mean(1)
|
35 |
+
wav = resample(wav, sr, target_sr)
|
36 |
+
wav_len = len(wav)
|
37 |
+
wav = wav.float().unsqueeze(0).to(device)
|
38 |
+
input_dict = {
|
39 |
+
"mode": "inference",
|
40 |
+
"wav": wav,
|
41 |
+
"wav_len": [wav_len],
|
42 |
+
"specaug": False,
|
43 |
+
"sample_method": "beam",
|
44 |
+
"beam_size": 3,
|
45 |
+
}
|
46 |
+
with torch.no_grad():
|
47 |
+
output_dict = model(input_dict)
|
48 |
+
seq = output_dict["seq"].cpu().numpy()
|
49 |
+
cap = tokenizer.decode(seq)[0]
|
50 |
+
return cap
|
51 |
+
|
52 |
+
# def input_toggle(input_type):
|
53 |
+
# if input_type == "file":
|
54 |
+
# return gr.update(visible=True), gr.update(visible=False)
|
55 |
+
# elif input_type == "mic":
|
56 |
+
# return gr.update(visible=False), gr.update(visible=True)
|
57 |
+
|
58 |
+
|
59 |
+
if __name__ == "__main__":
|
60 |
+
|
61 |
+
parser = argparse.ArgumentParser()
|
62 |
+
parser.add_argument("--share", action="store_true", default=False)
|
63 |
+
|
64 |
+
args = parser.parse_args()
|
65 |
+
|
66 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
67 |
+
exp_dir = Path("./checkpoints/audiocaps")
|
68 |
+
cfg = train_util.load_config(exp_dir / "config.yaml")
|
69 |
+
target_sr = cfg["target_sr"]
|
70 |
+
model, tokenizer = load_model(cfg, exp_dir / "ckpt.pth", device)
|
71 |
+
|
72 |
+
with gr.Blocks() as demo:
|
73 |
+
with gr.Row():
|
74 |
+
with gr.Column():
|
75 |
+
# radio = gr.Radio(
|
76 |
+
# ["file", "mic"],
|
77 |
+
# value="file",
|
78 |
+
# label="Select input type"
|
79 |
+
# )
|
80 |
+
file = gr.Audio(label="Input", visible=True)
|
81 |
+
# mic = gr.Microphone(label="Input", visible=False)
|
82 |
+
# radio.change(fn=input_toggle, inputs=radio, outputs=[file, mic])
|
83 |
+
btn = gr.Button("Run")
|
84 |
+
with gr.Column():
|
85 |
+
output = gr.Textbox(label="Output")
|
86 |
+
btn.click(
|
87 |
+
fn=partial(infer,
|
88 |
+
device=device,
|
89 |
+
model=model,
|
90 |
+
tokenizer=tokenizer,
|
91 |
+
target_sr=target_sr),
|
92 |
+
inputs=[file,],
|
93 |
+
outputs=output
|
94 |
+
)
|
95 |
+
|
96 |
+
demo.launch(share=args.share)
|
97 |
+
|
checkpoints/audiocaps/ckpt.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e1c435b1cf05a2b0058dae6f096c4eb4e71c685a19754ed84ea1ee812257434b
|
3 |
+
size 55293225
|
checkpoints/audiocaps/config.yaml
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
tokenizer:
|
2 |
+
type: text_tokenizer.DictTokenizer
|
3 |
+
args:
|
4 |
+
max_length: 20
|
5 |
+
|
6 |
+
target_sr: 16000
|
7 |
+
|
8 |
+
model:
|
9 |
+
args:
|
10 |
+
shared_dim: 1024
|
11 |
+
tchr_dim: 768
|
12 |
+
model:
|
13 |
+
args: {}
|
14 |
+
decoder:
|
15 |
+
args:
|
16 |
+
attn_emb_dim: 1408
|
17 |
+
dropout: 0.2
|
18 |
+
emb_dim: 256
|
19 |
+
fc_emb_dim: 1408
|
20 |
+
nlayers: 2
|
21 |
+
tie_weights: true
|
22 |
+
vocab_size: 4981
|
23 |
+
type: models.transformer_decoder.TransformerDecoder
|
24 |
+
encoder:
|
25 |
+
args:
|
26 |
+
freeze: false
|
27 |
+
pretrained: true
|
28 |
+
type: models.cnn_encoder.EfficientNetB2
|
29 |
+
type: models.transformer_model.TransformerModel
|
30 |
+
type: models.kd_wrapper.ContraEncoderKdWrapper
|
models/__init__.py
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
|
5 |
+
from utils.model_util import max_with_lens, mean_with_lens
|
6 |
+
|
7 |
+
|
8 |
+
def embedding_pooling(x, lens, pooling="mean"):
|
9 |
+
if pooling == "max":
|
10 |
+
fc_embs = max_with_lens(x, lens)
|
11 |
+
elif pooling == "mean":
|
12 |
+
fc_embs = mean_with_lens(x, lens)
|
13 |
+
elif pooling == "mean+max":
|
14 |
+
x_mean = mean_with_lens(x, lens)
|
15 |
+
x_max = max_with_lens(x, lens)
|
16 |
+
fc_embs = x_mean + x_max
|
17 |
+
elif pooling == "last":
|
18 |
+
indices = (lens - 1).reshape(-1, 1, 1).repeat(1, 1, x.size(-1))
|
19 |
+
# indices: [N, 1, hidden]
|
20 |
+
fc_embs = torch.gather(x, 1, indices).squeeze(1)
|
21 |
+
else:
|
22 |
+
raise Exception(f"pooling method {pooling} not support")
|
23 |
+
return fc_embs
|
24 |
+
|
25 |
+
|
26 |
+
class BaseEncoder(nn.Module):
|
27 |
+
|
28 |
+
"""
|
29 |
+
Encode the given audio into embedding
|
30 |
+
Base encoder class, cannot be called directly
|
31 |
+
All encoders should inherit from this class
|
32 |
+
"""
|
33 |
+
|
34 |
+
def __init__(self, spec_dim, fc_feat_dim, attn_feat_dim):
|
35 |
+
super(BaseEncoder, self).__init__()
|
36 |
+
self.spec_dim = spec_dim
|
37 |
+
self.fc_feat_dim = fc_feat_dim
|
38 |
+
self.attn_feat_dim = attn_feat_dim
|
39 |
+
|
40 |
+
|
41 |
+
def forward(self, x):
|
42 |
+
#########################
|
43 |
+
# Arguments:
|
44 |
+
# `x`: {
|
45 |
+
# (may contain)
|
46 |
+
# wav: [batch_size, n_samples],
|
47 |
+
# spec: [batch_size, n_frames, spec_dim],
|
48 |
+
# fc: [batch_size, fc_feat_dim],
|
49 |
+
# attn: [batch_size, attn_max_len, attn_feat_dim],
|
50 |
+
# attn_len: [batch_size,]
|
51 |
+
# ......
|
52 |
+
# }
|
53 |
+
#
|
54 |
+
# Returns:
|
55 |
+
# `encoded`: {
|
56 |
+
# fc_emb: [batch_size, fc_emb_dim],
|
57 |
+
# attn_emb: [batch_size, attn_max_len, attn_emb_dim],
|
58 |
+
# attn_emb_lens: [batch_size,]
|
59 |
+
# }
|
60 |
+
#########################
|
61 |
+
raise NotImplementedError
|
62 |
+
|
63 |
+
|
64 |
+
class BaseDecoder(nn.Module):
|
65 |
+
"""
|
66 |
+
Take word/audio embeddings and output the next word probs
|
67 |
+
"""
|
68 |
+
def __init__(self, emb_dim, vocab_size, fc_emb_dim,
|
69 |
+
attn_emb_dim, dropout=0.2, tie_weights=False):
|
70 |
+
super().__init__()
|
71 |
+
self.emb_dim = emb_dim
|
72 |
+
self.vocab_size = vocab_size
|
73 |
+
self.fc_emb_dim = fc_emb_dim
|
74 |
+
self.attn_emb_dim = attn_emb_dim
|
75 |
+
self.tie_weights = tie_weights
|
76 |
+
self.word_embedding = nn.Embedding(vocab_size, emb_dim)
|
77 |
+
self.in_dropout = nn.Dropout(dropout)
|
78 |
+
|
79 |
+
def forward(self, x):
|
80 |
+
raise NotImplementedError
|
81 |
+
|
82 |
+
def load_word_embedding(self, weight, freeze=True):
|
83 |
+
embedding = np.load(weight)
|
84 |
+
assert embedding.shape[0] == self.vocab_size, "vocabulary size mismatch"
|
85 |
+
assert embedding.shape[1] == self.emb_dim, "embed size mismatch"
|
86 |
+
|
87 |
+
# embeddings = torch.as_tensor(embeddings).float()
|
88 |
+
# self.word_embeddings.weight = nn.Parameter(embeddings)
|
89 |
+
# for para in self.word_embeddings.parameters():
|
90 |
+
# para.requires_grad = tune
|
91 |
+
self.word_embedding = nn.Embedding.from_pretrained(embedding,
|
92 |
+
freeze=freeze)
|
models/base.py
ADDED
@@ -0,0 +1,504 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
|
3 |
+
from typing import Dict
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
|
8 |
+
from utils.model_util import mean_with_lens, repeat_tensor
|
9 |
+
|
10 |
+
|
11 |
+
class CaptionMetaMixin:
|
12 |
+
pad_idx = 0
|
13 |
+
start_idx = 1
|
14 |
+
end_idx = 2
|
15 |
+
max_length = 20
|
16 |
+
|
17 |
+
@classmethod
|
18 |
+
def set_index(cls, start_idx, end_idx, pad_idx):
|
19 |
+
cls.start_idx = start_idx
|
20 |
+
cls.end_idx = end_idx
|
21 |
+
cls.pad_idx = pad_idx
|
22 |
+
|
23 |
+
|
24 |
+
class CaptionModel(nn.Module, CaptionMetaMixin):
|
25 |
+
"""
|
26 |
+
Encoder-decoder captioning model.
|
27 |
+
"""
|
28 |
+
|
29 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
30 |
+
super().__init__()
|
31 |
+
self.encoder = encoder
|
32 |
+
self.decoder = decoder
|
33 |
+
self.vocab_size = decoder.vocab_size
|
34 |
+
self.train_forward_keys = ["cap", "cap_len", "ss_ratio"]
|
35 |
+
self.inference_forward_keys = ["sample_method", "max_length", "temp"]
|
36 |
+
freeze_encoder = kwargs.get("freeze_encoder", False)
|
37 |
+
if freeze_encoder:
|
38 |
+
for param in self.encoder.parameters():
|
39 |
+
param.requires_grad = False
|
40 |
+
self.check_decoder_compatibility()
|
41 |
+
|
42 |
+
def check_decoder_compatibility(self):
|
43 |
+
compatible_decoders = [x.__class__.__name__ for x in self.compatible_decoders]
|
44 |
+
assert isinstance(self.decoder, self.compatible_decoders), \
|
45 |
+
f"{self.decoder.__class__.__name__} is incompatible with " \
|
46 |
+
f"{self.__class__.__name__}, please use decoder in {compatible_decoders} "
|
47 |
+
|
48 |
+
def forward(self, input_dict: Dict):
|
49 |
+
"""
|
50 |
+
input_dict: {
|
51 |
+
(required)
|
52 |
+
mode: train/inference,
|
53 |
+
[spec, spec_len],
|
54 |
+
[fc],
|
55 |
+
[attn, attn_len],
|
56 |
+
[wav, wav_len],
|
57 |
+
[sample_method: greedy],
|
58 |
+
[temp: 1.0] (in case of no teacher forcing)
|
59 |
+
|
60 |
+
(optional, mode=train)
|
61 |
+
cap,
|
62 |
+
cap_len,
|
63 |
+
ss_ratio,
|
64 |
+
|
65 |
+
(optional, mode=inference)
|
66 |
+
sample_method: greedy/beam,
|
67 |
+
max_length,
|
68 |
+
temp,
|
69 |
+
beam_size (optional, sample_method=beam),
|
70 |
+
n_best (optional, sample_method=beam),
|
71 |
+
}
|
72 |
+
"""
|
73 |
+
encoder_output_dict = self.encoder(input_dict)
|
74 |
+
output = self.forward_decoder(input_dict, encoder_output_dict)
|
75 |
+
return output
|
76 |
+
|
77 |
+
def forward_decoder(self, input_dict: Dict, encoder_output_dict: Dict):
|
78 |
+
if input_dict["mode"] == "train":
|
79 |
+
forward_dict = {
|
80 |
+
"mode": "train", "sample_method": "greedy", "temp": 1.0
|
81 |
+
}
|
82 |
+
for key in self.train_forward_keys:
|
83 |
+
forward_dict[key] = input_dict[key]
|
84 |
+
forward_dict.update(encoder_output_dict)
|
85 |
+
output = self.train_forward(forward_dict)
|
86 |
+
elif input_dict["mode"] == "inference":
|
87 |
+
forward_dict = {"mode": "inference"}
|
88 |
+
default_args = { "sample_method": "greedy", "max_length": self.max_length, "temp": 1.0 }
|
89 |
+
for key in self.inference_forward_keys:
|
90 |
+
if key in input_dict:
|
91 |
+
forward_dict[key] = input_dict[key]
|
92 |
+
else:
|
93 |
+
forward_dict[key] = default_args[key]
|
94 |
+
|
95 |
+
if forward_dict["sample_method"] == "beam":
|
96 |
+
forward_dict["beam_size"] = input_dict.get("beam_size", 3)
|
97 |
+
forward_dict["n_best"] = input_dict.get("n_best", False)
|
98 |
+
forward_dict["n_best_size"] = input_dict.get("n_best_size", forward_dict["beam_size"])
|
99 |
+
elif forward_dict["sample_method"] == "dbs":
|
100 |
+
forward_dict["beam_size"] = input_dict.get("beam_size", 6)
|
101 |
+
forward_dict["group_size"] = input_dict.get("group_size", 3)
|
102 |
+
forward_dict["diversity_lambda"] = input_dict.get("diversity_lambda", 0.5)
|
103 |
+
forward_dict["group_nbest"] = input_dict.get("group_nbest", True)
|
104 |
+
|
105 |
+
forward_dict.update(encoder_output_dict)
|
106 |
+
output = self.inference_forward(forward_dict)
|
107 |
+
else:
|
108 |
+
raise Exception("mode should be either 'train' or 'inference'")
|
109 |
+
output.update(encoder_output_dict)
|
110 |
+
return output
|
111 |
+
|
112 |
+
def prepare_output(self, input_dict):
|
113 |
+
output = {}
|
114 |
+
batch_size = input_dict["fc_emb"].size(0)
|
115 |
+
if input_dict["mode"] == "train":
|
116 |
+
max_length = input_dict["cap"].size(1) - 1
|
117 |
+
elif input_dict["mode"] == "inference":
|
118 |
+
max_length = input_dict["max_length"]
|
119 |
+
else:
|
120 |
+
raise Exception("mode should be either 'train' or 'inference'")
|
121 |
+
device = input_dict["fc_emb"].device
|
122 |
+
output["seq"] = torch.full((batch_size, max_length), self.end_idx,
|
123 |
+
dtype=torch.long)
|
124 |
+
output["logit"] = torch.empty(batch_size, max_length,
|
125 |
+
self.vocab_size).to(device)
|
126 |
+
output["sampled_logprob"] = torch.zeros(batch_size, max_length)
|
127 |
+
output["embed"] = torch.empty(batch_size, max_length,
|
128 |
+
self.decoder.d_model).to(device)
|
129 |
+
return output
|
130 |
+
|
131 |
+
def train_forward(self, input_dict):
|
132 |
+
if input_dict["ss_ratio"] != 1: # scheduled sampling training
|
133 |
+
input_dict["mode"] = "train"
|
134 |
+
return self.stepwise_forward(input_dict)
|
135 |
+
output = self.seq_forward(input_dict)
|
136 |
+
self.train_process(output, input_dict)
|
137 |
+
return output
|
138 |
+
|
139 |
+
def seq_forward(self, input_dict):
|
140 |
+
raise NotImplementedError
|
141 |
+
|
142 |
+
def train_process(self, output, input_dict):
|
143 |
+
pass
|
144 |
+
|
145 |
+
def inference_forward(self, input_dict):
|
146 |
+
if input_dict["sample_method"] == "beam":
|
147 |
+
return self.beam_search(input_dict)
|
148 |
+
elif input_dict["sample_method"] == "dbs":
|
149 |
+
return self.diverse_beam_search(input_dict)
|
150 |
+
return self.stepwise_forward(input_dict)
|
151 |
+
|
152 |
+
def stepwise_forward(self, input_dict):
|
153 |
+
"""Step-by-step decoding"""
|
154 |
+
output = self.prepare_output(input_dict)
|
155 |
+
max_length = output["seq"].size(1)
|
156 |
+
# start sampling
|
157 |
+
for t in range(max_length):
|
158 |
+
input_dict["t"] = t
|
159 |
+
self.decode_step(input_dict, output)
|
160 |
+
if input_dict["mode"] == "inference": # decide whether to stop when sampling
|
161 |
+
unfinished_t = output["seq"][:, t] != self.end_idx
|
162 |
+
if t == 0:
|
163 |
+
unfinished = unfinished_t
|
164 |
+
else:
|
165 |
+
unfinished *= unfinished_t
|
166 |
+
output["seq"][:, t][~unfinished] = self.end_idx
|
167 |
+
if unfinished.sum() == 0:
|
168 |
+
break
|
169 |
+
self.stepwise_process(output)
|
170 |
+
return output
|
171 |
+
|
172 |
+
def decode_step(self, input_dict, output):
|
173 |
+
"""Decoding operation of timestep t"""
|
174 |
+
decoder_input = self.prepare_decoder_input(input_dict, output)
|
175 |
+
# feed to the decoder to get logit
|
176 |
+
output_t = self.decoder(decoder_input)
|
177 |
+
logit_t = output_t["logit"]
|
178 |
+
# assert logit_t.ndim == 3
|
179 |
+
if logit_t.size(1) == 1:
|
180 |
+
logit_t = logit_t.squeeze(1)
|
181 |
+
embed_t = output_t["embed"].squeeze(1)
|
182 |
+
elif logit_t.size(1) > 1:
|
183 |
+
logit_t = logit_t[:, -1, :]
|
184 |
+
embed_t = output_t["embed"][:, -1, :]
|
185 |
+
else:
|
186 |
+
raise Exception("no logit output")
|
187 |
+
# sample the next input word and get the corresponding logit
|
188 |
+
sampled = self.sample_next_word(logit_t,
|
189 |
+
method=input_dict["sample_method"],
|
190 |
+
temp=input_dict["temp"])
|
191 |
+
|
192 |
+
output_t.update(sampled)
|
193 |
+
output_t["t"] = input_dict["t"]
|
194 |
+
output_t["logit"] = logit_t
|
195 |
+
output_t["embed"] = embed_t
|
196 |
+
self.stepwise_process_step(output, output_t)
|
197 |
+
|
198 |
+
def prepare_decoder_input(self, input_dict, output):
|
199 |
+
"""Prepare the inp ut dict for the decoder"""
|
200 |
+
raise NotImplementedError
|
201 |
+
|
202 |
+
def stepwise_process_step(self, output, output_t):
