Commit
•
16b7417
1
Parent(s):
6d3f4b0
First model version
Browse files- .gitignore +1 -0
- README.md +28 -0
- config.json +71 -0
- convert.py +150 -0
- pytorch_model.bin +3 -0
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
content-vec-best-legacy-500.pt
|
README.md
CHANGED
@@ -1,3 +1,31 @@
|
|
1 |
---
|
2 |
license: mit
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
---
|
4 |
+
|
5 |
+
# Content Vec Best
|
6 |
+
Official Repo: [ContentVec](https://github.com/auspicious3000/contentvec)
|
7 |
+
This repo brings fairseq ContentVec model to HuggingFace Transformers.
|
8 |
+
|
9 |
+
## How to use
|
10 |
+
To use this model, you need to define
|
11 |
+
```python
|
12 |
+
class HubertModelWithFinalProj(HubertModel):
|
13 |
+
def __init__(self, config):
|
14 |
+
super().__init__(config)
|
15 |
+
|
16 |
+
self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
|
17 |
+
```
|
18 |
+
|
19 |
+
and then load the model with
|
20 |
+
```python
|
21 |
+
model = HubertModelWithFinalProj.from_pretrained("lengyue233/content-vec-best")
|
22 |
+
|
23 |
+
x = model(audio)["last_hidden_state"]
|
24 |
+
x = model.final_proj(x)
|
25 |
+
```
|
26 |
+
|
27 |
+
## How to convert
|
28 |
+
You need to download the ContentVec_legacy model from the official repo, and then run
|
29 |
+
```bash
|
30 |
+
python convert.py
|
31 |
+
```
|
config.json
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"activation_dropout": 0.1,
|
3 |
+
"apply_spec_augment": true,
|
4 |
+
"architectures": [
|
5 |
+
"HubertModelWithFinalProj"
|
6 |
+
],
|
7 |
+
"attention_dropout": 0.1,
|
8 |
+
"bos_token_id": 1,
|
9 |
+
"classifier_proj_size": 256,
|
10 |
+
"conv_bias": false,
|
11 |
+
"conv_dim": [
|
12 |
+
512,
|
13 |
+
512,
|
14 |
+
512,
|
15 |
+
512,
|
16 |
+
512,
|
17 |
+
512,
|
18 |
+
512
|
19 |
+
],
|
20 |
+
"conv_kernel": [
|
21 |
+
10,
|
22 |
+
3,
|
23 |
+
3,
|
24 |
+
3,
|
25 |
+
3,
|
26 |
+
2,
|
27 |
+
2
|
28 |
+
],
|
29 |
+
"conv_stride": [
|
30 |
+
5,
|
31 |
+
2,
|
32 |
+
2,
|
33 |
+
2,
|
34 |
+
2,
|
35 |
+
2,
|
36 |
+
2
|
37 |
+
],
|
38 |
+
"ctc_loss_reduction": "sum",
|
39 |
+
"ctc_zero_infinity": false,
|
40 |
+
"do_stable_layer_norm": false,
|
41 |
+
"eos_token_id": 2,
|
42 |
+
"feat_extract_activation": "gelu",
|
43 |
+
"feat_extract_norm": "group",
|
44 |
+
"feat_proj_dropout": 0.0,
|
45 |
+
"feat_proj_layer_norm": true,
|
46 |
+
"final_dropout": 0.1,
|
47 |
+
"hidden_act": "gelu",
|
48 |
+
"hidden_dropout": 0.1,
|
49 |
+
"hidden_size": 768,
|
50 |
+
"initializer_range": 0.02,
|
51 |
+
"intermediate_size": 3072,
|
52 |
+
"layer_norm_eps": 1e-05,
|
53 |
+
"layerdrop": 0.1,
|
54 |
+
"mask_feature_length": 10,
|
55 |
+
"mask_feature_min_masks": 0,
|
56 |
+
"mask_feature_prob": 0.0,
|
57 |
+
"mask_time_length": 10,
|
58 |
+
"mask_time_min_masks": 2,
|
59 |
+
"mask_time_prob": 0.05,
|
60 |
+
"model_type": "hubert",
|
61 |
+
"num_attention_heads": 12,
|
62 |
+
"num_conv_pos_embedding_groups": 16,
|
63 |
+
"num_conv_pos_embeddings": 128,
|
64 |
+
"num_feat_extract_layers": 7,
