-
Notifications
You must be signed in to change notification settings - Fork 1
/
hubert.py
257 lines (239 loc) · 12.6 KB
/
hubert.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
from typing import Optional
import mindspore
import numpy as np
from mindspore import dtype as mstype
from mindspore import nn, ops
from mindspore.common.initializer import Uniform, initializer
from mindspore.common.tensor import Tensor
from mindspore.nn import DynamicLossScaleUpdateCell
from mindspore.nn.wrap import TrainOneStepWithLossScaleCell
from mindspore.ops import composite as C
from mindspore.ops import constexpr
from mindspore.ops import functional as F
from mindspore.ops import operations as P
from mindspore.common.initializer import HeUniform, Uniform, _calculate_correct_fan, initializer
from mindspore.nn.cell import Cell
import math
from .feature_extractor import ConvFeatureExtractionModel
from .layernorm import LayerNorm
from .encoder import Wav2VecTransformerEncoder
from .dropout import Dropout
class Dense(Cell):
"""A self-defined dense layer.
Args:
in_channel (int): The number of channels in the input space.
out_channel (int): The number of channels in the output space.
has_bias (bool): The trainable bias_init parameter.
activation (nn.Cell): activate function applied to the output of the fully connected layer.
negative_slope (int, float, bool): The negative slope of the rectifier used after this layer.
mode (str): Either "fan_in" or "fan_out".
nonlinerity (str): The non-linear function, use only with "relu" or "leaky_relu"
"""
def __init__(self,
in_channel,
out_channel,
has_bias=True,
activation=None,
negative_slope=math.sqrt(5),
mode='fan_in',
nonlinearity='leaky_relu',
type=mindspore.float32):
super(Dense, self).__init__()
kaiming_uniform_0 = initializer(HeUniform(negative_slope=negative_slope, mode=mode, nonlinearity=nonlinearity),
(out_channel, in_channel))
bias_init_0 = 'zeros'
if has_bias:
fan_in = _calculate_correct_fan((out_channel, in_channel), mode=mode)
scale = 1 / math.sqrt(fan_in)
bias_init_0 = initializer(Uniform(scale), [out_channel], type)
self.dense = nn.Dense(in_channel,
out_channel,
weight_init=kaiming_uniform_0,
bias_init=bias_init_0,
has_bias=True,
activation=activation).to_float(type)
def construct(self, x):
out = self.dense(x)
return out
@constexpr
def create_neg_inf(shape, dtype):
if dtype == mstype.float32:
return Tensor(np.full(shape, -3.4028235e+38, dtype=np.float32))
if dtype == mstype.float16:
return Tensor(np.full(shape, -65504, dtype=np.float16))
raise TypeError('dtype must be float16 or float32.')
class HubertModel(nn.Cell):
"""Definition of Hubert speech pretrained model.
Args:
enc_input_dim (int): Dimension of encoder input features.
enc_output_dim (int): Dimension of encoder output features.
n_classes (int): Number of kmeans clusters.
feature_extractor (nn.Cell): Feature extractor instance.
encoder (nn.Cell): Encoder instance.
n_classes (int): Number of pesudo label clusters.
final_dim (int): Output dimension of Hubert model.
logit_temp (float): scale factor for logits.
dropout_input (float): Dropout rate for dropping input.
dropout_features (float): Dropout rate for dropping features.
pred_masked_weight (float): Loss weight of masked prediction.
pred_unmasked_weight (float): Loss weight of unmasked prediction.
feature_pen_weight (float): Loss weight of feature penalty.
encoder_type (str): encoder type.
compute_type (int): Whether to use mix precision training.