|
203 |
+
"""Postprocessing (save output values) after each timestep t"""
|
204 |
+
t = output_t["t"]
|
205 |
+
output["logit"][:, t, :] = output_t["logit"]
|
206 |
+
output["seq"][:, t] = output_t["word"]
|
207 |
+
output["sampled_logprob"][:, t] = output_t["probs"]
|
208 |
+
output["embed"][:, t, :] = output_t["embed"]
|
209 |
+
|
210 |
+
def stepwise_process(self, output):
|
211 |
+
"""Postprocessing after the whole step-by-step autoregressive decoding"""
|
212 |
+
pass
|
213 |
+
|
214 |
+
def sample_next_word(self, logit, method, temp):
|
215 |
+
"""Sample the next word, given probs output by the decoder"""
|
216 |
+
logprob = torch.log_softmax(logit, dim=1)
|
217 |
+
if method == "greedy":
|
218 |
+
sampled_logprob, word = torch.max(logprob.detach(), 1)
|
219 |
+
elif method == "gumbel":
|
220 |
+
def sample_gumbel(shape, eps=1e-20):
|
221 |
+
U = torch.rand(shape).to(logprob.device)
|
222 |
+
return -torch.log(-torch.log(U + eps) + eps)
|
223 |
+
def gumbel_softmax_sample(logit, temperature):
|
224 |
+
y = logit + sample_gumbel(logit.size())
|
225 |
+
return torch.log_softmax(y / temperature, dim=-1)
|
226 |
+
_logprob = gumbel_softmax_sample(logprob, temp)
|
227 |
+
_, word = torch.max(_logprob.data, 1)
|
228 |
+
sampled_logprob = logprob.gather(1, word.unsqueeze(-1))
|
229 |
+
else:
|
230 |
+
logprob = logprob / temp
|
231 |
+
if method.startswith("top"):
|
232 |
+
top_num = float(method[3:])
|
233 |
+
if 0 < top_num < 1: # top-p sampling
|
234 |
+
probs = torch.softmax(logit, dim=1)
|
235 |
+
sorted_probs, sorted_indices = torch.sort(probs, descending=True, dim=1)
|
236 |
+
_cumsum = sorted_probs.cumsum(1)
|
237 |
+
mask = _cumsum < top_num
|
238 |
+
mask = torch.cat([torch.ones_like(mask[:,:1]), mask[:,:-1]], 1)
|
239 |
+
sorted_probs = sorted_probs * mask.to(sorted_probs)
|
240 |
+
sorted_probs = sorted_probs / sorted_probs.sum(1, keepdim=True)
|
241 |
+
logprob.scatter_(1, sorted_indices, sorted_probs.log())
|
242 |
+
else: # top-k sampling
|
243 |
+
k = int(top_num)
|
244 |
+
tmp = torch.empty_like(logprob).fill_(float('-inf'))
|
245 |
+
topk, indices = torch.topk(logprob, k, dim=1)
|
246 |
+
tmp = tmp.scatter(1, indices, topk)
|
247 |
+
logprob = tmp
|
248 |
+
word = torch.distributions.Categorical(logits=logprob.detach()).sample()
|
249 |
+
sampled_logprob = logprob.gather(1, word.unsqueeze(-1)).squeeze(1)
|
250 |
+
word = word.detach().long()
|
251 |
+
# sampled_logprob: [N,], word: [N,]
|
252 |
+
return {"word": word, "probs": sampled_logprob}
|
253 |
+
|
254 |
+
def beam_search(self, input_dict):
|
255 |
+
output = self.prepare_output(input_dict)
|
256 |
+
max_length = input_dict["max_length"]
|
257 |
+
beam_size = input_dict["beam_size"]
|
258 |
+
if input_dict["n_best"]:
|
259 |
+
n_best_size = input_dict["n_best_size"]
|
260 |
+
batch_size, max_length = output["seq"].size()
|
261 |
+
output["seq"] = torch.full((batch_size, n_best_size, max_length),
|
262 |
+
self.end_idx, dtype=torch.long)
|
263 |
+
|
264 |
+
temp = input_dict["temp"]
|
265 |
+
# instance by instance beam seach
|
266 |
+
for i in range(output["seq"].size(0)):
|
267 |
+
output_i = self.prepare_beamsearch_output(input_dict)
|
268 |
+
input_dict["sample_idx"] = i
|
269 |
+
for t in range(max_length):
|
270 |
+
input_dict["t"] = t
|
271 |
+
output_t = self.beamsearch_step(input_dict, output_i)
|
272 |
+
#######################################
|
273 |
+
# merge with previous beam and select the current max prob beam
|
274 |
+
#######################################
|
275 |
+
logit_t = output_t["logit"]
|
276 |
+
if logit_t.size(1) == 1:
|
277 |
+
logit_t = logit_t.squeeze(1)
|
278 |
+
elif logit_t.size(1) > 1:
|
279 |
+
logit_t = logit_t[:, -1, :]
|
280 |
+
else:
|
281 |
+
raise Exception("no logit output")
|
282 |
+
logprob_t = torch.log_softmax(logit_t, dim=1)
|
283 |
+
logprob_t = torch.log_softmax(logprob_t / temp, dim=1)
|
284 |
+
logprob_t = output_i["topk_logprob"].unsqueeze(1) + logprob_t
|
285 |
+
if t == 0: # for the first step, all k seq will have the same probs
|
286 |
+
topk_logprob, topk_words = logprob_t[0].topk(
|
287 |
+
beam_size, 0, True, True)
|
288 |
+
else: # unroll and find top logprob, and their unrolled indices
|
289 |
+
topk_logprob, topk_words = logprob_t.view(-1).topk(
|
290 |
+
beam_size, 0, True, True)
|
291 |
+
topk_words = topk_words.cpu()
|
292 |
+
output_i["topk_logprob"] = topk_logprob
|
293 |
+
# output_i["prev_words_beam"] = topk_words // self.vocab_size # [beam_size,]
|
294 |
+
output_i["prev_words_beam"] = torch.div(topk_words, self.vocab_size,
|
295 |
+
rounding_mode='trunc')
|
296 |
+
output_i["next_word"] = topk_words % self.vocab_size # [beam_size,]
|
297 |
+
if t == 0:
|
298 |
+
output_i["seq"] = output_i["next_word"].unsqueeze(1)
|
299 |
+
else:
|
300 |
+
output_i["seq"] = torch.cat([
|
301 |
+
output_i["seq"][output_i["prev_words_beam"]],
|
302 |
+
output_i["next_word"].unsqueeze(1)], dim=1)
|
303 |
+
|
304 |
+
# add finished beams to results
|
305 |
+
is_end = output_i["next_word"] == self.end_idx
|
306 |
+
if t == max_length - 1:
|
307 |
+
is_end.fill_(1)
|
308 |
+
|
309 |
+
for beam_idx in range(beam_size):
|
310 |
+
if is_end[beam_idx]:
|
311 |
+
final_beam = {
|
312 |
+
"seq": output_i["seq"][beam_idx].clone(),
|
313 |
+
"score": output_i["topk_logprob"][beam_idx].item()
|
314 |
+
}
|
315 |
+
final_beam["score"] = final_beam["score"] / (t + 1)
|
316 |
+
output_i["done_beams"].append(final_beam)
|
317 |
+
output_i["topk_logprob"][is_end] -= 1000
|
318 |
+
|
319 |
+
self.beamsearch_process_step(output_i, output_t)
|
320 |
+
|
321 |
+
self.beamsearch_process(output, output_i, input_dict)
|
322 |
+
return output
|
323 |
+
|
324 |
+
def prepare_beamsearch_output(self, input_dict):
|
325 |
+
beam_size = input_dict["beam_size"]
|
326 |
+
device = input_dict["fc_emb"].device
|
327 |
+
output = {
|
328 |
+
"topk_logprob": torch.zeros(beam_size).to(device),
|
329 |
+
"seq": None,
|
330 |
+
"prev_words_beam": None,
|
331 |
+
"next_word": None,
|
332 |
+
"done_beams": [],
|
333 |
+
}
|
334 |
+
return output
|
335 |
+
|
336 |
+
def beamsearch_step(self, input_dict, output_i):
|
337 |
+
decoder_input = self.prepare_beamsearch_decoder_input(input_dict, output_i)
|
338 |
+
output_t = self.decoder(decoder_input)
|
339 |
+
output_t["t"] = input_dict["t"]
|
340 |
+
return output_t
|
341 |
+
|
342 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
343 |
+
raise NotImplementedError
|
344 |
+
|
345 |
+
def beamsearch_process_step(self, output_i, output_t):
|
346 |
+
pass
|
347 |
+
|
348 |
+
def beamsearch_process(self, output, output_i, input_dict):
|
349 |
+
i = input_dict["sample_idx"]
|
350 |
+
done_beams = sorted(output_i["done_beams"], key=lambda x: -x["score"])
|
351 |
+
if input_dict["n_best"]:
|
352 |
+
done_beams = done_beams[:input_dict["n_best_size"]]
|
353 |
+
for out_idx, done_beam in enumerate(done_beams):
|
354 |
+
seq = done_beam["seq"]
|
355 |
+
output["seq"][i][out_idx, :len(seq)] = seq
|
356 |
+
else:
|
357 |
+
seq = done_beams[0]["seq"]
|
358 |
+
output["seq"][i][:len(seq)] = seq
|
359 |
+
|
360 |
+
def diverse_beam_search(self, input_dict):
|
361 |
+
|
362 |
+
def add_diversity(seq_table, logprob, t, divm, diversity_lambda, bdash):
|
363 |
+
local_time = t - divm
|
364 |
+
unaug_logprob = logprob.clone()
|
365 |
+
|
366 |
+
if divm > 0:
|
367 |
+
change = torch.zeros(logprob.size(-1))
|
368 |
+
for prev_choice in range(divm):
|
369 |
+
prev_decisions = seq_table[prev_choice][..., local_time]
|
370 |
+
for prev_labels in range(bdash):
|
371 |
+
change.scatter_add_(0, prev_decisions[prev_labels], change.new_ones(1))
|
372 |
+
|
373 |
+
change = change.to(logprob.device)
|
374 |
+
logprob = logprob - repeat_tensor(change, bdash) * diversity_lambda
|
375 |
+
|
376 |
+
return logprob, unaug_logprob
|
377 |
+
|
378 |
+
output = self.prepare_output(input_dict)
|
379 |
+
group_size = input_dict["group_size"]
|
380 |
+
batch_size = output["seq"].size(0)
|
381 |
+
beam_size = input_dict["beam_size"]
|
382 |
+
bdash = beam_size // group_size
|
383 |
+
input_dict["bdash"] = bdash
|
384 |
+
diversity_lambda = input_dict["diversity_lambda"]
|
385 |
+
device = input_dict["fc_emb"].device
|
386 |
+
max_length = input_dict["max_length"]
|
387 |
+
temp = input_dict["temp"]
|
388 |
+
group_nbest = input_dict["group_nbest"]
|
389 |
+
batch_size, max_length = output["seq"].size()
|
390 |
+
if group_nbest:
|
391 |
+
output["seq"] = torch.full((batch_size, beam_size, max_length),
|
392 |
+
self.end_idx, dtype=torch.long)
|
393 |
+
else:
|
394 |
+
output["seq"] = torch.full((batch_size, group_size, max_length),
|
395 |
+
self.end_idx, dtype=torch.long)
|
396 |
+
|
397 |
+
|
398 |
+
for i in range(batch_size):
|
399 |
+
input_dict["sample_idx"] = i
|
400 |
+
seq_table = [torch.LongTensor(bdash, 0) for _ in range(group_size)] # group_size x [bdash, 0]
|
401 |
+
logprob_table = [torch.zeros(bdash).to(device) for _ in range(group_size)]
|
402 |
+
done_beams_table = [[] for _ in range(group_size)]
|
403 |
+
|
404 |
+
output_i = {
|
405 |
+
"prev_words_beam": [None for _ in range(group_size)],
|
406 |
+
"next_word": [None for _ in range(group_size)],
|
407 |
+
"state": [None for _ in range(group_size)]
|
408 |
+
}
|
409 |
+
|
410 |
+
for t in range(max_length + group_size - 1):
|
411 |
+
input_dict["t"] = t
|
412 |
+
for divm in range(group_size):
|
413 |
+
input_dict["divm"] = divm
|
414 |
+
if t >= divm and t <= max_length + divm - 1:
|
415 |
+
local_time = t - divm
|
416 |
+
decoder_input = self.prepare_dbs_decoder_input(input_dict, output_i)
|
417 |
+
output_t = self.decoder(decoder_input)
|
418 |
+
output_t["divm"] = divm
|
419 |
+
logit_t = output_t["logit"]
|
420 |
+
if logit_t.size(1) == 1:
|
421 |
+
logit_t = logit_t.squeeze(1)
|
422 |
+
elif logit_t.size(1) > 1:
|
423 |
+
logit_t = logit_t[:, -1, :]
|
424 |
+
else:
|
425 |
+
raise Exception("no logit output")
|
426 |
+
logprob_t = torch.log_softmax(logit_t, dim=1)
|
427 |
+
logprob_t = torch.log_softmax(logprob_t / temp, dim=1)
|
428 |
+
logprob_t, unaug_logprob_t = add_diversity(seq_table, logprob_t, t, divm, diversity_lambda, bdash)
|
429 |
+
logprob_t = logprob_table[divm].unsqueeze(-1) + logprob_t
|
430 |
+
if local_time == 0: # for the first step, all k seq will have the same probs
|
431 |
+
topk_logprob, topk_words = logprob_t[0].topk(
|
432 |
+
bdash, 0, True, True)
|
433 |
+
else: # unroll and find top logprob, and their unrolled indices
|
434 |
+
topk_logprob, topk_words = logprob_t.view(-1).topk(
|
435 |
+
bdash, 0, True, True)
|
436 |
+
topk_words = topk_words.cpu()
|
437 |
+
logprob_table[divm] = topk_logprob
|
438 |
+
output_i["prev_words_beam"][divm] = topk_words // self.vocab_size # [bdash,]
|
439 |
+
output_i["next_word"][divm] = topk_words % self.vocab_size # [bdash,]
|
440 |
+
if local_time > 0:
|
441 |
+
seq_table[divm] = seq_table[divm][output_i["prev_words_beam"][divm]]
|
442 |
+
seq_table[divm] = torch.cat([
|
443 |
+
seq_table[divm],
|
444 |
+
output_i["next_word"][divm].unsqueeze(-1)], -1)
|
445 |
+
|
446 |
+
is_end = seq_table[divm][:, t-divm] == self.end_idx
|
447 |
+
assert seq_table[divm].shape[-1] == t - divm + 1
|
448 |
+
if t == max_length + divm - 1:
|
449 |
+
is_end.fill_(1)
|
450 |
+
for beam_idx in range(bdash):
|
451 |
+
if is_end[beam_idx]:
|
452 |
+
final_beam = {
|
453 |
+
"seq": seq_table[divm][beam_idx].clone(),
|
454 |
+
"score": logprob_table[divm][beam_idx].item()
|
455 |
+
}
|
456 |
+
final_beam["score"] = final_beam["score"] / (t - divm + 1)
|
457 |
+
done_beams_table[divm].append(final_beam)
|
458 |
+
logprob_table[divm][is_end] -= 1000
|
459 |
+
self.dbs_process_step(output_i, output_t)
|
460 |
+
done_beams_table = [sorted(done_beams_table[divm], key=lambda x: -x["score"])[:bdash] for divm in range(group_size)]
|
461 |
+
if group_nbest:
|
462 |
+
done_beams = sum(done_beams_table, [])
|
463 |
+
else:
|
464 |
+
done_beams = [group_beam[0] for group_beam in done_beams_table]
|
465 |
+
for _, done_beam in enumerate(done_beams):
|
466 |
+
output["seq"][i, _, :len(done_beam["seq"])] = done_beam["seq"]
|
467 |
+
|
468 |
+
return output
|
469 |
+
|
470 |
+
def prepare_dbs_decoder_input(self, input_dict, output_i):
|
471 |
+
raise NotImplementedError
|
472 |
+
|
473 |
+
def dbs_process_step(self, output_i, output_t):
|
474 |
+
pass
|
475 |
+
|
476 |
+
|
477 |
+
class CaptionSequenceModel(nn.Module, CaptionMetaMixin):
|
478 |
+
|
479 |
+
def __init__(self, model, seq_output_size):
|
480 |
+
super().__init__()
|
481 |
+
self.model = model
|
482 |
+
if model.decoder.d_model != seq_output_size:
|
483 |
+
self.output_transform = nn.Linear(model.decoder.d_model, seq_output_size)
|
484 |
+
else:
|
485 |
+
self.output_transform = lambda x: x
|
486 |
+
|
487 |
+
def forward(self, input_dict):
|
488 |
+
output = self.model(input_dict)
|
489 |
+
|
490 |
+
if input_dict["mode"] == "train":
|
491 |
+
lens = input_dict["cap_len"] - 1
|
492 |
+
# seq_outputs: [N, d_model]
|
493 |
+
elif input_dict["mode"] == "inference":
|
494 |
+
if "sample_method" in input_dict and input_dict["sample_method"] == "beam":
|
495 |
+
return output
|
496 |
+
seq = output["seq"]
|
497 |
+
lens = torch.where(seq == self.model.end_idx, torch.zeros_like(seq), torch.ones_like(seq)).sum(dim=1)
|
498 |
+
else:
|
499 |
+
raise Exception("mode should be either 'train' or 'inference'")
|
500 |
+
seq_output = mean_with_lens(output["embed"], lens)
|
501 |
+
seq_output = self.output_transform(seq_output)
|
502 |
+
output["seq_output"] = seq_output
|
503 |
+
return output
|
504 |
+
|
models/cnn_encoder.py
ADDED
@@ -0,0 +1,810 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
|
3 |
+
from einops import rearrange
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import torch.nn.functional as F
|
8 |
+
from torchaudio import transforms
|
9 |
+
|
10 |
+
from utils.model_util import mean_with_lens, max_with_lens
|
11 |
+
from utils.train_util import merge_load_state_dict
|
12 |
+
|
13 |
+
|
14 |
+
def init_layer(layer):
|
15 |
+
"""Initialize a Linear or Convolutional layer. """
|
16 |
+
nn.init.xavier_uniform_(layer.weight)
|
17 |
+
|
18 |
+
if hasattr(layer, 'bias'):
|
19 |
+
if layer.bias is not None:
|
20 |
+
layer.bias.data.fill_(0.)
|
21 |
+
|
22 |
+
|
23 |
+
def init_bn(bn):
|
24 |
+
"""Initialize a Batchnorm layer. """
|
25 |
+
bn.bias.data.fill_(0.)
|
26 |
+
bn.weight.data.fill_(1.)