|
65 |
+
"num_hidden_layers": 12,
|
66 |
+
"pad_token_id": 0,
|
67 |
+
"torch_dtype": "float32",
|
68 |
+
"transformers_version": "4.27.3",
|
69 |
+
"use_weighted_layer_sum": false,
|
70 |
+
"vocab_size": 32
|
71 |
+
}
|
convert.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn
|
3 |
+
from transformers import HubertConfig, HubertModel
|
4 |
+
import logging
|
5 |
+
|
6 |
+
# Ignore fairseq's logger
|
7 |
+
logging.getLogger("fairseq").setLevel(logging.WARNING)
|
8 |
+
logging.getLogger("torch.distributed.nn.jit.instantiator").setLevel(logging.WARNING)
|
9 |
+
|
10 |
+
from fairseq import checkpoint_utils
|
11 |
+
|
12 |
+
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
|
13 |
+
["content-vec-best-legacy-500.pt"], suffix=""
|
14 |
+
)
|
15 |
+
model = models[0]
|
16 |
+
model.eval()
|
17 |
+
model.eval()
|
18 |
+
|
19 |
+
|
20 |
+
class HubertModelWithFinalProj(HubertModel):
|
21 |
+
def __init__(self, config):
|
22 |
+
super().__init__(config)
|
23 |
+
|
24 |
+
self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
|
25 |
+
|
26 |
+
|
27 |
+
# Default Config
|
28 |
+
hubert = HubertModelWithFinalProj(HubertConfig())
|
29 |
+
|
30 |
+
# huggingface: fairseq
|
31 |
+
mapping = {
|
32 |
+
"masked_spec_embed": "mask_emb",
|
33 |
+
"encoder.layer_norm.bias": "encoder.layer_norm.bias",
|
34 |
+
"encoder.layer_norm.weight": "encoder.layer_norm.weight",
|
35 |
+
"encoder.pos_conv_embed.conv.bias": "encoder.pos_conv.0.bias",
|
36 |
+
"encoder.pos_conv_embed.conv.weight_g": "encoder.pos_conv.0.weight_g",
|
37 |
+
"encoder.pos_conv_embed.conv.weight_v": "encoder.pos_conv.0.weight_v",
|
38 |
+
"feature_projection.layer_norm.bias": "layer_norm.bias",
|
39 |
+
"feature_projection.layer_norm.weight": "layer_norm.weight",
|
40 |
+
"feature_projection.projection.bias": "post_extract_proj.bias",
|
41 |
+
"feature_projection.projection.weight": "post_extract_proj.weight",
|
42 |
+
"final_proj.bias": "final_proj.bias",
|
43 |
+
"final_proj.weight": "final_proj.weight",
|
44 |
+
}
|
45 |
+
|
46 |
+
# Convert encoder
|
47 |
+
for layer in range(12):
|
48 |
+
for j in ["q", "k", "v"]:
|
49 |
+
mapping[
|
50 |
+
f"encoder.layers.{layer}.attention.{j}_proj.weight"
|
51 |
+
] = f"encoder.layers.{layer}.self_attn.{j}_proj.weight"
|
52 |
+
mapping[
|
53 |
+
f"encoder.layers.{layer}.attention.{j}_proj.bias"
|
54 |
+
] = f"encoder.layers.{layer}.self_attn.{j}_proj.bias"
|
55 |
+
|
56 |
+
mapping[
|
57 |
+
f"encoder.layers.{layer}.final_layer_norm.bias"
|
58 |
+
] = f"encoder.layers.{layer}.final_layer_norm.bias"
|
59 |
+
mapping[
|
60 |
+
f"encoder.layers.{layer}.final_layer_norm.weight"
|
61 |
+
] = f"encoder.layers.{layer}.final_layer_norm.weight"
|
62 |
+
|
63 |
+
mapping[
|
64 |
+
f"encoder.layers.{layer}.layer_norm.bias"
|
65 |
+
] = f"encoder.layers.{layer}.self_attn_layer_norm.bias"
|
66 |
+
mapping[
|
67 |
+
f"encoder.layers.{layer}.layer_norm.weight"
|
68 |
+
] = f"encoder.layers.{layer}.self_attn_layer_norm.weight"
|
69 |
+
|
70 |
+
mapping[