|
27 |
+
|
28 |
+
|
29 |
+
class ConvBlock(nn.Module):
|
30 |
+
def __init__(self, in_channels, out_channels):
|
31 |
+
|
32 |
+
super(ConvBlock, self).__init__()
|
33 |
+
|
34 |
+
self.conv1 = nn.Conv2d(in_channels=in_channels,
|
35 |
+
out_channels=out_channels,
|
36 |
+
kernel_size=(3, 3), stride=(1, 1),
|
37 |
+
padding=(1, 1), bias=False)
|
38 |
+
|
39 |
+
self.conv2 = nn.Conv2d(in_channels=out_channels,
|
40 |
+
out_channels=out_channels,
|
41 |
+
kernel_size=(3, 3), stride=(1, 1),
|
42 |
+
padding=(1, 1), bias=False)
|
43 |
+
|
44 |
+
self.bn1 = nn.BatchNorm2d(out_channels)
|
45 |
+
self.bn2 = nn.BatchNorm2d(out_channels)
|
46 |
+
|
47 |
+
self.init_weight()
|
48 |
+
|
49 |
+
def init_weight(self):
|
50 |
+
init_layer(self.conv1)
|
51 |
+
init_layer(self.conv2)
|
52 |
+
init_bn(self.bn1)
|
53 |
+
init_bn(self.bn2)
|
54 |
+
|
55 |
+
|
56 |
+
def forward(self, input, pool_size=(2, 2), pool_type='avg'):
|
57 |
+
|
58 |
+
x = input
|
59 |
+
x = F.relu_(self.bn1(self.conv1(x)))
|
60 |
+
x = F.relu_(self.bn2(self.conv2(x)))
|
61 |
+
if pool_type == 'max':
|
62 |
+
x = F.max_pool2d(x, kernel_size=pool_size)
|
63 |
+
elif pool_type == 'avg':
|
64 |
+
x = F.avg_pool2d(x, kernel_size=pool_size)
|
65 |
+
elif pool_type == 'avg+max':
|
66 |
+
x1 = F.avg_pool2d(x, kernel_size=pool_size)
|
67 |
+
x2 = F.max_pool2d(x, kernel_size=pool_size)
|
68 |
+
x = x1 + x2
|
69 |
+
else:
|
70 |
+
raise Exception('Incorrect argument!')
|
71 |
+
|
72 |
+
return x
|
73 |
+
|
74 |
+
|
75 |
+
class ConvBlock5x5(nn.Module):
|
76 |
+
def __init__(self, in_channels, out_channels):
|
77 |
+
|
78 |
+
super(ConvBlock5x5, self).__init__()
|
79 |
+
|
80 |
+
self.conv1 = nn.Conv2d(in_channels=in_channels,
|
81 |
+
out_channels=out_channels,
|
82 |
+
kernel_size=(5, 5), stride=(1, 1),
|
83 |
+
padding=(2, 2), bias=False)
|
84 |
+
|
85 |
+
self.bn1 = nn.BatchNorm2d(out_channels)
|
86 |
+
|
87 |
+
self.init_weight()
|
88 |
+
|
89 |
+
def init_weight(self):
|
90 |
+
init_layer(self.conv1)
|
91 |
+
init_bn(self.bn1)
|
92 |
+
|
93 |
+
def forward(self, input, pool_size=(2, 2), pool_type='avg'):
|
94 |
+
|
95 |
+
x = input
|
96 |
+
x = F.relu_(self.bn1(self.conv1(x)))
|
97 |
+
if pool_type == 'max':
|
98 |
+
x = F.max_pool2d(x, kernel_size=pool_size)
|
99 |
+
elif pool_type == 'avg':
|
100 |
+
x = F.avg_pool2d(x, kernel_size=pool_size)
|
101 |
+
elif pool_type == 'avg+max':
|
102 |
+
x1 = F.avg_pool2d(x, kernel_size=pool_size)
|
103 |
+
x2 = F.max_pool2d(x, kernel_size=pool_size)
|
104 |
+
x = x1 + x2
|
105 |
+
else:
|
106 |
+
raise Exception('Incorrect argument!')
|
107 |
+
|
108 |
+
return x
|
109 |
+
|
110 |
+
|
111 |
+
class Cnn6Encoder(nn.Module):
|
112 |
+
|
113 |
+
def __init__(self, sample_rate=32000, freeze=False):
|
114 |
+
super().__init__()
|
115 |
+
|
116 |
+
sr_to_fmax = {
|
117 |
+
32000: 14000,
|
118 |
+
16000: 8000
|
119 |
+
}
|
120 |
+
# Logmel spectrogram extractor
|
121 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
122 |
+
sample_rate=sample_rate,
|
123 |
+
n_fft=32 * sample_rate // 1000,
|
124 |
+
win_length=32 * sample_rate // 1000,
|
125 |
+
hop_length=10 * sample_rate // 1000,
|
126 |
+
f_min=50,
|
127 |
+
f_max=sr_to_fmax[sample_rate],
|
128 |
+
n_mels=64,
|
129 |
+
norm="slaney",
|
130 |
+
mel_scale="slaney"
|
131 |
+
)
|
132 |
+
self.hop_length = 10 * sample_rate // 1000
|
133 |
+
self.db_transform = transforms.AmplitudeToDB()
|
134 |
+
|
135 |
+
self.bn0 = nn.BatchNorm2d(64)
|
136 |
+
|
137 |
+
self.conv_block1 = ConvBlock5x5(in_channels=1, out_channels=64)
|
138 |
+
self.conv_block2 = ConvBlock5x5(in_channels=64, out_channels=128)
|
139 |
+
self.conv_block3 = ConvBlock5x5(in_channels=128, out_channels=256)
|
140 |
+
self.conv_block4 = ConvBlock5x5(in_channels=256, out_channels=512)
|
141 |
+
|
142 |
+
self.downsample_ratio = 16
|
143 |
+
|
144 |
+
self.fc1 = nn.Linear(512, 512, bias=True)
|
145 |
+
self.fc_emb_size = 512
|
146 |
+
self.init_weight()
|
147 |
+
self.freeze = freeze
|
148 |
+
|
149 |
+
def init_weight(self):
|
150 |
+
init_bn(self.bn0)
|
151 |
+
init_layer(self.fc1)
|
152 |
+
|
153 |
+
def load_pretrained(self, pretrained, output_fn):
|
154 |
+
checkpoint = torch.load(pretrained, map_location="cpu")
|
155 |
+
|
156 |
+
if "model" in checkpoint:
|
157 |
+
state_dict = checkpoint["model"]
|
158 |
+
else:
|
159 |
+
raise Exception("Unkown checkpoint format")
|
160 |
+
|
161 |
+
loaded_keys = merge_load_state_dict(state_dict, self, output_fn)
|
162 |
+
if self.freeze:
|
163 |
+
for name, param in self.named_parameters():
|
164 |
+
if name in loaded_keys:
|
165 |
+
param.requires_grad = False
|
166 |
+
else:
|
167 |
+
param.requires_grad = True
|
168 |
+
|
169 |
+
def forward(self, input_dict):
|
170 |
+
waveform = input_dict["wav"]
|
171 |
+
wave_length = input_dict["wav_len"]
|
172 |
+
specaug = input_dict["specaug"]
|
173 |
+
x = self.melspec_extractor(waveform)
|
174 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
175 |
+
x = x.transpose(1, 2)
|
176 |
+
x = x.unsqueeze(1) # (batch_size, 1, time_steps, mel_bins)
|
177 |
+
|
178 |
+
x = x.transpose(1, 3)
|
179 |
+
x = self.bn0(x)
|
180 |
+
x = x.transpose(1, 3)
|
181 |
+
|
182 |
+
x = self.conv_block1(x, pool_size=(2, 2), pool_type='avg')
|
183 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
184 |
+
x = self.conv_block2(x, pool_size=(2, 2), pool_type='avg')
|
185 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
186 |
+
x = self.conv_block3(x, pool_size=(2, 2), pool_type='avg')
|
187 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
188 |
+
x = self.conv_block4(x, pool_size=(2, 2), pool_type='avg')
|
189 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
190 |
+
|
191 |
+
x = torch.mean(x, dim=3)
|
192 |
+
attn_emb = x.transpose(1, 2)
|
193 |
+
wave_length = torch.as_tensor(wave_length)
|
194 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
195 |
+
rounding_mode="floor") + 1
|
196 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
197 |
+
rounding_mode="floor")
|
198 |
+
x_max = max_with_lens(attn_emb, feat_length)
|
199 |
+
x_mean = mean_with_lens(attn_emb, feat_length)
|
200 |
+
x = x_max + x_mean
|
201 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
202 |
+
x = F.relu_(self.fc1(x))
|
203 |
+
fc_emb = F.dropout(x, p=0.5, training=self.training)
|
204 |
+
|
205 |
+
return {
|
206 |
+
"attn_emb": attn_emb,
|
207 |
+
"fc_emb": fc_emb,
|
208 |
+
"attn_emb_len": feat_length
|
209 |
+
}
|
210 |
+
|
211 |
+
|
212 |
+
class Cnn10Encoder(nn.Module):
|
213 |
+
|
214 |
+
def __init__(self, sample_rate=32000, freeze=False):
|
215 |
+
super().__init__()
|
216 |
+
|
217 |
+
sr_to_fmax = {
|
218 |
+
32000: 14000,
|
219 |
+
16000: 8000
|
220 |
+
}
|
221 |
+
# Logmel spectrogram extractor
|
222 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
223 |
+
sample_rate=sample_rate,
|
224 |
+
n_fft=32 * sample_rate // 1000,
|
225 |
+
win_length=32 * sample_rate // 1000,
|
226 |
+
hop_length=10 * sample_rate // 1000,
|
227 |
+
f_min=50,
|
228 |
+
f_max=sr_to_fmax[sample_rate],
|
229 |
+
n_mels=64,
|
230 |
+
norm="slaney",
|
231 |
+
mel_scale="slaney"
|
232 |
+
)
|
233 |
+
self.hop_length = 10 * sample_rate // 1000
|
234 |
+
self.db_transform = transforms.AmplitudeToDB()
|
235 |
+
|
236 |
+
self.bn0 = nn.BatchNorm2d(64)
|
237 |
+
|
238 |
+
self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
|
239 |
+
self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
|
240 |
+
self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
|
241 |
+
self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
|
242 |
+
|
243 |
+
self.downsample_ratio = 16
|
244 |
+
|
245 |
+
self.fc1 = nn.Linear(512, 512, bias=True)
|
246 |
+
self.fc_emb_size = 512
|
247 |
+
self.init_weight()
|
248 |
+
self.freeze = freeze
|
249 |
+
|
250 |
+
def init_weight(self):
|
251 |
+
init_bn(self.bn0)
|
252 |
+
init_layer(self.fc1)
|
253 |
+
|
254 |
+
def load_pretrained(self, pretrained, output_fn):
|
255 |
+
checkpoint = torch.load(pretrained, map_location="cpu")
|
256 |
+
|
257 |
+
if "model" in checkpoint:
|
258 |
+
state_dict = checkpoint["model"]
|
259 |
+
else:
|
260 |
+
raise Exception("Unkown checkpoint format")
|
261 |
+
|
262 |
+
loaded_keys = merge_load_state_dict(state_dict, self, output_fn)
|
263 |
+
if self.freeze:
|
264 |
+
for name, param in self.named_parameters():
|
265 |
+
if name in loaded_keys:
|
266 |
+
param.requires_grad = False
|
267 |
+
else:
|
268 |
+
param.requires_grad = True
|
269 |
+
|
270 |
+
def forward(self, input_dict):
|
271 |
+
waveform = input_dict["wav"]
|
272 |
+
wave_length = input_dict["wav_len"]
|
273 |
+
specaug = input_dict["specaug"]
|
274 |
+
x = self.melspec_extractor(waveform)
|
275 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
276 |
+
x = x.transpose(1, 2)
|
277 |
+
x = x.unsqueeze(1) # (batch_size, 1, time_steps, mel_bins)
|
278 |
+
|
279 |
+
x = x.transpose(1, 3)
|
280 |
+
x = self.bn0(x)
|
281 |
+
x = x.transpose(1, 3)
|
282 |
+
|
283 |
+
x = self.conv_block1(x, pool_size=(2, 2), pool_type='avg')
|
284 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
285 |
+
x = self.conv_block2(x, pool_size=(2, 2), pool_type='avg')
|
286 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
287 |
+
x = self.conv_block3(x, pool_size=(2, 2), pool_type='avg')
|
288 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
289 |
+
x = self.conv_block4(x, pool_size=(2, 2), pool_type='avg')
|
290 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
291 |
+
|
292 |
+
x = torch.mean(x, dim=3)
|
293 |
+
attn_emb = x.transpose(1, 2)
|
294 |
+
wave_length = torch.as_tensor(wave_length)
|
295 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
296 |
+
rounding_mode="floor") + 1
|
297 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
298 |
+
rounding_mode="floor")
|
299 |
+
x_max = max_with_lens(attn_emb, feat_length)
|
300 |
+
x_mean = mean_with_lens(attn_emb, feat_length)
|
301 |
+
x = x_max + x_mean
|
302 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
303 |
+
x = F.relu_(self.fc1(x))
|
304 |
+
fc_emb = F.dropout(x, p=0.5, training=self.training)
|
305 |
+
|
306 |
+
return {
|
307 |
+
"attn_emb": attn_emb,
|
308 |
+
"fc_emb": fc_emb,
|
309 |
+
"attn_emb_len": feat_length
|
310 |
+
}
|
311 |
+
|
312 |
+
|
313 |
+
class Cnn14Encoder(nn.Module):
|
314 |
+
def __init__(self, sample_rate=32000, freeze=False):
|
315 |
+
super().__init__()
|
316 |
+
sr_to_fmax = {
|
317 |
+
32000: 14000,
|
318 |
+
16000: 8000
|
319 |
+
}
|
320 |
+
# Logmel spectrogram extractor
|
321 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
322 |
+
sample_rate=sample_rate,
|
323 |
+
n_fft=32 * sample_rate // 1000,
|
324 |
+
win_length=32 * sample_rate // 1000,
|
325 |
+
hop_length=10 * sample_rate // 1000,
|
326 |
+
f_min=50,
|
327 |
+
f_max=sr_to_fmax[sample_rate],
|
328 |
+
n_mels=64,
|
329 |
+
norm="slaney",
|
330 |
+
mel_scale="slaney"
|
331 |
+
)
|
332 |
+
self.hop_length = 10 * sample_rate // 1000
|
333 |
+
self.db_transform = transforms.AmplitudeToDB()
|
334 |
+
|
335 |
+
self.bn0 = nn.BatchNorm2d(64)
|
336 |
+
|
337 |
+
self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
|
338 |
+
self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
|
339 |
+
self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
|
340 |
+
self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
|
341 |
+
self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
|
342 |
+
self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
|
343 |
+
|
344 |
+
self.downsample_ratio = 32
|
345 |
+
|
346 |
+
self.fc1 = nn.Linear(2048, 2048, bias=True)
|
347 |
+
self.fc_emb_size = 2048
|
348 |
+
|
349 |
+
self.init_weight()
|
350 |
+
self.freeze = freeze
|
351 |
+
|
352 |
+
def init_weight(self):
|
353 |
+
init_bn(self.bn0)
|
354 |
+
init_layer(self.fc1)
|
355 |
+
|
356 |
+
def load_pretrained(self, pretrained, output_fn):
|
357 |
+
checkpoint = torch.load(pretrained, map_location="cpu")
|
358 |
+
|
359 |
+
if "model" in checkpoint:
|
360 |
+
state_keys = checkpoint["model"].keys()
|
361 |
+
backbone = False
|
362 |
+
for key in state_keys:
|
363 |
+
if key.startswith("backbone."):
|
364 |
+
backbone = True
|
365 |
+
break
|
366 |
+
|
367 |
+
if backbone: # COLA
|
368 |
+
state_dict = {}
|
369 |
+
for key, value in checkpoint["model"].items():
|
370 |
+
if key.startswith("backbone."):
|
371 |
+
model_key = key.replace("backbone.", "")
|
372 |
+
state_dict[model_key] = value
|
373 |
+
else: # PANNs
|
374 |
+
state_dict = checkpoint["model"]
|
375 |
+
elif "state_dict" in checkpoint: # BLAT
|
376 |
+
state_dict = checkpoint["state_dict"]
|
377 |
+
state_dict_keys = list(filter(
|
378 |
+
lambda x: "audio_encoder" in x, state_dict.keys()))
|
379 |
+
state_dict = {
|
380 |
+
key.replace('audio_encoder.', ''): state_dict[key]
|
381 |
+
for key in state_dict_keys
|
382 |
+
}
|
383 |
+
else:
|
384 |
+
raise Exception("Unkown checkpoint format")
|
385 |
+
|
386 |
+
loaded_keys = merge_load_state_dict(state_dict, self, output_fn)
|
387 |
+
if self.freeze:
|
388 |
+
for name, param in self.named_parameters():
|
389 |
+
if name in loaded_keys:
|
390 |
+
param.requires_grad = False
|
391 |
+
else:
|
392 |
+
param.requires_grad = True
|
393 |
+
|
394 |
+
def forward(self, input_dict):
|
395 |
+
waveform = input_dict["wav"]
|
396 |
+
wave_length = input_dict["wav_len"]
|
397 |
+
specaug = input_dict["specaug"]
|
398 |
+
x = self.melspec_extractor(waveform)
|
399 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
400 |
+
x = x.transpose(1, 2)
|
401 |
+
x = x.unsqueeze(1) # (batch_size, 1, time_steps, mel_bins)
|
402 |
+
|
403 |
+
x = x.transpose(1, 3)
|
404 |
+
x = self.bn0(x)
|
405 |
+
x = x.transpose(1, 3)
|
406 |
+
|
407 |
+
x = self.conv_block1(x, pool_size=(2, 2), pool_type='avg')
|
408 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
409 |
+
x = self.conv_block2(x, pool_size=(2, 2), pool_type='avg')
|
410 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
411 |
+
x = self.conv_block3(x, pool_size=(2, 2), pool_type='avg')
|
412 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
413 |
+
x = self.conv_block4(x, pool_size=(2, 2), pool_type='avg')
|
414 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
415 |
+
x = self.conv_block5(x, pool_size=(2, 2), pool_type='avg')
|
416 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
417 |
+
x = self.conv_block6(x, pool_size=(1, 1), pool_type='avg')
|
418 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
419 |
+
x = torch.mean(x, dim=3)
|
420 |
+
attn_emb = x.transpose(1, 2)
|
421 |
+
|
422 |
+
wave_length = torch.as_tensor(wave_length)
|
423 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
424 |
+
rounding_mode="floor") + 1
|
425 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
426 |
+
rounding_mode="floor")
|
427 |
+
x_max = max_with_lens(attn_emb, feat_length)
|
428 |
+
x_mean = mean_with_lens(attn_emb, feat_length)
|
429 |
+
x = x_max + x_mean
|
430 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
431 |
+
x = F.relu_(self.fc1(x))
|
432 |
+
fc_emb = F.dropout(x, p=0.5, training=self.training)
|
433 |
+
|
434 |
+
output_dict = {
|
435 |
+
'fc_emb': fc_emb,
|
436 |
+
'attn_emb': attn_emb,
|
437 |
+
'attn_emb_len': feat_length
|
438 |
+
}
|
439 |
+
|
440 |
+
return output_dict
|
441 |
+
|
442 |
+
|
443 |
+
class InvertedResidual(nn.Module):
|
444 |
+
|
445 |
+
def __init__(self, inp, oup, stride, expand_ratio):
|
446 |
+
super().__init__()
|
447 |
+
self.stride = stride
|
448 |
+
assert stride in [1, 2]
|
449 |
+
|
450 |
+
hidden_dim = round(inp * expand_ratio)
|
451 |
+
self.use_res_connect = self.stride == 1 and inp == oup
|
452 |
+
|
453 |
+
if expand_ratio == 1:
|
454 |
+
_layers = [
|
455 |
+
nn.Conv2d(hidden_dim, hidden_dim, 3, 1, 1, groups=hidden_dim, bias=False),
|
456 |
+
nn.AvgPool2d(stride),
|
457 |
+
nn.BatchNorm2d(hidden_dim),
|
458 |
+
nn.ReLU6(inplace=True),
|
459 |
+
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
|
460 |
+
nn.BatchNorm2d(oup)
|
461 |
+
]
|
462 |
+
_layers = nn.Sequential(*_layers)
|
463 |
+
init_layer(_layers[0])
|
464 |
+
init_bn(_layers[2])
|
465 |
+
init_layer(_layers[4])
|
466 |
+
init_bn(_layers[5])
|
467 |
+
self.conv = _layers
|
468 |
+
else:
|
469 |
+
_layers = [
|
470 |
+
nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
|
471 |
+
nn.BatchNorm2d(hidden_dim),
|
472 |
+
nn.ReLU6(inplace=True),
|
473 |
+
nn.Conv2d(hidden_dim, hidden_dim, 3, 1, 1, groups=hidden_dim, bias=False),
|
474 |
+
nn.AvgPool2d(stride),
|
475 |
+
nn.BatchNorm2d(hidden_dim),
|
476 |
+
nn.ReLU6(inplace=True),
|
477 |
+
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
|
478 |
+
nn.BatchNorm2d(oup)
|
479 |
+
]
|
480 |
+
_layers = nn.Sequential(*_layers)
|
481 |
+
init_layer(_layers[0])
|
482 |
+
init_bn(_layers[1])
|
483 |
+
init_layer(_layers[3])
|
484 |
+
init_bn(_layers[5])
|
485 |
+
init_layer(_layers[7])
|
486 |
+
init_bn(_layers[8])
|
487 |
+
self.conv = _layers
|
488 |
+
|
489 |
+
def forward(self, x):
|
490 |
+
if self.use_res_connect:
|
491 |
+
return x + self.conv(x)
|
492 |
+
else:
|
493 |
+
return self.conv(x)
|
494 |
+
|
495 |
+
|
496 |
+
class MobileNetV2(nn.Module):
|
497 |
+
def __init__(self, sample_rate):
|
498 |
+
|
499 |
+
super().__init__()
|
500 |
+
|
501 |
+
sr_to_fmax = {
|
502 |
+
32000: 14000,
|
503 |
+
16000: 8000
|
504 |
+
}
|
505 |
+
# Logmel spectrogram extractor
|
506 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
507 |
+
sample_rate=sample_rate,
|
508 |
+
n_fft=32 * sample_rate // 1000,
|
509 |
+
win_length=32 * sample_rate // 1000,
|
510 |
+
hop_length=10 * sample_rate // 1000,
|
511 |
+
f_min=50,
|
512 |
+
f_max=sr_to_fmax[sample_rate],
|
513 |
+
n_mels=64,
|
514 |
+
norm="slaney",
|
515 |
+
mel_scale="slaney"
|
516 |
+
)
|
517 |
+
self.hop_length = 10 * sample_rate // 1000
|
518 |
+
self.db_transform = transforms.AmplitudeToDB()
|
519 |
+
|
520 |
+
self.bn0 = nn.BatchNorm2d(64)
|
521 |
+
|
522 |
+
width_mult=1.
|
523 |
+
block = InvertedResidual
|
524 |
+
input_channel = 32
|
525 |
+
last_channel = 1280
|
526 |
+
interverted_residual_setting = [
|
527 |
+
# t, c, n, s
|
528 |
+
[1, 16, 1, 1],
|
529 |
+
[6, 24, 2, 2],
|
530 |
+
[6, 32, 3, 2],
|
531 |
+
[6, 64, 4, 2],
|
532 |
+
[6, 96, 3, 2],
|
533 |
+
[6, 160, 3, 1],
|
534 |
+
[6, 320, 1, 1],
|
535 |
+
]
|
536 |
+
|
537 |
+
self.downsample_ratio = 32
|
538 |
+
|
539 |
+
def conv_bn(inp, oup, stride):
|
540 |
+
_layers = [
|
541 |
+
nn.Conv2d(inp, oup, 3, 1, 1, bias=False),
|
542 |
+
nn.AvgPool2d(stride),
|
543 |
+
nn.BatchNorm2d(oup),
|
544 |
+
nn.ReLU6(inplace=True)
|
545 |
+
]
|
546 |
+
_layers = nn.Sequential(*_layers)
|
547 |
+
init_layer(_layers[0])
|
548 |
+
init_bn(_layers[2])
|
549 |
+
return _layers
|
550 |
+
|
551 |
+
|
552 |
+
def conv_1x1_bn(inp, oup):
|
553 |
+
_layers = nn.Sequential(
|
554 |
+
nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
|
555 |
+
nn.BatchNorm2d(oup),
|
556 |
+
nn.ReLU6(inplace=True)
|
557 |
+
)
|
558 |
+
init_layer(_layers[0])
|
559 |
+
init_bn(_layers[1])
|
560 |
+
return _layers
|
561 |
+
|
562 |
+
# building first layer
|
563 |
+
input_channel = int(input_channel * width_mult)
|
564 |
+
self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
|
565 |
+
self.features = [conv_bn(1, input_channel, 2)]
|
566 |
+
# building inverted residual blocks
|
567 |
+
for t, c, n, s in interverted_residual_setting:
|
568 |
+
output_channel = int(c * width_mult)
|
569 |
+
for i in range(n):
|
570 |
+
if i == 0:
|
571 |
+
self.features.append(block(input_channel, output_channel, s, expand_ratio=t))
|
572 |
+
else:
|
573 |
+
self.features.append(block(input_channel, output_channel, 1, expand_ratio=t))
|
574 |
+
input_channel = output_channel
|
575 |
+
# building last several layers
|
576 |
+
self.features.append(conv_1x1_bn(input_channel, self.last_channel))
|
577 |
+
# make it nn.Sequential
|
578 |
+
self.features = nn.Sequential(*self.features)
|
579 |
+
|
580 |
+
self.fc1 = nn.Linear(1280, 1024, bias=True)
|
581 |
+
|
582 |
+
self.init_weight()
|
583 |
+
|
584 |
+
def init_weight(self):
|
585 |
+
init_bn(self.bn0)
|
586 |
+
init_layer(self.fc1)
|
587 |
+
|
588 |
+
def forward(self, input_dict):
|
589 |
+
|
590 |
+
waveform = input_dict["wav"]
|
591 |
+
wave_length = input_dict["wav_len"]
|
592 |
+
specaug = input_dict["specaug"]
|
593 |
+
x = self.melspec_extractor(waveform)
|
594 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
595 |
+
x = x.transpose(1, 2)
|
596 |
+
x = x.unsqueeze(1) # (batch_size, 1, time_steps, mel_bins)
|
597 |
+
|
598 |
+
x = x.transpose(1, 3)
|
599 |
+
x = self.bn0(x)
|
600 |
+
x = x.transpose(1, 3)
|
601 |
+
|
602 |
+
x = self.features(x)
|
603 |
+
|
604 |
+
x = torch.mean(x, dim=3)
|
605 |
+
attn_emb = x.transpose(1, 2)
|
606 |
+
|
607 |
+
wave_length = torch.as_tensor(wave_length)
|
608 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
609 |
+
rounding_mode="floor") + 1
|
610 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
611 |
+
rounding_mode="floor")
|
612 |
+
x_max = max_with_lens(attn_emb, feat_length)
|
613 |
+
x_mean = mean_with_lens(attn_emb, feat_length)
|
614 |
+
x = x_max + x_mean
|
615 |
+
# TODO: the original PANNs code does not have dropout here, why?
|
616 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
617 |
+
x = F.relu_(self.fc1(x))
|
618 |
+
fc_emb = F.dropout(x, p=0.5, training=self.training)
|
619 |
+
|
620 |
+
output_dict = {
|
621 |
+
'fc_emb': fc_emb,
|
622 |
+
'attn_emb': attn_emb,
|
623 |
+
'attn_emb_len': feat_length
|
624 |
+
}
|
625 |
+
|
626 |
+
return output_dict
|
627 |
+
|
628 |
+
|
629 |
+
class MobileNetV3(nn.Module):
|
630 |
+
|
631 |
+
def __init__(self,
|
632 |
+
sample_rate,
|
633 |
+
model_name,
|
634 |
+
n_mels=64,
|
635 |
+
win_length=32,
|
636 |
+
pretrained=True,
|
637 |
+
freeze=False,
|
638 |
+
pooling="mean_max_fc"):
|
639 |
+
|
640 |
+
from captioning.models.eff_at_encoder import get_model, NAME_TO_WIDTH
|
641 |
+
|
642 |
+
super().__init__()
|
643 |
+
sr_to_fmax = {
|
644 |
+
32000: 14000,
|
645 |
+
16000: 8000
|
646 |
+
}
|
647 |
+
self.n_mels = n_mels
|
648 |
+
# Logmel spectrogram extractor
|
649 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
650 |
+
sample_rate=sample_rate,
|
651 |
+
n_fft=32 * sample_rate // 1000,
|
652 |
+
win_length=win_length * sample_rate // 1000,
|
653 |
+
hop_length=10 * sample_rate // 1000,
|
654 |
+
f_min=50,
|
655 |
+
f_max=sr_to_fmax[sample_rate],
|
656 |
+
n_mels=n_mels,
|
657 |
+
norm="slaney",
|
658 |
+
mel_scale="slaney"
|
659 |
+
)
|
660 |
+
self.hop_length = 10 * sample_rate // 1000
|
661 |
+
self.db_transform = transforms.AmplitudeToDB()
|
662 |
+
|
663 |
+
self.bn0 = nn.BatchNorm2d(n_mels)
|
664 |
+
|
665 |
+
width_mult = NAME_TO_WIDTH(model_name)
|
666 |
+
self.features = get_model(model_name=model_name,
|
667 |
+
pretrained=pretrained,
|
668 |
+
width_mult=width_mult).features
|
669 |
+
self.downsample_ratio = 32
|
670 |
+
|
671 |
+
if pooling == "mean_max_fc":
|
672 |
+
self.fc_emb_size = 512
|
673 |
+
self.fc1 = nn.Linear(self.features[-1].out_channels, 512, bias=True)
|
674 |
+
elif pooling == "mean":
|
675 |
+
self.fc_emb_size = self.features[-1].out_channels
|
676 |
+
self.init_weight()
|
677 |
+
|
678 |
+
if freeze:
|
679 |
+
for param in self.parameters():
|
680 |
+
param.requires_grad = False
|
681 |
+
|
682 |
+
self.pooling = pooling
|
683 |
+
|
684 |
+
def init_weight(self):
|
685 |
+
init_bn(self.bn0)
|
686 |
+
if hasattr(self, "fc1"):
|
687 |
+
init_layer(self.fc1)
|
688 |
+
|
689 |
+
def forward(self, input_dict):
|
690 |
+
|
691 |
+
waveform = input_dict["wav"]
|
692 |
+
wave_length = input_dict["wav_len"]
|
693 |
+
specaug = input_dict["specaug"]
|
694 |
+
x = self.melspec_extractor(waveform)
|
695 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
696 |
+
x = x.transpose(1, 2)
|
697 |
+
x = x.unsqueeze(1) # (batch_size, 1, time_steps, mel_bins)
|
698 |
+
|
699 |
+
x = x.transpose(1, 3)
|
700 |
+
x = self.bn0(x)
|
701 |
+
x = x.transpose(1, 3)
|
702 |
+
|
703 |
+
x = self.features(x)
|
704 |
+
|
705 |
+
x = torch.mean(x, dim=3)
|
706 |
+
attn_emb = x.transpose(1, 2)
|
707 |
+
|
708 |
+
wave_length = torch.as_tensor(wave_length)
|
709 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
710 |
+
rounding_mode="floor") + 1
|
711 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
712 |
+
rounding_mode="floor")
|
713 |
+
|
714 |
+
if self.pooling == "mean_max_fc":
|
715 |
+
x_max = max_with_lens(attn_emb, feat_length)
|
716 |
+
x_mean = mean_with_lens(attn_emb, feat_length)
|
717 |
+
x = x_max + x_mean
|
718 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
719 |
+
x = F.relu_(self.fc1(x))
|
720 |
+
fc_emb = F.dropout(x, p=0.5, training=self.training)
|
721 |
+
elif self.pooling == "mean":
|
722 |
+
fc_emb = mean_with_lens(attn_emb, feat_length)
|
723 |
+
|
724 |
+
output_dict = {
|
725 |
+
'fc_emb': fc_emb,
|
726 |
+
'attn_emb': attn_emb,
|
727 |
+
'attn_emb_len': feat_length
|
728 |
+
}
|
729 |
+
|
730 |
+
return output_dict
|
731 |
+
|
732 |
+
|
733 |
+
class EfficientNetB2(nn.Module):
|
734 |
+
|
735 |
+
def __init__(self,
|
736 |
+
n_mels: int = 64,
|
737 |
+
win_length: int = 32,
|
738 |
+
hop_length: int = 10,
|
739 |
+
f_min: int = 0,
|
740 |
+
pretrained: bool = False,
|
741 |
+
prune_ratio: float = 0.0,
|
742 |
+
prune_se: bool = True,
|
743 |
+
prune_start_layer: int = 0,
|
744 |
+
prune_method: str = "operator_norm",
|
745 |
+
freeze: bool = False,):
|
746 |
+
from models.eff_latent_encoder import get_model, get_pruned_model
|
747 |
+
super().__init__()
|
748 |
+
sample_rate = 16000
|
749 |
+
self.melspec_extractor = transforms.MelSpectrogram(
|
750 |
+
sample_rate=sample_rate,
|
751 |
+
n_fft=win_length * sample_rate // 1000,
|
752 |
+
win_length=win_length * sample_rate // 1000,
|
753 |
+
hop_length=hop_length * sample_rate // 1000,
|
754 |
+
f_min=f_min,
|
755 |
+
n_mels=n_mels,
|
756 |
+
)
|
757 |
+
self.hop_length = 10 * sample_rate // 1000
|
758 |
+
self.db_transform = transforms.AmplitudeToDB(top_db=120)
|
759 |
+
if prune_ratio > 0:
|
760 |
+
self.backbone = get_pruned_model(pretrained=pretrained,
|
761 |
+
prune_ratio=prune_ratio,
|
762 |
+
prune_start_layer=prune_start_layer,
|
763 |
+
prune_se=prune_se,
|
764 |
+
prune_method=prune_method)
|
765 |
+
else:
|
766 |
+
self.backbone = get_model(pretrained=pretrained)
|
767 |
+
self.fc_emb_size = self.backbone.eff_net._conv_head.out_channels
|
768 |
+
self.downsample_ratio = 32
|
769 |
+
if freeze:
|
770 |
+
for param in self.parameters():
|
771 |
+
param.requires_grad = False
|
772 |
+
|
773 |
+
def forward(self, input_dict):
|
774 |
+
|
775 |
+
waveform = input_dict["wav"]
|
776 |
+
wave_length = input_dict["wav_len"]
|
777 |
+
specaug = input_dict["specaug"]
|
778 |
+
x = self.melspec_extractor(waveform)
|
779 |
+
x = self.db_transform(x) # (batch_size, mel_bins, time_steps)
|
780 |
+
|
781 |
+
x = self.backbone(x)
|
782 |
+
attn_emb = x
|
783 |
+
|
784 |
+
wave_length = torch.as_tensor(wave_length)
|
785 |
+
feat_length = torch.div(wave_length, self.hop_length,
|
786 |
+
rounding_mode="floor") + 1
|
787 |
+
feat_length = torch.div(feat_length, self.downsample_ratio,
|
788 |
+
rounding_mode="floor")
|
789 |
+
fc_emb = mean_with_lens(attn_emb, feat_length)
|
790 |
+
|
791 |
+
output_dict = {
|
792 |
+
'fc_emb': fc_emb,
|
793 |
+
'attn_emb': attn_emb,
|
794 |
+
'attn_emb_len': feat_length
|
795 |
+
}
|
796 |
+
return output_dict
|
797 |
+
|
798 |
+
|
799 |
+
if __name__ == "__main__":
|
800 |
+
encoder = MobileNetV3(32000, "mn10_as")
|
801 |
+
print(encoder)
|
802 |
+
input_dict = {
|
803 |
+
"wav": torch.randn(4, 320000),
|
804 |
+
"wav_len": torch.tensor([320000, 280000, 160000, 300000]),
|
805 |
+
"specaug": True
|
806 |
+
}
|
807 |
+
output_dict = encoder(input_dict)
|
808 |
+
print("attn embed: ", output_dict["attn_emb"].shape)
|
809 |
+
print("fc embed: ", output_dict["fc_emb"].shape)
|
810 |
+
print("attn embed length: ", output_dict["attn_emb_len"])
|
models/eff_latent_encoder.py
ADDED
@@ -0,0 +1,347 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from tqdm import tqdm
|
6 |
+
from efficientnet_pytorch import EfficientNet
|
7 |
+
from efficientnet_pytorch.model import MBConvBlock
|
8 |
+
from efficientnet_pytorch import utils as efficientnet_utils
|
9 |
+
from efficientnet_pytorch.utils import (
|
10 |
+
round_filters,
|
11 |
+
round_repeats,
|
12 |
+
get_same_padding_conv2d,
|
13 |
+
calculate_output_image_size,
|
14 |
+
MemoryEfficientSwish,
|
15 |
+
)
|
16 |
+
from einops import rearrange, reduce
|
17 |
+
from torch.hub import load_state_dict_from_url
|
18 |
+
|
19 |
+
|
20 |
+
model_dir = "./"
|
21 |
+
|
22 |
+
|
23 |
+
class _EffiNet(nn.Module):
|
24 |
+
"""A proxy for efficient net models"""
|
25 |
+
def __init__(self,
|
26 |
+
blocks_args=None,
|
27 |
+
global_params=None,
|
28 |
+
prune_start_layer: int = 0,
|
29 |
+
prune_se: bool = True,
|
30 |
+
prune_ratio: float = 0.0
|
31 |
+
) -> None:
|
32 |
+
super().__init__()
|
33 |
+
if prune_ratio > 0:
|
34 |
+
self.eff_net = EfficientNetB2Pruned(blocks_args=blocks_args,
|
35 |
+
global_params=global_params,
|
36 |
+
prune_start_layer=prune_start_layer,
|
37 |
+
prune_se=prune_se,
|
38 |
+
prune_ratio=prune_ratio)
|
39 |
+
else:
|
40 |
+
self.eff_net = EfficientNet(blocks_args=blocks_args,
|
41 |
+
global_params=global_params)
|
42 |
+
|
43 |
+
|
44 |
+
def forward(self, x: torch.Tensor):
|
45 |
+
x = rearrange(x, 'b f t -> b 1 f t')
|
46 |
+
x = self.eff_net.extract_features(x)
|
47 |
+
return reduce(x, 'b c f t -> b t c', 'mean')
|
48 |
+
|
49 |
+
|
50 |
+
def get_model(pretrained=True) -> _EffiNet:
|
51 |
+
blocks_args, global_params = efficientnet_utils.get_model_params(
|
52 |
+
'efficientnet-b2', {'include_top': False})
|
53 |
+
model = _EffiNet(blocks_args=blocks_args,
|
54 |
+
global_params=global_params)
|
55 |
+
model.eff_net._change_in_channels(1)
|
56 |
+
if pretrained:
|
57 |
+
model_path = os.path.join(model_dir, "effb2.pt")
|
58 |
+
if not os.path.exists(model_path):
|
59 |
+
state_dict = load_state_dict_from_url(
|
60 |
+
'https://github.com/richermans/HEAR2021_EfficientLatent/releases/download/v0.0.1/effb2.pt',
|
61 |
+
progress=True,
|
62 |
+
model_dir=model_dir)
|
63 |
+
else:
|
64 |
+
state_dict = torch.load(model_path)
|
65 |
+
del_keys = [key for key in state_dict if key.startswith("front_end")]
|
66 |
+
for key in del_keys:
|
67 |
+
del state_dict[key]
|
68 |
+
model.eff_net.load_state_dict(state_dict)
|
69 |
+