|
71 |
+
f"encoder.layers.{layer}.attention.out_proj.bias"
|
72 |
+
] = f"encoder.layers.{layer}.self_attn.out_proj.bias"
|
73 |
+
mapping[
|
74 |
+
f"encoder.layers.{layer}.attention.out_proj.weight"
|
75 |
+
] = f"encoder.layers.{layer}.self_attn.out_proj.weight"
|
76 |
+
|
77 |
+
mapping[
|
78 |
+
f"encoder.layers.{layer}.feed_forward.intermediate_dense.bias"
|
79 |
+
] = f"encoder.layers.{layer}.fc1.bias"
|
80 |
+
mapping[
|
81 |
+
f"encoder.layers.{layer}.feed_forward.intermediate_dense.weight"
|
82 |
+
] = f"encoder.layers.{layer}.fc1.weight"
|
83 |
+
|
84 |
+
mapping[
|
85 |
+
f"encoder.layers.{layer}.feed_forward.output_dense.bias"
|
86 |
+
] = f"encoder.layers.{layer}.fc2.bias"
|
87 |
+
mapping[
|
88 |
+
f"encoder.layers.{layer}.feed_forward.output_dense.weight"
|
89 |
+
] = f"encoder.layers.{layer}.fc2.weight"
|
90 |
+
|
91 |
+
# Convert Conv Layers
|
92 |
+
for layer in range(7):
|
93 |
+
mapping[
|
94 |
+
f"feature_extractor.conv_layers.{layer}.conv.weight"
|
95 |
+
] = f"feature_extractor.conv_layers.{layer}.0.weight"
|
96 |
+
|
97 |
+
if layer != 0:
|
98 |
+
continue
|
99 |
+
|
100 |
+
mapping[
|
101 |
+
f"feature_extractor.conv_layers.{layer}.layer_norm.weight"
|
102 |
+
] = f"feature_extractor.conv_layers.{layer}.2.weight"
|
103 |
+
mapping[
|
104 |
+
f"feature_extractor.conv_layers.{layer}.layer_norm.bias"
|
105 |
+
] = f"feature_extractor.conv_layers.{layer}.2.bias"
|
106 |
+
|
107 |
+
hf_keys = set(hubert.state_dict().keys())
|
108 |
+
fair_keys = set(model.state_dict().keys())
|
109 |
+
|
110 |
+
hf_keys -= set(mapping.keys())
|
111 |
+
fair_keys -= set(mapping.values())
|
112 |
+
|
113 |
+
for i, j in zip(sorted(hf_keys), sorted(fair_keys)):
|
114 |
+
print(i, j)
|
115 |
+
|
116 |
+
print(hf_keys, fair_keys)
|
117 |
+
print(len(hf_keys), len(fair_keys))
|
118 |
+
|
119 |
+
# try loading the weights
|
120 |
+
new_state_dict = {}
|
121 |
+
for k, v in mapping.items():
|
122 |
+
new_state_dict[k] = model.state_dict()[v]
|
123 |
+
|
124 |
+
x = hubert.load_state_dict(new_state_dict, strict=False)
|
125 |
+
print(x)
|
126 |
+
hubert.eval()
|
127 |
+
|
128 |
+
with torch.no_grad():
|
129 |
+
new_input = torch.randn(1, 16384)
|
130 |
+
|
131 |
+
result1 = hubert(new_input, output_hidden_states=True)["hidden_states"][9]
|
132 |
+
result1 = hubert.final_proj(result1)
|
133 |
+
|
134 |
+
result2 = model.extract_features(
|
135 |
+
**{
|
136 |
+
"source": new_input,
|
137 |
+
"padding_mask": torch.zeros(1, 16384, dtype=torch.bool),
|
138 |
+
# "features_only": True,
|
139 |
+
"output_layer": 9,
|
140 |
+
}
|
141 |
+
)[0]
|
142 |
+
result2 = model.final_proj(result2)
|
143 |
+
|
144 |
+
assert torch.allclose(result1, result2, atol=1e-3)
|
145 |
+
|
146 |
+
print("Sanity check passed")
|
147 |
+
|
148 |
+
# Save huggingface model
|
149 |
+
hubert.save_pretrained(".")
|
150 |
+
print("Saved model")
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d8dd400e054ddf4e6be75dab5a2549db748cc99e756a097c496c099f65a4854e
|
3 |
+
size 378342945
|