return model
|
70 |
+
|
71 |
+
|
72 |
+
class MBConvBlockPruned(MBConvBlock):
|
73 |
+
|
74 |
+
def __init__(self, block_args, global_params, image_size=None, prune_ratio=0.5, prune_se=True):
|
75 |
+
super(MBConvBlock, self).__init__()
|
76 |
+
self._block_args = block_args
|
77 |
+
self._bn_mom = 1 - global_params.batch_norm_momentum # pytorch's difference from tensorflow
|
78 |
+
self._bn_eps = global_params.batch_norm_epsilon
|
79 |
+
self.has_se = (self._block_args.se_ratio is not None) and (0 < self._block_args.se_ratio <= 1)
|
80 |
+
self.id_skip = block_args.id_skip # whether to use skip connection and drop connect
|
81 |
+
|
82 |
+
# Expansion phase (Inverted Bottleneck)
|
83 |
+
inp = self._block_args.input_filters # number of input channels
|
84 |
+
oup = self._block_args.input_filters * self._block_args.expand_ratio # number of output channels
|
85 |
+
if self._block_args.expand_ratio != 1:
|
86 |
+
oup = int(oup * (1 - prune_ratio))
|
87 |
+
Conv2d = get_same_padding_conv2d(image_size=image_size)
|
88 |
+
self._expand_conv = Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)
|
89 |
+
self._bn0 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
|
90 |
+
# image_size = calculate_output_image_size(image_size, 1) <-- this wouldn't modify image_size
|
91 |
+
|
92 |
+
# Depthwise convolution phase
|
93 |
+
k = self._block_args.kernel_size
|
94 |
+
s = self._block_args.stride
|
95 |
+
Conv2d = get_same_padding_conv2d(image_size=image_size)
|
96 |
+
self._depthwise_conv = Conv2d(
|
97 |
+
in_channels=oup, out_channels=oup, groups=oup, # groups makes it depthwise
|
98 |
+
kernel_size=k, stride=s, bias=False)
|
99 |
+
self._bn1 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
|
100 |
+
image_size = calculate_output_image_size(image_size, s)
|
101 |
+
|
102 |
+
# Squeeze and Excitation layer, if desired
|
103 |
+
if self.has_se:
|
104 |
+
Conv2d = get_same_padding_conv2d(image_size=(1, 1))
|
105 |
+
num_squeezed_channels = max(1, int(self._block_args.input_filters * self._block_args.se_ratio))
|
106 |
+
if prune_se:
|
107 |
+
num_squeezed_channels = int(num_squeezed_channels * (1 - prune_ratio))
|
108 |
+
self._se_reduce = Conv2d(in_channels=oup, out_channels=num_squeezed_channels, kernel_size=1)
|
109 |
+
self._se_expand = Conv2d(in_channels=num_squeezed_channels, out_channels=oup, kernel_size=1)
|
110 |
+
|
111 |
+
# Pointwise convolution phase
|
112 |
+
final_oup = self._block_args.output_filters
|
113 |
+
Conv2d = get_same_padding_conv2d(image_size=image_size)
|
114 |
+
self._project_conv = Conv2d(in_channels=oup, out_channels=final_oup, kernel_size=1, bias=False)
|
115 |
+
self._bn2 = nn.BatchNorm2d(num_features=final_oup, momentum=self._bn_mom, eps=self._bn_eps)
|
116 |
+
self._swish = MemoryEfficientSwish()
|
117 |
+
|
118 |
+
|
119 |
+
class EfficientNetB2Pruned(EfficientNet):
|
120 |
+
|
121 |
+
def __init__(self, blocks_args=None, global_params=None,
|
122 |
+
prune_start_layer=0, prune_ratio=0.5, prune_se=True):
|
123 |
+
super(EfficientNet, self).__init__()
|
124 |
+
assert isinstance(blocks_args, list), 'blocks_args should be a list'
|
125 |
+
assert len(blocks_args) > 0, 'block args must be greater than 0'
|
126 |
+
self._global_params = global_params
|
127 |
+
self._blocks_args = blocks_args
|
128 |
+
|
129 |
+
# Batch norm parameters
|
130 |
+
bn_mom = 1 - self._global_params.batch_norm_momentum
|
131 |
+
bn_eps = self._global_params.batch_norm_epsilon
|
132 |
+
|
133 |
+
# Get stem static or dynamic convolution depending on image size
|
134 |
+
image_size = global_params.image_size
|
135 |
+
Conv2d = get_same_padding_conv2d(image_size=image_size)
|
136 |
+
|
137 |
+
n_build_blks = 0
|
138 |
+
# Stem
|
139 |
+
in_channels = 1 # spectrogram
|
140 |
+
|
141 |
+
p = 0.0 if n_build_blks < prune_start_layer else prune_ratio
|
142 |
+
out_channels = round_filters(32 * (1 - p),
|
143 |
+
self._global_params) # number of output channels
|
144 |
+
self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)
|
145 |
+
self._bn0 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
|
146 |
+
image_size = calculate_output_image_size(image_size, 2)
|
147 |
+
n_build_blks += 1
|
148 |
+
|
149 |
+
# Build blocks
|
150 |
+
self._blocks = nn.ModuleList([])
|
151 |
+
for block_args in self._blocks_args:
|
152 |
+
|
153 |
+
p = 0.0 if n_build_blks < prune_start_layer else prune_ratio
|
154 |
+
orig_input_filters = block_args.input_filters
|
155 |
+
# Update block input and output filters based on depth multiplier.
|
156 |
+
block_args = block_args._replace(
|
157 |
+
input_filters=round_filters(
|
158 |
+
block_args.input_filters * (1 - p),
|
159 |
+
self._global_params),
|
160 |
+
output_filters=round_filters(
|
161 |
+
block_args.output_filters * (1 - p),
|
162 |
+
self._global_params),
|
163 |
+
num_repeat=round_repeats(block_args.num_repeat, self._global_params)
|
164 |
+
)
|
165 |
+
|
166 |
+
if n_build_blks == prune_start_layer:
|
167 |
+
block_args = block_args._replace(input_filters=round_filters(
|
168 |
+
orig_input_filters,
|
169 |
+
self._global_params)
|
170 |
+
)
|
171 |
+
|
172 |
+
# The first block needs to take care of stride and filter size increase.
|
173 |
+
self._blocks.append(MBConvBlockPruned(block_args, self._global_params,
|
174 |
+
image_size=image_size, prune_ratio=p,
|
175 |
+
prune_se=prune_se))
|
176 |
+
n_build_blks += 1
|
177 |
+
|
178 |
+
image_size = calculate_output_image_size(image_size, block_args.stride)
|
179 |
+
if block_args.num_repeat > 1: # modify block_args to keep same output size
|
180 |
+
block_args = block_args._replace(input_filters=block_args.output_filters, stride=1)
|
181 |
+
for _ in range(block_args.num_repeat - 1):
|
182 |
+
self._blocks.append(MBConvBlockPruned(block_args,
|
183 |
+
self._global_params,
|
184 |
+
image_size=image_size,
|
185 |
+
prune_ratio=p,
|
186 |
+
prune_se=prune_se))
|
187 |
+
# image_size = calculate_output_image_size(image_size, block_args.stride) # stride = 1
|
188 |
+
|
189 |
+
# Head
|
190 |
+
in_channels = block_args.output_filters # output of final block
|
191 |
+
p = 0.0 if n_build_blks < prune_start_layer else prune_ratio
|
192 |
+
out_channels = round_filters(1280 * (1 - p), self._global_params)
|
193 |
+
Conv2d = get_same_padding_conv2d(image_size=image_size)
|
194 |
+
self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
|
195 |
+
self._bn1 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
|
196 |
+
|
197 |
+
# Final linear layer
|
198 |
+
self._avg_pooling = nn.AdaptiveAvgPool2d(1)
|
199 |
+
if self._global_params.include_top:
|
200 |
+
self._dropout = nn.Dropout(self._global_params.dropout_rate)
|
201 |
+
self._fc = nn.Linear(out_channels, self._global_params.num_classes)
|
202 |
+
|
203 |
+
# set activation to memory efficient swish by default
|
204 |
+
self._swish = MemoryEfficientSwish()
|
205 |
+
|
206 |
+
|
207 |
+
def get_pruned_model(pretrained: bool = True,
|
208 |
+
prune_ratio: float = 0.5,
|
209 |
+
prune_start_layer: int = 0,
|
210 |
+
prune_se: bool = True,
|
211 |
+
prune_method: str = "operator_norm") -> _EffiNet:
|
212 |
+
|
213 |
+
import captioning.models.conv_filter_pruning as pruning_lib
|
214 |
+
|
215 |
+
blocks_args, global_params = efficientnet_utils.get_model_params(
|
216 |
+
'efficientnet-b2', {'include_top': False})
|
217 |
+
# print("num blocks: ", len(blocks_args))
|
218 |
+
# print("block args: ")
|
219 |
+
# for block_arg in blocks_args:
|
220 |
+
# print(block_arg)
|
221 |
+
model = _EffiNet(blocks_args=blocks_args,
|
222 |
+
global_params=global_params,
|
223 |
+
prune_start_layer=prune_start_layer,
|
224 |
+
prune_se=prune_se,
|
225 |
+
prune_ratio=prune_ratio)
|
226 |
+
|
227 |
+
if prune_method == "operator_norm":
|
228 |
+
filter_pruning = pruning_lib.operator_norm_pruning
|
229 |
+
elif prune_method == "interspeech":
|
230 |
+
filter_pruning = pruning_lib.cs_interspeech
|
231 |
+
elif prune_method == "iclr_l1":
|
232 |
+
filter_pruning = pruning_lib.iclr_l1
|
233 |
+
elif prune_method == "iclr_gm":
|
234 |
+
filter_pruning = pruning_lib.iclr_gm
|
235 |
+
elif prune_method == "cs_waspaa":
|
236 |
+
filter_pruning = pruning_lib.cs_waspaa
|
237 |
+
|
238 |
+
|
239 |
+
if isinstance(pretrained, str):
|
240 |
+
ckpt = torch.load(pretrained, "cpu")
|
241 |
+
state_dict = {}
|
242 |
+
for key in ckpt["model"].keys():
|
243 |
+
if key.startswith("model.encoder.backbone"):
|
244 |
+
state_dict[key[len("model.encoder.backbone.eff_net."):]] = ckpt["model"][key]
|
245 |
+
elif isinstance(pretrained, bool):
|
246 |
+
model_path = os.path.join(model_dir, "effb2.pt")
|
247 |
+
if not os.path.exists(model_path):
|
248 |
+
state_dict = load_state_dict_from_url(
|
249 |
+
'https://github.com/richermans/HEAR2021_EfficientLatent/releases/download/v0.0.1/effb2.pt',
|
250 |
+
progress=True,
|
251 |
+
model_dir=model_dir)
|
252 |
+
else:
|
253 |
+
state_dict = torch.load(model_path)
|
254 |
+
del_keys = [key for key in state_dict if key.startswith("front_end")]
|
255 |
+
for key in del_keys:
|
256 |
+
del state_dict[key]
|
257 |
+
|
258 |
+
# load pretrained model with corresponding filters
|
259 |
+
# rule:
|
260 |
+
# * depthwise_conv: in_ch_idx = out_ch_idx = prev_conv_idx
|
261 |
+
mod_dep_path = [
|
262 |
+
"_conv_stem",
|
263 |
+
]
|
264 |
+
conv_to_bn = {"_conv_stem": "_bn0"}
|
265 |
+
for i in range(2):
|
266 |
+
mod_dep_path.extend([
|
267 |
+
f"_blocks.{i}._depthwise_conv",
|
268 |
+
f"_blocks.{i}._se_reduce",
|
269 |
+
f"_blocks.{i}._se_expand",
|
270 |
+
f"_blocks.{i}._project_conv",
|
271 |
+
])
|
272 |
+
conv_to_bn[f"_blocks.{i}._depthwise_conv"] = f"_blocks.{i}._bn1"
|
273 |
+
conv_to_bn[f"_blocks.{i}._project_conv"] = f"_blocks.{i}._bn2"
|
274 |
+
|
275 |
+
for i in range(2, 23):
|
276 |
+
mod_dep_path.extend([
|
277 |
+
f"_blocks.{i}._expand_conv",
|
278 |
+
f"_blocks.{i}._depthwise_conv",
|
279 |
+
f"_blocks.{i}._se_reduce",
|
280 |
+
f"_blocks.{i}._se_expand",
|
281 |
+
f"_blocks.{i}._project_conv"
|
282 |
+
])
|
283 |
+
conv_to_bn[f"_blocks.{i}._expand_conv"] = f"_blocks.{i}._bn0"
|
284 |
+
conv_to_bn[f"_blocks.{i}._depthwise_conv"] = f"_blocks.{i}._bn1"
|
285 |
+
conv_to_bn[f"_blocks.{i}._project_conv"] = f"_blocks.{i}._bn2"
|
286 |
+
|
287 |
+
mod_dep_path.append("_conv_head")
|
288 |
+
conv_to_bn["_conv_head"] = "_bn1"
|
289 |
+
|
290 |
+
# print(mod_dep_path)
|
291 |
+
# print(conv_to_bn)
|
292 |
+
|
293 |
+
key_to_w_b_idx = {}
|
294 |
+
model_dict = model.eff_net.state_dict()
|
295 |
+
for conv_key in tqdm(mod_dep_path):
|
296 |
+
weight = state_dict[f"{conv_key}.weight"]
|
297 |
+
ptr_n_filter = weight.size(0)
|
298 |
+
model_n_filter = model_dict[f"{conv_key}.weight"].size(0)
|
299 |
+
if model_n_filter < ptr_n_filter:
|
300 |
+
key_to_w_b_idx[conv_key] = filter_pruning(weight.numpy())[:model_n_filter]
|
301 |
+
else:
|
302 |
+
key_to_w_b_idx[conv_key] = slice(None)
|
303 |
+
|
304 |
+
pruned_state_dict = {}
|
305 |
+
for conv_key, prev_conv_key in zip(mod_dep_path, [None] + mod_dep_path[:-1]):
|
306 |
+
|
307 |
+
for sub_key in ["weight", "bias"]: # adjust the conv layer
|
308 |
+
cur_key = f"{conv_key}.{sub_key}"
|
309 |
+
|
310 |
+
if cur_key not in state_dict:
|
311 |
+
continue
|
312 |
+
|
313 |
+
if prev_conv_key is None or conv_key.endswith("_depthwise_conv"):
|
314 |
+
conv_in_idx = slice(None)
|
315 |
+
else:
|
316 |
+
conv_in_idx = key_to_w_b_idx[prev_conv_key]
|
317 |
+
|
318 |
+
# the first pruned layer
|
319 |
+
if model_dict[cur_key].ndim > 1 and model_dict[cur_key].size(1) == state_dict[cur_key].size(1):
|
320 |
+
conv_in_idx = slice(None)
|
321 |
+
|
322 |
+
if conv_key.endswith("_depthwise_conv"):
|
323 |
+
conv_out_idx = key_to_w_b_idx[prev_conv_key]
|
324 |
+
else:
|
325 |
+
conv_out_idx = key_to_w_b_idx[conv_key]
|
326 |
+
|
327 |
+
# if conv_key == "_blocks.16._se_reduce":
|
328 |
+
# print(len(conv_out_idx), len(conv_in_idx))
|
329 |
+
|
330 |
+
if sub_key == "weight":
|
331 |
+
pruned_state_dict[cur_key] = state_dict[cur_key][
|
332 |
+
conv_out_idx, ...][:, conv_in_idx, ...]
|
333 |
+
else:
|
334 |
+
pruned_state_dict[cur_key] = state_dict[cur_key][
|
335 |
+
conv_out_idx, ...]
|
336 |
+
|
337 |
+
if conv_key in conv_to_bn: # adjust the corresponding bn layer
|
338 |
+
for sub_key in ["weight", "bias", "running_mean", "running_var"]:
|
339 |
+
cur_key = f"{conv_to_bn[conv_key]}.{sub_key}"
|
340 |
+
if cur_key not in state_dict:
|
341 |
+
continue
|
342 |
+
pruned_state_dict[cur_key] = state_dict[cur_key][
|
343 |
+
key_to_w_b_idx[conv_key], ...]
|
344 |
+
|
345 |
+
model.eff_net.load_state_dict(pruned_state_dict)
|
346 |
+
|
347 |
+
return model
|
models/kd_wrapper.py
ADDED
@@ -0,0 +1,226 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
from einops import repeat
|
8 |
+
|
9 |
+
from models.base import CaptionMetaMixin
|
10 |
+
from utils.model_util import init
|
11 |
+
|
12 |
+
|
13 |
+
class WmlEncoderKdWrapper(nn.Module, CaptionMetaMixin):
|
14 |
+
|
15 |
+
def __init__(self,
|
16 |
+
model: nn.Module,
|
17 |
+
shared_dim: int,
|
18 |
+
tchr_layer_to_dims: Dict[str, int],
|
19 |
+
loss_type: str = "mse",):
|
20 |
+
super().__init__()
|
21 |
+
self.model = model
|
22 |
+
self.tchr_layers = list(tchr_layer_to_dims.keys())
|
23 |
+
self.stdnt_qv_proj = nn.Linear(model.encoder.fc_emb_size,
|
24 |
+
2 * shared_dim)
|
25 |
+
self.stdnt_qv_proj.apply(init)
|
26 |
+
for layer, dim in tchr_layer_to_dims.items():
|
27 |
+
self.add_module(f'tchr_kv_proj_{layer}', nn.Linear(dim, 2 * shared_dim))
|
28 |
+
getattr(self, f'tchr_kv_proj_{layer}').apply(init)
|
29 |
+
if loss_type == "mse":
|
30 |
+
self.loss_fn = nn.MSELoss(reduction="none")
|
31 |
+
|
32 |
+
def forward(self, input_dict: Dict):
|
33 |
+
output_dict = self.model(input_dict)
|
34 |
+
if "tchr_output" in input_dict:
|
35 |
+
stdnt_emb = output_dict["fc_emb"]
|
36 |
+
stdnt_qv = self.stdnt_qv_proj(stdnt_emb)
|
37 |
+
stdnt_q, stdnt_v = torch.chunk(stdnt_qv, 2, dim=-1)
|
38 |
+
|
39 |
+
tchr_output = input_dict["tchr_output"]
|
40 |
+
layer_ks, layer_vs = [], []
|
41 |
+
for layer in self.tchr_layers:
|
42 |
+
layer_kv = getattr(self, f'tchr_kv_proj_{layer}')(tchr_output[layer])
|
43 |
+
layer_k, layer_v = torch.chunk(layer_kv, 2, dim=-1)
|
44 |
+
layer_ks.append(layer_k)
|
45 |
+
layer_vs.append(layer_v)
|
46 |
+
layer_ks = torch.stack(layer_ks, dim=1)
|
47 |
+
layer_vs = torch.stack(layer_vs, dim=1)
|
48 |
+
weights = torch.softmax(stdnt_q.unsqueeze(1) @ layer_ks.transpose(1, 2), dim=-1)
|
49 |
+
stdnt_v = repeat(stdnt_v, 'b d -> b n d', n=len(self.tchr_layers))
|
50 |
+
loss = self.loss_fn(stdnt_v, layer_vs).mean(dim=-1, keepdim=True)
|
51 |
+
loss = (weights @ loss).mean()
|
52 |
+
output_dict["enc_kd_loss"] = loss
|
53 |
+
return output_dict
|
54 |
+
|
55 |
+
|
56 |
+
class MseEncoderKdWrapper(nn.Module, CaptionMetaMixin):
|
57 |
+
|
58 |
+
def __init__(self,
|
59 |
+
model: nn.Module,
|
60 |
+
shared_dim: int,
|
61 |
+
tchr_dim: int,
|
62 |
+
use_tchr_proj: bool = True,
|
63 |
+
l2_norm: bool = False,
|
64 |
+
):
|
65 |
+
super().__init__()
|
66 |
+
self.model = model
|
67 |
+
self.use_tchr_proj = use_tchr_proj
|
68 |
+
if not use_tchr_proj:
|
69 |
+
assert shared_dim == tchr_dim
|
70 |
+
self.tchr_dim = tchr_dim
|
71 |
+
self.l2_norm = l2_norm
|
72 |
+
if hasattr(model, "encoder"):
|
73 |
+
self.stdnt_proj = nn.Linear(model.encoder.fc_emb_size,
|
74 |
+
shared_dim)
|
75 |
+
else:
|
76 |
+
self.stdnt_proj = nn.Linear(model.fc_emb_size,
|
77 |
+
shared_dim)
|
78 |
+
self.stdnt_proj.apply(init)
|
79 |
+
if use_tchr_proj:
|
80 |
+
self.tchr_proj = nn.Linear(tchr_dim, shared_dim)
|
81 |
+
self.tchr_proj.apply(init)
|
82 |
+
else:
|
83 |
+
self.tchr_proj = nn.Identity()
|
84 |
+
|
85 |
+
def forward(self, input_dict: Dict):
|
86 |
+
unsup = input_dict.get("unsup", False)
|
87 |
+
if unsup is False:
|
88 |
+
if self.use_tchr_proj:
|
89 |
+
output_dict = self.model(input_dict)
|
90 |
+
stdnt_emb = output_dict["fc_emb"]
|
91 |
+
else:
|
92 |
+
encoder_output = self.model.encoder(input_dict)
|
93 |
+
stdnt_emb = encoder_output["fc_emb"]
|
94 |
+
encoder_output["fc_emb"] = self.stdnt_proj(encoder_output["fc_emb"])
|
95 |
+
encoder_output["attn_emb"] = self.stdnt_proj(encoder_output["attn_emb"])
|
96 |
+
output_dict = self.model.forward_decoder(input_dict, encoder_output)
|
97 |
+
else:
|
98 |
+
output_dict = self.model.encoder(input_dict)
|
99 |
+
stdnt_emb = output_dict["fc_emb"]
|
100 |
+
if "tchr_output" in input_dict:
|
101 |
+
stdnt_emb = self.stdnt_proj(stdnt_emb)
|
102 |
+
tchr_emb = input_dict["tchr_output"]["embedding"]
|
103 |
+
thcr_emb = self.tchr_proj(tchr_emb)
|
104 |
+
|
105 |
+
if self.l2_norm:
|
106 |
+
stdnt_emb = F.normalize(stdnt_emb, dim=-1)
|
107 |
+
thcr_emb = F.normalize(thcr_emb, dim=-1)
|
108 |
+
|
109 |
+
loss = F.mse_loss(stdnt_emb, thcr_emb)
|
110 |
+
output_dict["enc_kd_loss"] = loss
|
111 |
+
return output_dict
|
112 |
+
|
113 |
+
|
114 |
+
class ContraEncoderKdWrapper(nn.Module, CaptionMetaMixin):
|
115 |
+
|
116 |
+
def __init__(self,
|
117 |
+
model: nn.Module,
|
118 |
+
shared_dim: int,
|
119 |
+
tchr_dim: int,
|
120 |
+
):
|
121 |
+
super().__init__()
|
122 |
+
self.model = model
|
123 |
+
self.tchr_dim = tchr_dim
|
124 |
+
if hasattr(model, "encoder"):
|
125 |
+
self.stdnt_proj = nn.Linear(model.encoder.fc_emb_size,
|
126 |
+
shared_dim)
|
127 |
+
else:
|
128 |
+
self.stdnt_proj = nn.Linear(model.fc_emb_size,
|
129 |
+
shared_dim)
|
130 |
+
self.stdnt_proj.apply(init)
|
131 |
+
self.tchr_proj = nn.Linear(tchr_dim, shared_dim)
|
132 |
+
self.tchr_proj.apply(init)
|
133 |
+
self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
|
134 |
+
|
135 |
+
def forward(self, input_dict: Dict):
|
136 |
+
unsup = input_dict.get("unsup", False)
|
137 |
+
if unsup is False:
|
138 |
+
output_dict = self.model(input_dict)
|
139 |
+
else:
|
140 |
+
output_dict = self.model.encoder(input_dict)
|
141 |
+
if "tchr_output" in input_dict:
|
142 |
+
stdnt_emb = output_dict["fc_emb"]
|
143 |
+
stdnt_emb = self.stdnt_proj(stdnt_emb)
|
144 |
+
tchr_emb = input_dict["tchr_output"]["embedding"]
|
145 |
+
thcr_emb = self.tchr_proj(tchr_emb)
|
146 |
+
|
147 |
+
stdnt_emb = F.normalize(stdnt_emb, dim=-1)
|
148 |
+
thcr_emb = F.normalize(thcr_emb, dim=-1)
|
149 |
+
|
150 |
+
unscaled_logit = stdnt_emb @ thcr_emb.transpose(0, 1)
|
151 |
+
logit = self.logit_scale * unscaled_logit
|
152 |
+
label = torch.arange(logit.shape[0]).to(logit.device)
|
153 |
+
loss1 = F.cross_entropy(logit, label)
|
154 |
+
loss2 = F.cross_entropy(logit.transpose(0, 1), label)
|
155 |
+
loss = (loss1 + loss2) / 2
|
156 |
+
output_dict["enc_kd_loss"] = loss
|
157 |
+
return output_dict
|
158 |
+
|
159 |
+
|
160 |
+
class ContraMseEncoderKdWrapper(nn.Module, CaptionMetaMixin):
|
161 |
+
|
162 |
+
def __init__(self,
|
163 |
+
model: nn.Module,
|
164 |
+
shared_dim: int,
|
165 |
+
tchr_dim: int,
|
166 |
+
use_tchr_proj: bool = True,
|
167 |
+
l2_norm: bool = False,
|
168 |
+
):
|
169 |
+
super().__init__()
|
170 |
+
self.model = model
|
171 |
+
self.use_tchr_proj = use_tchr_proj
|
172 |
+
if not use_tchr_proj:
|
173 |
+
assert shared_dim == tchr_dim
|
174 |
+
self.tchr_dim = tchr_dim
|
175 |
+
self.l2_norm = l2_norm
|
176 |
+
if hasattr(model, "encoder"):
|
177 |
+
self.stdnt_proj = nn.Linear(model.encoder.fc_emb_size,
|
178 |
+
shared_dim)
|
179 |
+
else:
|
180 |
+
self.stdnt_proj = nn.Linear(model.fc_emb_size,
|
181 |
+
shared_dim)
|
182 |
+
self.stdnt_proj.apply(init)
|
183 |
+
if use_tchr_proj:
|
184 |
+
self.tchr_proj = nn.Linear(tchr_dim, shared_dim)
|
185 |
+
self.tchr_proj.apply(init)
|
186 |
+
else:
|
187 |
+
self.tchr_proj = nn.Identity()
|
188 |
+
self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
|
189 |
+
|
190 |
+
def forward(self, input_dict: Dict):
|
191 |
+
unsup = input_dict.get("unsup", False)
|
192 |
+
if unsup is False:
|
193 |
+
if self.use_tchr_proj:
|
194 |
+
output_dict = self.model(input_dict)
|
195 |
+
stdnt_emb = output_dict["fc_emb"]
|
196 |
+
else:
|
197 |
+
encoder_output = self.model.encoder(input_dict)
|
198 |
+
stdnt_emb = encoder_output["fc_emb"]
|
199 |
+
encoder_output["fc_emb"] = self.stdnt_proj(encoder_output["fc_emb"])
|
200 |
+
encoder_output["attn_emb"] = self.stdnt_proj(encoder_output["attn_emb"])
|
201 |
+
output_dict = self.model.forward_decoder(input_dict, encoder_output)
|
202 |
+
else:
|
203 |
+
output_dict = self.model.encoder(input_dict)
|
204 |
+
stdnt_emb = output_dict["fc_emb"]
|
205 |
+
if "tchr_output" in input_dict:
|
206 |
+
stdnt_emb = self.stdnt_proj(stdnt_emb)
|
207 |
+
tchr_emb = input_dict["tchr_output"]["embedding"]
|
208 |
+
thcr_emb = self.tchr_proj(tchr_emb)
|
209 |
+
|
210 |
+
if self.l2_norm:
|
211 |
+
stdnt_emb = F.normalize(stdnt_emb, dim=-1)
|
212 |
+
thcr_emb = F.normalize(thcr_emb, dim=-1)
|
213 |
+
|
214 |
+
mse_loss = F.mse_loss(stdnt_emb, thcr_emb)
|
215 |
+
|
216 |
+
stdnt_emb = F.normalize(stdnt_emb, dim=-1)
|
217 |
+
thcr_emb = F.normalize(thcr_emb, dim=-1)
|
218 |
+
unscaled_logit = stdnt_emb @ thcr_emb.transpose(0, 1)
|
219 |
+
logit = self.logit_scale * unscaled_logit
|
220 |
+
label = torch.arange(logit.shape[0]).to(logit.device)
|
221 |
+
loss1 = F.cross_entropy(logit, label)
|
222 |
+
loss2 = F.cross_entropy(logit.transpose(0, 1), label)
|
223 |
+
cntr_loss = (loss1 + loss2) / 2
|
224 |
+
output_dict["enc_kd_loss"] = mse_loss + cntr_loss
|
225 |
+
|
226 |
+
return output_dict
|
models/transformer_decoder.py
ADDED
@@ -0,0 +1,214 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
|
6 |
+
from models import BaseDecoder
|
7 |
+
from utils.model_util import generate_length_mask, PositionalEncoding
|
8 |
+
from utils.train_util import merge_load_state_dict
|
9 |
+
|
10 |
+
|
11 |
+
class TransformerDecoder(BaseDecoder):
|
12 |
+
|
13 |
+
def __init__(self,
|
14 |
+
emb_dim,
|
15 |
+
vocab_size,
|
16 |
+
fc_emb_dim,
|
17 |
+
attn_emb_dim,
|
18 |
+
dropout,
|
19 |
+
freeze=False,
|
20 |
+
tie_weights=False,
|
21 |
+
**kwargs):
|
22 |
+
super().__init__(emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
23 |
+
dropout=dropout, tie_weights=tie_weights)
|
24 |
+
self.d_model = emb_dim
|
25 |
+
self.nhead = kwargs.get("nhead", self.d_model // 64)
|
26 |
+
self.nlayers = kwargs.get("nlayers", 2)
|
27 |
+
self.dim_feedforward = kwargs.get("dim_feedforward", self.d_model * 4)
|
28 |
+
|
29 |
+
self.pos_encoder = PositionalEncoding(self.d_model, dropout)
|
30 |
+
layer = nn.TransformerDecoderLayer(d_model=self.d_model,
|
31 |
+
nhead=self.nhead,
|
32 |
+
dim_feedforward=self.dim_feedforward,
|
33 |
+
dropout=dropout)
|
34 |
+
self.model = nn.TransformerDecoder(layer, self.nlayers)
|
35 |
+
self.classifier = nn.Linear(self.d_model, vocab_size, bias=False)
|
36 |
+
if tie_weights:
|
37 |
+
self.classifier.weight = self.word_embedding.weight
|
38 |
+
self.attn_proj = nn.Sequential(
|
39 |
+
nn.Linear(self.attn_emb_dim, self.d_model),
|
40 |
+
nn.ReLU(),
|
41 |
+
nn.Dropout(dropout),
|
42 |
+
nn.LayerNorm(self.d_model)
|
43 |
+
)
|
44 |
+
self.init_params()
|
45 |
+
|
46 |
+
self.freeze = freeze
|
47 |
+
if freeze:
|
48 |
+
for p in self.parameters():
|
49 |
+
p.requires_grad = False
|
50 |
+
|
51 |
+
def init_params(self):
|
52 |
+
for p in self.parameters():
|
53 |
+
if p.dim() > 1:
|
54 |
+
nn.init.xavier_uniform_(p)
|
55 |
+
|
56 |
+
def load_pretrained(self, pretrained, output_fn):
|
57 |
+
checkpoint = torch.load(pretrained, map_location="cpu")
|
58 |
+
|
59 |
+
if "model" in checkpoint:
|
60 |
+
checkpoint = checkpoint["model"]
|
61 |
+
if next(iter(checkpoint)).startswith("decoder."):
|
62 |
+
state_dict = {}
|
63 |
+
for k, v in checkpoint.items():
|
64 |
+
state_dict[k[8:]] = v
|
65 |
+
|
66 |
+
loaded_keys = merge_load_state_dict(state_dict, self, output_fn)
|
67 |
+
if self.freeze:
|
68 |
+
for name, param in self.named_parameters():
|
69 |
+
if name in loaded_keys:
|
70 |
+
param.requires_grad = False
|
71 |
+
else:
|
72 |
+
param.requires_grad = True
|
73 |
+
|
74 |
+
|
75 |
+
def generate_square_subsequent_mask(self, max_length):
|
76 |
+
mask = (torch.triu(torch.ones(max_length, max_length)) == 1).transpose(0, 1)
|
77 |
+
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
|
78 |
+
return mask
|
79 |
+
|
80 |
+
def forward(self, input_dict):
|
81 |
+
word = input_dict["word"]
|
82 |
+
attn_emb = input_dict["attn_emb"]
|
83 |
+
attn_emb_len = input_dict["attn_emb_len"]
|
84 |
+
cap_padding_mask = input_dict["cap_padding_mask"]
|
85 |
+
|
86 |
+
p_attn_emb = self.attn_proj(attn_emb)
|
87 |
+
p_attn_emb = p_attn_emb.transpose(0, 1) # [T_src, N, emb_dim]
|
88 |
+
word = word.to(attn_emb.device)
|
89 |
+
embed = self.in_dropout(self.word_embedding(word)) * math.sqrt(self.emb_dim) # [N, T, emb_dim]
|
90 |
+
embed = embed.transpose(0, 1) # [T, N, emb_dim]
|
91 |
+
embed = self.pos_encoder(embed)
|
92 |
+
|
93 |
+
tgt_mask = self.generate_square_subsequent_mask(embed.size(0)).to(attn_emb.device)
|
94 |
+
memory_key_padding_mask = ~generate_length_mask(attn_emb_len, attn_emb.size(1)).to(attn_emb.device)
|
95 |
+
output = self.model(embed, p_attn_emb, tgt_mask=tgt_mask,
|
96 |
+
tgt_key_padding_mask=cap_padding_mask,
|
97 |
+
memory_key_padding_mask=memory_key_padding_mask)
|
98 |
+
output = output.transpose(0, 1)
|
99 |
+
output = {
|
100 |
+
"embed": output,
|
101 |
+
"logit": self.classifier(output),
|
102 |
+
}
|
103 |
+
return output
|
104 |
+
|
105 |
+
|
106 |
+
class M2TransformerDecoder(BaseDecoder):
|
107 |
+
|
108 |
+
def __init__(self, vocab_size, fc_emb_dim, attn_emb_dim, dropout=0.1, **kwargs):
|
109 |
+
super().__init__(attn_emb_dim, vocab_size, fc_emb_dim, attn_emb_dim, dropout=dropout,)
|
110 |
+
try:
|
111 |
+
from m2transformer.models.transformer import MeshedDecoder
|
112 |
+
except:
|
113 |
+
raise ImportError("meshed-memory-transformer not installed; please run `pip install git+https://github.com/ruotianluo/meshed-memory-transformer.git`")
|
114 |
+
del self.word_embedding
|
115 |
+
del self.in_dropout
|
116 |
+
|
117 |
+
self.d_model = attn_emb_dim
|
118 |
+
self.nhead = kwargs.get("nhead", self.d_model // 64)
|
119 |
+
self.nlayers = kwargs.get("nlayers", 2)
|
120 |
+
self.dim_feedforward = kwargs.get("dim_feedforward", self.d_model * 4)
|
121 |
+
self.model = MeshedDecoder(vocab_size, 100, self.nlayers, 0,
|
122 |
+
d_model=self.d_model,
|
123 |
+
h=self.nhead,
|
124 |
+
d_ff=self.dim_feedforward,
|
125 |
+
dropout=dropout)
|
126 |
+
self.init_params()
|
127 |
+
|
128 |
+
def init_params(self):
|
129 |
+
for p in self.parameters():
|
130 |
+
if p.dim() > 1:
|
131 |
+
nn.init.xavier_uniform_(p)
|
132 |
+
|
133 |
+
def forward(self, input_dict):
|
134 |
+
word = input_dict["word"]
|
135 |
+
attn_emb = input_dict["attn_emb"]
|
136 |
+
attn_emb_mask = input_dict["attn_emb_mask"]
|
137 |
+
word = word.to(attn_emb.device)
|
138 |
+
embed, logit = self.model(word, attn_emb, attn_emb_mask)
|
139 |
+
output = {
|
140 |
+
"embed": embed,
|
141 |
+
"logit": logit,
|
142 |
+
}
|
143 |
+
return output
|
144 |
+
|
145 |
+
|
146 |
+
class EventTransformerDecoder(TransformerDecoder):
|
147 |
+
|
148 |
+
def forward(self, input_dict):
|
149 |
+
word = input_dict["word"] # index of word embeddings
|
150 |
+
attn_emb = input_dict["attn_emb"]
|
151 |
+
attn_emb_len = input_dict["attn_emb_len"]
|
152 |
+
cap_padding_mask = input_dict["cap_padding_mask"]
|
153 |
+
event_emb = input_dict["event"] # [N, emb_dim]
|
154 |
+
|
155 |
+
p_attn_emb = self.attn_proj(attn_emb)
|
156 |
+
p_attn_emb = p_attn_emb.transpose(0, 1) # [T_src, N, emb_dim]
|
157 |
+
word = word.to(attn_emb.device)
|
158 |
+
embed = self.in_dropout(self.word_embedding(word)) * math.sqrt(self.emb_dim) # [N, T, emb_dim]
|
159 |
+
|
160 |
+
embed = embed.transpose(0, 1) # [T, N, emb_dim]
|
161 |
+
embed += event_emb
|
162 |
+
embed = self.pos_encoder(embed)
|
163 |
+
|
164 |
+
tgt_mask = self.generate_square_subsequent_mask(embed.size(0)).to(attn_emb.device)
|
165 |
+
memory_key_padding_mask = ~generate_length_mask(attn_emb_len, attn_emb.size(1)).to(attn_emb.device)
|
166 |
+
output = self.model(embed, p_attn_emb, tgt_mask=tgt_mask,
|
167 |
+
tgt_key_padding_mask=cap_padding_mask,
|
168 |
+
memory_key_padding_mask=memory_key_padding_mask)
|
169 |
+
output = output.transpose(0, 1)
|
170 |
+
output = {
|
171 |
+
"embed": output,
|
172 |
+
"logit": self.classifier(output),
|
173 |
+
}
|
174 |
+
return output
|
175 |
+
|
176 |
+
|
177 |
+
class KeywordProbTransformerDecoder(TransformerDecoder):
|
178 |
+
|
179 |
+
def __init__(self, emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
180 |
+
dropout, keyword_classes_num, **kwargs):
|
181 |
+
super().__init__(emb_dim, vocab_size, fc_emb_dim, attn_emb_dim,
|
182 |
+
dropout, **kwargs)
|
183 |
+
self.keyword_proj = nn.Linear(keyword_classes_num, self.d_model)
|
184 |
+
self.word_keyword_norm = nn.LayerNorm(self.d_model)
|
185 |
+
|
186 |
+
def forward(self, input_dict):
|
187 |
+
word = input_dict["word"] # index of word embeddings
|
188 |
+
attn_emb = input_dict["attn_emb"]
|
189 |
+
attn_emb_len = input_dict["attn_emb_len"]
|
190 |
+
cap_padding_mask = input_dict["cap_padding_mask"]
|
191 |
+
keyword = input_dict["keyword"] # [N, keyword_classes_num]
|
192 |
+
|
193 |
+
p_attn_emb = self.attn_proj(attn_emb)
|
194 |
+
p_attn_emb = p_attn_emb.transpose(0, 1) # [T_src, N, emb_dim]
|
195 |
+
word = word.to(attn_emb.device)
|
196 |
+
embed = self.in_dropout(self.word_embedding(word)) * math.sqrt(self.emb_dim) # [N, T, emb_dim]
|
197 |
+
|
198 |
+
embed = embed.transpose(0, 1) # [T, N, emb_dim]
|
199 |
+
embed += self.keyword_proj(keyword)
|
200 |
+
embed = self.word_keyword_norm(embed)
|
201 |
+
|
202 |
+
embed = self.pos_encoder(embed)
|
203 |
+
|
204 |
+
tgt_mask = self.generate_square_subsequent_mask(embed.size(0)).to(attn_emb.device)
|
205 |
+
memory_key_padding_mask = ~generate_length_mask(attn_emb_len, attn_emb.size(1)).to(attn_emb.device)
|
206 |
+
output = self.model(embed, p_attn_emb, tgt_mask=tgt_mask,
|
207 |
+
tgt_key_padding_mask=cap_padding_mask,
|
208 |
+
memory_key_padding_mask=memory_key_padding_mask)
|
209 |
+
output = output.transpose(0, 1)
|
210 |
+
output = {
|
211 |
+
"embed": output,
|
212 |
+
"logit": self.classifier(output),
|
213 |
+
}
|
214 |
+
return output
|
models/transformer_model.py
ADDED
@@ -0,0 +1,264 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
import random
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
|
6 |
+
from models.base import CaptionModel
|
7 |
+
from utils.model_util import repeat_tensor
|
8 |
+
import models.transformer_decoder
|
9 |
+
|
10 |
+
|
11 |
+
class TransformerModel(CaptionModel):
|
12 |
+
|
13 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
14 |
+
if not hasattr(self, "compatible_decoders"):
|
15 |
+
self.compatible_decoders = (
|
16 |
+
models.transformer_decoder.TransformerDecoder,
|
17 |
+
)
|
18 |
+
super().__init__(encoder, decoder, **kwargs)
|
19 |
+
|
20 |
+
def seq_forward(self, input_dict):
|
21 |
+
cap = input_dict["cap"]
|
22 |
+
cap_padding_mask = (cap == self.pad_idx).to(cap.device)
|
23 |
+
cap_padding_mask = cap_padding_mask[:, :-1]
|
24 |
+
output = self.decoder(
|
25 |
+
{
|
26 |
+
"word": cap[:, :-1],
|
27 |
+
"attn_emb": input_dict["attn_emb"],
|
28 |
+
"attn_emb_len": input_dict["attn_emb_len"],
|
29 |
+
"cap_padding_mask": cap_padding_mask
|
30 |
+
}
|
31 |
+
)
|
32 |
+
return output
|
33 |
+
|
34 |
+
def prepare_decoder_input(self, input_dict, output):
|
35 |
+
decoder_input = {
|
36 |
+
"attn_emb": input_dict["attn_emb"],
|
37 |
+
"attn_emb_len": input_dict["attn_emb_len"]
|
38 |
+
}
|
39 |
+
t = input_dict["t"]
|
40 |
+
|
41 |
+
###############
|
42 |
+
# determine input word
|
43 |
+
################
|
44 |
+
if input_dict["mode"] == "train" and random.random() < input_dict["ss_ratio"]: # training, scheduled sampling
|
45 |
+
word = input_dict["cap"][:, :t+1]
|
46 |
+
else:
|
47 |
+
start_word = torch.tensor([self.start_idx,] * input_dict["attn_emb"].size(0)).unsqueeze(1).long()
|
48 |
+
if t == 0:
|
49 |
+
word = start_word
|
50 |
+
else:
|
51 |
+
word = torch.cat((start_word, output["seq"][:, :t]), dim=-1)
|
52 |
+
# word: [N, T]
|
53 |
+
decoder_input["word"] = word
|
54 |
+
|
55 |
+
cap_padding_mask = (word == self.pad_idx).to(input_dict["attn_emb"].device)
|
56 |
+
decoder_input["cap_padding_mask"] = cap_padding_mask
|
57 |
+
return decoder_input
|
58 |
+
|
59 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
60 |
+
decoder_input = {}
|
61 |
+
t = input_dict["t"]
|
62 |
+
i = input_dict["sample_idx"]
|
63 |
+
beam_size = input_dict["beam_size"]
|
64 |
+
###############
|
65 |
+
# prepare attn embeds
|
66 |
+
################
|
67 |
+
if t == 0:
|
68 |
+
attn_emb = repeat_tensor(input_dict["attn_emb"][i], beam_size)
|
69 |
+
attn_emb_len = repeat_tensor(input_dict["attn_emb_len"][i], beam_size)
|
70 |
+
output_i["attn_emb"] = attn_emb
|
71 |
+
output_i["attn_emb_len"] = attn_emb_len
|
72 |
+
decoder_input["attn_emb"] = output_i["attn_emb"]
|
73 |
+
decoder_input["attn_emb_len"] = output_i["attn_emb_len"]
|
74 |
+
###############
|
75 |
+
# determine input word
|
76 |
+
################
|
77 |
+
start_word = torch.tensor([self.start_idx,] * beam_size).unsqueeze(1).long()
|
78 |
+
if t == 0:
|
79 |
+
word = start_word
|
80 |
+
else:
|
81 |
+
word = torch.cat((start_word, output_i["seq"]), dim=-1)
|
82 |
+
decoder_input["word"] = word
|
83 |
+
cap_padding_mask = (word == self.pad_idx).to(input_dict["attn_emb"].device)
|
84 |
+
decoder_input["cap_padding_mask"] = cap_padding_mask
|
85 |
+
|
86 |
+
return decoder_input
|
87 |
+
|
88 |
+
|
89 |
+
class M2TransformerModel(CaptionModel):
|
90 |
+
|
91 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
92 |
+
if not hasattr(self, "compatible_decoders"):
|
93 |
+
self.compatible_decoders = (
|
94 |
+
models.transformer_decoder.M2TransformerDecoder,
|
95 |
+
)
|
96 |
+
super().__init__(encoder, decoder, **kwargs)
|
97 |
+
self.check_encoder_compatibility()
|
98 |
+
|
99 |
+
def check_encoder_compatibility(self):
|
100 |
+
assert isinstance(self.encoder, models.encoder.M2TransformerEncoder), \
|
101 |
+
f"only M2TransformerModel is compatible with {self.__class__.__name__}"
|
102 |
+
|
103 |
+
def seq_forward(self, input_dict):
|
104 |
+
cap = input_dict["cap"]
|
105 |
+
output = self.decoder(
|
106 |
+
{
|
107 |
+
"word": cap[:, :-1],
|
108 |
+
"attn_emb": input_dict["attn_emb"],
|
109 |
+
"attn_emb_mask": input_dict["attn_emb_mask"],
|
110 |
+
}
|
111 |
+
)
|
112 |
+
return output
|
113 |
+
|
114 |
+
def prepare_decoder_input(self, input_dict, output):
|
115 |
+
decoder_input = {
|
116 |
+
"attn_emb": input_dict["attn_emb"],
|
117 |
+
"attn_emb_mask": input_dict["attn_emb_mask"]
|
118 |
+
}
|
119 |
+
t = input_dict["t"]
|
120 |
+
|
121 |
+
###############
|
122 |
+
# determine input word
|
123 |
+
################
|
124 |
+
if input_dict["mode"] == "train" and random.random() < input_dict["ss_ratio"]: # training, scheduled sampling
|
125 |
+
word = input_dict["cap"][:, :t+1]
|
126 |
+
else:
|
127 |
+
start_word = torch.tensor([self.start_idx,] * input_dict["attn_emb"].size(0)).unsqueeze(1).long()
|
128 |
+
if t == 0:
|
129 |
+
word = start_word
|
130 |
+
else:
|
131 |
+
word = torch.cat((start_word, output["seq"][:, :t]), dim=-1)
|
132 |
+
# word: [N, T]
|
133 |
+
decoder_input["word"] = word
|
134 |
+
|
135 |
+
return decoder_input
|
136 |
+
|
137 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
138 |
+
decoder_input = {}
|
139 |
+
t = input_dict["t"]
|
140 |
+
i = input_dict["sample_idx"]
|
141 |
+
beam_size = input_dict["beam_size"]
|
142 |
+
###############
|
143 |
+
# prepare attn embeds
|
144 |
+
################
|
145 |
+
if t == 0:
|
146 |
+
attn_emb = repeat_tensor(input_dict["attn_emb"][i], beam_size)
|
147 |
+
attn_emb_mask = repeat_tensor(input_dict["attn_emb_mask"][i], beam_size)
|
148 |
+
output_i["attn_emb"] = attn_emb
|
149 |
+
output_i["attn_emb_mask"] = attn_emb_mask
|
150 |
+
decoder_input["attn_emb"] = output_i["attn_emb"]
|
151 |
+
decoder_input["attn_emb_mask"] = output_i["attn_emb_mask"]
|
152 |
+
###############
|
153 |
+
# determine input word
|
154 |
+
################
|
155 |
+
start_word = torch.tensor([self.start_idx,] * beam_size).unsqueeze(1).long()
|
156 |
+
if t == 0:
|
157 |
+
word = start_word
|
158 |
+
else:
|
159 |
+
word = torch.cat((start_word, output_i["seq"]), dim=-1)
|
160 |
+
decoder_input["word"] = word
|
161 |
+
|
162 |
+
return decoder_input
|
163 |
+
|
164 |
+
|
165 |
+
class EventEncoder(nn.Module):
|
166 |
+
"""
|
167 |
+
Encode the Label information in AudioCaps and AudioSet
|
168 |
+
"""
|
169 |
+
def __init__(self, emb_dim, vocab_size=527):
|
170 |
+
super(EventEncoder, self).__init__()
|
171 |
+
self.label_embedding = nn.Parameter(
|
172 |
+
torch.randn((vocab_size, emb_dim)), requires_grad=True)
|
173 |
+
|
174 |
+
def forward(self, word_idxs):
|
175 |
+
indices = word_idxs / word_idxs.sum(dim=1, keepdim=True)
|
176 |
+
embeddings = indices @ self.label_embedding
|
177 |
+
return embeddings
|
178 |
+
|
179 |
+
|
180 |
+
class EventCondTransformerModel(TransformerModel):
|
181 |
+
|
182 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
183 |
+
if not hasattr(self, "compatible_decoders"):
|
184 |
+
self.compatible_decoders = (
|
185 |
+
models.transformer_decoder.EventTransformerDecoder,
|
186 |
+
)
|
187 |
+
super().__init__(encoder, decoder, **kwargs)
|
188 |
+
self.label_encoder = EventEncoder(decoder.emb_dim, 527)
|
189 |
+
self.train_forward_keys += ["events"]
|
190 |
+
self.inference_forward_keys += ["events"]
|
191 |
+
|
192 |
+
# def seq_forward(self, input_dict):
|
193 |
+
# cap = input_dict["cap"]
|
194 |
+
# cap_padding_mask = (cap == self.pad_idx).to(cap.device)
|
195 |
+
# cap_padding_mask = cap_padding_mask[:, :-1]
|
196 |
+
# output = self.decoder(
|
197 |
+
# {
|
198 |
+
# "word": cap[:, :-1],
|
199 |
+
# "attn_emb": input_dict["attn_emb"],
|
200 |
+
# "attn_emb_len": input_dict["attn_emb_len"],
|
201 |
+
# "cap_padding_mask": cap_padding_mask
|
202 |
+
# }
|
203 |
+
# )
|
204 |
+
# return output
|
205 |
+
|
206 |
+
def prepare_decoder_input(self, input_dict, output):
|
207 |
+
decoder_input = super().prepare_decoder_input(input_dict, output)
|
208 |
+
decoder_input["events"] = self.label_encoder(input_dict["events"])
|
209 |
+
return decoder_input
|
210 |
+
|
211 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
212 |
+
decoder_input = super().prepare_beamsearch_decoder_input(input_dict, output_i)
|
213 |
+
t = input_dict["t"]
|
214 |
+
i = input_dict["sample_idx"]
|
215 |
+
beam_size = input_dict["beam_size"]
|
216 |
+
if t == 0:
|
217 |
+
output_i["events"] = repeat_tensor(self.label_encoder(input_dict["events"])[i], beam_size)
|
218 |
+
decoder_input["events"] = output_i["events"]
|
219 |
+
return decoder_input
|
220 |
+
|
221 |
+
|
222 |
+
class KeywordCondTransformerModel(TransformerModel):
|
223 |
+
|
224 |
+
def __init__(self, encoder: nn.Module, decoder: nn.Module, **kwargs):
|
225 |
+
if not hasattr(self, "compatible_decoders"):
|
226 |
+
self.compatible_decoders = (
|
227 |
+
models.transformer_decoder.KeywordProbTransformerDecoder,
|
228 |
+
)
|
229 |
+
super().__init__(encoder, decoder, **kwargs)
|
230 |
+
self.train_forward_keys += ["keyword"]
|
231 |
+
self.inference_forward_keys += ["keyword"]
|
232 |
+
|
233 |
+
def seq_forward(self, input_dict):
|
234 |
+
cap = input_dict["cap"]
|
235 |
+
cap_padding_mask = (cap == self.pad_idx).to(cap.device)
|
236 |
+
cap_padding_mask = cap_padding_mask[:, :-1]
|
237 |
+
keyword = input_dict["keyword"]
|
238 |
+
output = self.decoder(
|
239 |
+
{
|
240 |
+
"word": cap[:, :-1],
|
241 |
+
"attn_emb": input_dict["attn_emb"],
|
242 |
+
"attn_emb_len": input_dict["attn_emb_len"],
|
243 |
+
"keyword": keyword,
|
244 |
+
"cap_padding_mask": cap_padding_mask
|
245 |
+
}
|
246 |
+
)
|
247 |
+
return output
|
248 |
+
|
249 |
+
def prepare_decoder_input(self, input_dict, output):
|
250 |
+
decoder_input = super().prepare_decoder_input(input_dict, output)
|
251 |
+
decoder_input["keyword"] = input_dict["keyword"]
|
252 |
+
return decoder_input
|
253 |
+
|
254 |
+
def prepare_beamsearch_decoder_input(self, input_dict, output_i):
|
255 |
+
decoder_input = super().prepare_beamsearch_decoder_input(input_dict, output_i)
|
256 |
+
t = input_dict["t"]
|
257 |
+
i = input_dict["sample_idx"]
|
258 |
+
beam_size = input_dict["beam_size"]
|
259 |
+
if t == 0:
|
260 |
+
output_i["keyword"] = repeat_tensor(input_dict["keyword"][i],
|
261 |
+
beam_size)
|
262 |
+
decoder_input["keyword"] = output_i["keyword"]
|
263 |
+
return decoder_input
|
264 |
+
|
requirements.txt
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
efficientnet_pytorch
|
2 |
+
PyYAML
|
text_tokenizer.py
ADDED
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pickle
|
2 |
+
from pathlib import Path
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
from utils.train_util import pad_sequence
|
6 |
+
|
7 |
+
|
8 |
+
class DictTokenizer:
|
9 |
+
|
10 |
+
def __init__(self,
|
11 |
+
tokenizer_path: str = None,
|
12 |
+
max_length: int = 20) -> None:
|
13 |
+
self.word2idx = {}
|
14 |
+
self.idx2word = {}
|
15 |
+
self.idx = 0
|
16 |
+
self.add_word("<pad>")
|
17 |
+
self.add_word("<start>")
|
18 |
+
self.add_word("<end>")
|
19 |
+
self.add_word("<unk>")
|
20 |
+
if tokenizer_path is not None and Path(tokenizer_path).exists():
|
21 |
+
state_dict = pickle.load(open(tokenizer_path, "rb"))
|
22 |
+
self.load_state_dict(state_dict)
|
23 |
+
self.loaded = True
|
24 |
+
else:
|
25 |
+
self.loaded = False
|
26 |
+
self.bos, self.eos = self.word2idx["<start>"], self.word2idx["<end>"]
|
27 |
+
self.pad = self.word2idx["<pad>"]
|
28 |
+
self.max_length = max_length
|
29 |
+
|
30 |
+
def add_word(self, word):
|
31 |
+
if not word in self.word2idx:
|
32 |
+
self.word2idx[word] = self.idx
|
33 |
+
self.idx2word[self.idx] = word
|
34 |
+
self.idx += 1
|
35 |
+
|
36 |
+
def encode_word(self, word):
|
37 |
+
if word in self.word2idx:
|
38 |
+
return self.word2idx[word]
|
39 |
+
else:
|
40 |
+
return self.word2idx["<unk>"]
|
41 |
+
|
42 |
+
def __call__(self, texts):
|
43 |
+
assert isinstance(texts, list), "the input must be List[str]"
|
44 |
+
batch_tokens = []
|
45 |
+
for text in texts:
|
46 |
+
tokens = [self.encode_word(token) for token in text.split()][:self.max_length]
|
47 |
+
tokens = [self.bos] + tokens + [self.eos]
|
48 |
+
tokens = np.array(tokens)
|
49 |
+
batch_tokens.append(tokens)
|
50 |
+
caps, cap_lens = pad_sequence(batch_tokens, self.pad)
|
51 |
+
return {
|
52 |
+
"cap": caps,
|
53 |
+
"cap_len": cap_lens
|
54 |
+
}
|
55 |
+
|
56 |
+
def decode(self, batch_token_ids):
|
57 |
+
output = []
|
58 |
+
for token_ids in batch_token_ids:
|
59 |
+
tokens = []
|
60 |
+
for token_id in token_ids:
|
61 |
+
if token_id == self.eos:
|
62 |
+
break
|
63 |
+
elif token_id == self.bos:
|
64 |
+
continue
|
65 |
+
tokens.append(self.idx2word[token_id])
|
66 |
+
output.append(" ".join(tokens))
|
67 |
+
return output
|
68 |
+
|
69 |
+
def __len__(self):
|
70 |
+
return len(self.word2idx)
|
71 |
+
|
72 |
+
def state_dict(self):
|
73 |
+
return self.word2idx
|
74 |
+
|
75 |
+
def load_state_dict(self, state_dict):
|
76 |
+
self.word2idx = state_dict
|
77 |
+
self.idx2word = {idx: word for word, idx in self.word2idx.items()}
|
78 |
+
self.idx = len(self.word2idx)
|
79 |
+
|
80 |
+
|
81 |
+
class HuggingfaceTokenizer:
|
82 |
+
|
83 |
+
def __init__(self,
|
84 |
+
model_name_or_path,
|
85 |
+
max_length) -> None:
|
86 |
+
from transformers import AutoTokenizer
|
87 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
|
88 |
+
self.max_length = max_length
|
89 |
+
self.bos, self.eos = self.tokenizer.bos_token_id, self.tokenizer.eos_token_id
|
90 |
+
self.pad = self.tokenizer.pad_token_id
|
91 |
+
self.loaded = True
|
92 |
+
|
93 |
+
def __call__(self, texts):
|
94 |
+
assert isinstance(texts, list), "the input must be List[str]"
|
95 |
+
batch_token_dict = self.tokenizer(texts,
|
96 |
+
padding=True,
|
97 |
+
truncation=True,
|
98 |
+
max_length=self.max_length,
|
99 |
+
return_tensors="pt")
|
100 |
+
batch_token_dict["cap"] = batch_token_dict["input_ids"]
|
101 |
+
cap_lens = batch_token_dict["attention_mask"].sum(dim=1)
|
102 |
+
cap_lens = cap_lens.numpy().astype(np.int32)
|
103 |
+
batch_token_dict["cap_len"] = cap_lens
|
104 |
+
return batch_token_dict
|
105 |
+
|
106 |
+
def decode(self, batch_token_ids):
|
107 |
+
return self.tokenizer.batch_decode(batch_token_ids, skip_special_tokens=True)
|
utils/model_util.py
ADDED
@@ -0,0 +1,186 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
|
7 |
+
from torch.nn.utils.rnn import PackedSequence, pack_padded_sequence, pad_packed_sequence
|
8 |
+
|
9 |
+
|
10 |
+
def sort_pack_padded_sequence(input, lengths):
|
11 |
+
sorted_lengths, indices = torch.sort(lengths, descending=True)
|
12 |
+
tmp = pack_padded_sequence(input[indices], sorted_lengths.cpu(), batch_first=True)
|
13 |
+
inv_ix = indices.clone()
|
14 |
+
inv_ix[indices] = torch.arange(0,len(indices)).type_as(inv_ix)
|
15 |
+
return tmp, inv_ix
|
16 |
+
|
17 |
+
def pad_unsort_packed_sequence(input, inv_ix):
|
18 |
+
tmp, _ = pad_packed_sequence(input, batch_first=True)
|
19 |
+
tmp = tmp[inv_ix]
|
20 |
+
return tmp
|
21 |
+
|
22 |
+
def pack_wrapper(module, attn_feats, attn_feat_lens):
|
23 |
+
packed, inv_ix = sort_pack_padded_sequence(attn_feats, attn_feat_lens)
|
24 |
+
if isinstance(module, torch.nn.RNNBase):
|
25 |
+
return pad_unsort_packed_sequence(module(packed)[0], inv_ix)
|
26 |
+
else:
|
27 |
+
return pad_unsort_packed_sequence(PackedSequence(module(packed[0]), packed[1]), inv_ix)
|
28 |
+
|
29 |
+
def generate_length_mask(lens, max_length=None):
|
30 |
+
lens = torch.as_tensor(lens)
|
31 |
+
N = lens.size(0)
|
32 |
+
if max_length is None:
|
33 |
+
max_length = max(lens)
|
34 |
+
if isinstance(max_length, torch.Tensor):
|
35 |
+
max_length = max_length.item()
|
36 |
+
idxs = torch.arange(max_length).repeat(N).view(N, max_length)
|
37 |
+
idxs = idxs.to(lens.device)
|
38 |
+
mask = (idxs < lens.view(-1, 1))
|
39 |
+
return mask
|
40 |
+
|
41 |
+
def mean_with_lens(features, lens):
|
42 |
+
"""
|
43 |
+
features: [N, T, ...] (assume the second dimension represents length)
|
44 |
+
lens: [N,]
|
45 |
+
"""
|
46 |
+
lens = torch.as_tensor(lens)
|
47 |
+
if max(lens) != features.size(1):
|
48 |
+
max_length = features.size(1)
|
49 |
+
mask = generate_length_mask(lens, max_length)
|
50 |
+
else:
|
51 |
+
mask = generate_length_mask(lens)
|
52 |
+
mask = mask.to(features.device) # [N, T]
|
53 |
+
|
54 |
+
while mask.ndim < features.ndim:
|
55 |
+
mask = mask.unsqueeze(-1)
|
56 |
+
feature_mean = features * mask
|
57 |
+
feature_mean = feature_mean.sum(1)
|
58 |
+
while lens.ndim < feature_mean.ndim:
|
59 |
+
lens = lens.unsqueeze(1)
|
60 |
+
feature_mean = feature_mean / lens.to(features.device)
|
61 |
+
# feature_mean = features * mask.unsqueeze(-1)
|
62 |
+
# feature_mean = feature_mean.sum(1) / lens.unsqueeze(1).to(features.device)
|
63 |
+
return feature_mean
|
64 |
+
|
65 |
+
def max_with_lens(features, lens):
|
66 |
+
"""
|
67 |
+
features: [N, T, ...] (assume the second dimension represents length)
|
68 |
+
lens: [N,]
|
69 |
+
"""
|
70 |
+
lens = torch.as_tensor(lens)
|
71 |
+
if max(lens) != features.size(1):
|
72 |
+
max_length = features.size(1)
|
73 |
+
mask = generate_length_mask(lens, max_length)
|
74 |
+
else:
|
75 |
+
mask = generate_length_mask(lens)
|
76 |
+
mask = mask.to(features.device) # [N, T]
|
77 |
+
|
78 |
+
feature_max = features.clone()
|
79 |
+
feature_max[~mask] = float("-inf")
|
80 |
+
feature_max, _ = feature_max.max(1)
|
81 |
+
return feature_max
|
82 |
+
|
83 |
+
def repeat_tensor(x, n):
|
84 |
+
return x.unsqueeze(0).repeat(n, *([1] * len(x.shape)))
|
85 |
+
|
86 |
+
def init(m, method="kaiming"):
|
87 |
+
if isinstance(m, (nn.Conv2d, nn.Conv1d)):
|
88 |
+
if method == "kaiming":
|
89 |
+
nn.init.kaiming_uniform_(m.weight)
|
90 |
+
elif method == "xavier":
|
91 |
+
nn.init.xavier_uniform_(m.weight)
|
92 |
+
else:
|
93 |
+
raise Exception(f"initialization method {method} not supported")
|
94 |
+
if m.bias is not None:
|
95 |
+
nn.init.constant_(m.bias, 0)
|
96 |
+
elif isinstance(m, (nn.BatchNorm2d, nn.BatchNorm1d)):
|
97 |
+
nn.init.constant_(m.weight, 1)
|
98 |
+
if m.bias is not None:
|
99 |
+
nn.init.constant_(m.bias, 0)
|
100 |
+
elif isinstance(m, nn.Linear):
|
101 |
+
if method == "kaiming":
|
102 |
+
nn.init.kaiming_uniform_(m.weight)
|
103 |
+
elif method == "xavier":
|
104 |
+
nn.init.xavier_uniform_(m.weight)
|
105 |
+
else:
|
106 |
+
raise Exception(f"initialization method {method} not supported")
|
107 |
+
if m.bias is not None:
|
108 |
+
nn.init.constant_(m.bias, 0)
|
109 |
+
elif isinstance(m, nn.Embedding):
|
110 |
+
if method == "kaiming":
|
111 |
+
nn.init.kaiming_uniform_(m.weight)
|
112 |
+
elif method == "xavier":
|
113 |
+
nn.init.xavier_uniform_(m.weight)
|
114 |
+
else:
|
115 |
+
raise Exception(f"initialization method {method} not supported")
|
116 |
+
|
117 |
+
def compute_batch_score(decode_res,
|
118 |
+
key2refs,
|
119 |
+
keys,
|
120 |
+
start_idx,
|
121 |
+
end_idx,
|
122 |
+
vocabulary,
|
123 |
+
scorer):
|
124 |
+
"""
|
125 |
+
Args:
|
126 |
+
decode_res: decoding results of model, [N, max_length]
|
127 |
+
key2refs: references of all samples, dict(<key> -> [ref_1, ref_2, ..., ref_n]
|
128 |
+
keys: keys of this batch, used to match decode results and refs
|
129 |
+
Return:
|
130 |
+
scores of this batch, [N,]
|
131 |
+
"""
|
132 |
+
|
133 |
+
if scorer is None:
|
134 |
+
from pycocoevalcap.cider.cider import Cider
|
135 |
+
scorer = Cider()
|
136 |
+
|
137 |
+
hypothesis = {}
|
138 |
+
references = {}
|
139 |
+
|
140 |
+
for i in range(len(keys)):
|
141 |
+
|
142 |
+
if keys[i] in hypothesis.keys():
|
143 |
+
continue
|
144 |
+
|
145 |
+
# prepare candidate sentence
|
146 |
+
candidate = []
|
147 |
+
for w_t in decode_res[i]:
|
148 |
+
if w_t == start_idx:
|
149 |
+
continue
|
150 |
+
elif w_t == end_idx:
|
151 |
+
break
|
152 |
+
candidate.append(vocabulary.idx2word[w_t])
|
153 |
+
|
154 |
+
hypothesis[keys[i]] = [" ".join(candidate), ]
|
155 |
+
|
156 |
+
# prepare reference sentences
|
157 |
+
references[keys[i]] = key2refs[keys[i]]
|
158 |
+
|
159 |
+
score, scores = scorer.compute_score(references, hypothesis)
|
160 |
+
key2score = {key: scores[i] for i, key in enumerate(references.keys())}
|
161 |
+
results = np.zeros(decode_res.shape[0])
|
162 |
+
for i in range(decode_res.shape[0]):
|
163 |
+
results[i] = key2score[keys[i]]
|
164 |
+
return results
|
165 |
+
|
166 |
+
|
167 |
+
class PositionalEncoding(nn.Module):
|
168 |
+
|
169 |
+
def __init__(self, d_model, dropout=0.1, max_len=100):
|
170 |
+
super(PositionalEncoding, self).__init__()
|
171 |
+
self.dropout = nn.Dropout(p=dropout)
|
172 |
+
|
173 |
+
pe = torch.zeros(max_len, d_model)
|
174 |
+
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
|
175 |
+
div_term = torch.exp(torch.arange(0, d_model, 2).float() * \
|
176 |
+
(-math.log(10000.0) / d_model))
|
177 |
+
pe[:, 0::2] = torch.sin(position * div_term)
|
178 |
+
pe[:, 1::2] = torch.cos(position * div_term)
|
179 |
+
pe = pe.unsqueeze(0).transpose(0, 1)
|
180 |
+
# self.register_buffer("pe", pe)
|
181 |
+
self.register_parameter("pe", nn.Parameter(pe, requires_grad=False))
|
182 |
+
|
183 |
+
def forward(self, x):
|
184 |
+
# x: [T, N, E]
|
185 |
+
x = x + self.pe[:x.size(0), :]
|
186 |
+
return self.dropout(x)
|
utils/train_util.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import importlib
|
2 |
+
import os
|
3 |
+
import sys
|
4 |
+
from typing import Callable, Dict, Union
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
import yaml
|
8 |
+
import torch
|
9 |
+
|
10 |
+
|
11 |
+
def merge_a_into_b(a, b):
|
12 |
+
# merge dict a into dict b. values in a will overwrite b.
|
13 |
+
for k, v in a.items():
|
14 |
+
if isinstance(v, dict) and k in b:
|
15 |
+
assert isinstance(
|
16 |
+
b[k], dict
|
17 |
+
), "Cannot inherit key '{}' from base!".format(k)
|
18 |
+
merge_a_into_b(v, b[k])
|
19 |
+
else:
|
20 |
+
b[k] = v
|
21 |
+
|
22 |
+
|
23 |
+
def load_config(config_file):
|
24 |
+
with open(config_file, "r") as reader:
|
25 |
+
config = yaml.load(reader, Loader=yaml.FullLoader)
|
26 |
+
if "inherit_from" in config:
|
27 |
+
base_config_file = config["inherit_from"]
|
28 |
+
base_config_file = os.path.join(
|
29 |
+
os.path.dirname(config_file), base_config_file
|
30 |
+
)
|
31 |
+
assert not os.path.samefile(config_file, base_config_file), \
|
32 |
+
"inherit from itself"
|
33 |
+
base_config = load_config(base_config_file)
|
34 |
+
del config["inherit_from"]
|
35 |
+
merge_a_into_b(config, base_config)
|
36 |
+
return base_config
|
37 |
+
return config
|
38 |
+
|
39 |
+
def get_cls_from_str(string, reload=False):
|
40 |
+
module_name, cls_name = string.rsplit(".", 1)
|
41 |
+
if reload:
|
42 |
+
module_imp = importlib.import_module(module_name)
|
43 |
+
importlib.reload(module_imp)
|
44 |
+
return getattr(importlib.import_module(module_name, package=None), cls_name)
|
45 |
+
|
46 |
+
def init_obj_from_dict(config, **kwargs):
|
47 |
+
obj_args = config["args"].copy()
|
48 |
+
obj_args.update(kwargs)
|
49 |
+
for k in config:
|
50 |
+
if k not in ["type", "args"] and isinstance(config[k], dict) and k not in kwargs:
|
51 |
+
obj_args[k] = init_obj_from_dict(config[k])
|
52 |
+
try:
|
53 |
+
obj = get_cls_from_str(config["type"])(**obj_args)
|
54 |
+
return obj
|
55 |
+
except Exception as e:
|
56 |
+
print(f"Initializing {config} failed, detailed error stack: ")
|
57 |
+
raise e
|
58 |
+
|
59 |
+
def init_model_from_config(config, print_fn=sys.stdout.write):
|
60 |
+
kwargs = {}
|
61 |
+
for k in config:
|
62 |
+
if k not in ["type", "args", "pretrained"]:
|
63 |
+
sub_model = init_model_from_config(config[k], print_fn)
|
64 |
+
if "pretrained" in config[k]:
|
65 |
+
load_pretrained_model(sub_model,
|
66 |
+
config[k]["pretrained"],
|
67 |
+
print_fn)
|
68 |
+
kwargs[k] = sub_model
|
69 |
+
model = init_obj_from_dict(config, **kwargs)
|
70 |
+
return model
|
71 |
+
|
72 |
+
def merge_load_state_dict(state_dict,
|
73 |
+
model: torch.nn.Module,
|
74 |
+
output_fn: Callable = sys.stdout.write):
|
75 |
+
model_dict = model.state_dict()
|
76 |
+
pretrained_dict = {}
|
77 |
+
mismatch_keys = []
|
78 |
+
for key, value in state_dict.items():
|
79 |
+
if key in model_dict and model_dict[key].shape == value.shape:
|
80 |
+
pretrained_dict[key] = value
|
81 |
+
else:
|
82 |
+
mismatch_keys.append(key)
|
83 |
+
output_fn(f"Loading pre-trained model, with mismatched keys {mismatch_keys}")
|
84 |
+
model_dict.update(pretrained_dict)
|
85 |
+
model.load_state_dict(model_dict, strict=True)
|
86 |
+
return pretrained_dict.keys()
|
87 |
+
|
88 |
+
|
89 |
+
def load_pretrained_model(model: torch.nn.Module,
|
90 |
+
pretrained: Union[str, Dict],
|
91 |
+
output_fn: Callable = sys.stdout.write):
|
92 |
+
if not isinstance(pretrained, dict) and not os.path.exists(pretrained):
|
93 |
+
output_fn(f"pretrained {pretrained} not exist!")
|
94 |
+
return
|
95 |
+
|
96 |
+
if hasattr(model, "load_pretrained"):
|
97 |
+
model.load_pretrained(pretrained, output_fn)
|
98 |
+
return
|
99 |
+
|
100 |
+
if isinstance(pretrained, dict):
|
101 |
+
state_dict = pretrained
|
102 |
+
else:
|
103 |
+
state_dict = torch.load(pretrained, map_location="cpu")
|
104 |
+
|
105 |
+
if "model" in state_dict:
|
106 |
+
state_dict = state_dict["model"]
|
107 |
+
|
108 |
+
merge_load_state_dict(state_dict, model, output_fn)
|
109 |
+
|
110 |
+
def pad_sequence(data, pad_value=0):
|
111 |
+
if isinstance(data[0], (np.ndarray, torch.Tensor)):
|
112 |
+
data = [torch.as_tensor(arr) for arr in data]
|
113 |
+
padded_seq = torch.nn.utils.rnn.pad_sequence(data,
|
114 |
+
batch_first=True,
|
115 |
+
padding_value=pad_value)
|
116 |
+
length = np.array([x.shape[0] for x in data])
|
117 |
+
return padded_seq, length
|