"""
def __init__(self,
enc_input_dim,
enc_output_dim,
feature_extractor,
encoder,
n_classes,
final_dim=2048,
logit_temp=0.1,
dropout_input=0.1,
dropout_features=0.1,
pred_masked_weight=1.0,
pred_unmasked_weight=0.0,
feature_pen_weight=10,
encoder_type='conformer',
compute_type=mindspore.float16,
feature_grad_mult=0.1,
stride_list=None,
label_rate=None,
sample_rate=16000,
mask_conf=None):
super(HubertModel, self).__init__()
# definition of modules
self.feature_extractor = feature_extractor
self.layer_norm = LayerNorm((enc_input_dim,), epsilon=1e-5).to_float(mstype.float32)
self.post_extract_proj = (Dense(enc_input_dim, enc_output_dim).to_float(compute_type)
if enc_input_dim != enc_output_dim else None)
self.encoder = encoder
self.dropout_input = Dropout(1.0 - dropout_input)
self.dropout_feature = Dropout(1.0 - dropout_features)
self.squeeze = ops.Squeeze(1)
self.n_classes = n_classes
self.logit_temp = logit_temp
self.pred_masked_weight = pred_masked_weight
self.pred_unmasked_weight = pred_unmasked_weight
self.feature_pen_weight = feature_pen_weight
self.encoder_type = encoder_type
def forward_align(
self,
source: Tensor,
x_len: Tensor,
output_layer: Optional[int] = None,
padding_mask=None
) -> Tensor:
# np.save("/Work20/2023/wangtianrui/codes/projects/ProgRE/supplementary_results/ms/ms_conv_inp.npy", source.asnumpy())
features = self.feature_extractor(source, x_len)
# np.save("/Work20/2023/wangtianrui/codes/projects/ProgRE/supplementary_results/ms/ms_conv_out.npy", features.asnumpy())
features = features.transpose(0, 2, 1)
features = self.layer_norm(features)
if self.post_extract_proj is not None:
features = self.post_extract_proj(features)
# conv_out = features.copy()
# feature: (B, T, D), float
# target: (B, T), long
# x: (B, T, D), float
# padding_mask: (B, T), bool
# mask_indices: (B, T), bool
# np.save("/Work20/2023/wangtianrui/codes/projects/ProgRE/supplementary_results/ms/ms_encoder_inp.npy", features.asnumpy())
features, results = self.encoder(
x=features,
padding_mask=padding_mask,
output_layer=None if output_layer is None else output_layer - 1,
)
# results.insert(0, conv_out)
# np.save("/Work20/2023/wangtianrui/codes/projects/ProgRE/supplementary_results/ms/ms_encoder_out.npy", features.asnumpy())
return features, results
class HuBERTTraining(nn.Cell):
def __init__(self,
enc_input_dim,
enc_output_dim,
feature_extractor,
encoder,
n_classes,
final_dim=2048,
logit_temp=0.1,
dropout_input=0.1,
dropout_features=0.1,
pred_masked_weight=1.0,
pred_unmasked_weight=0.0,
feature_pen_weight=10,
encoder_type='transformer',
compute_type=mindspore.float16,
feature_grad_mult=0.1,
stride_list=None,
label_rate=None,
mask_conf=None):
super(HuBERTTraining, self).__init__()
self.acc_net = HubertModel(enc_input_dim=enc_input_dim,
enc_output_dim=enc_output_dim,
feature_extractor=feature_extractor,
encoder=encoder,
n_classes=n_classes,
final_dim=final_dim,
logit_temp=logit_temp,
dropout_input=dropout_input,
dropout_features=dropout_features,
pred_masked_weight=pred_masked_weight,
pred_unmasked_weight=pred_unmasked_weight,
feature_pen_weight=feature_pen_weight,
encoder_type=encoder_type,
compute_type=compute_type,
feature_grad_mult=feature_grad_mult,
stride_list=stride_list,
label_rate=label_rate,
mask_conf=mask_conf)
def construct(self, xs, xs_len, ys, padding_mask, hubert_mask, hubert_unmask):
loss = self.acc_net(xs, xs_len, ys, padding_mask, hubert_mask, hubert_unmask)
return loss
def init_hubert_model(config, input_dim):
"""Init a Hubert model."""
extractor_extra_args = config['extractor_conf']
hubert_extra_args = config['hubert_conf']
encoder_extra_args = config['encoder_conf']
label_rate = extractor_extra_args['label_rate']
mask_conf = config["collate_conf"]["mask_conf"]
feature_extractor = ConvFeatureExtractionModel(stride_list=extractor_extra_args['stride_list'],
kernel_size_list=extractor_extra_args['kernel_size_list'],
dim=extractor_extra_args['output_dim'],
dropout=extractor_extra_args['dropout'],
layer_norm=extractor_extra_args['layer_norm'],
has_bias=extractor_extra_args['has_bias'],
compute_type=mindspore.float32)
encoder = Wav2VecTransformerEncoder(encoder_embed_dim=encoder_extra_args['encoder_embed_dim'],
encoder_ffn_embed_dim=encoder_extra_args['encoder_ffn_embed_dim'],
encoder_num_attention_heads=encoder_extra_args['encoder_attention_heads'],
num_encoder_layers=encoder_extra_args['encoder_layers'],
conv_pos_kernel_size=encoder_extra_args['conv_pos'],
conv_pos_groups=encoder_extra_args['conv_pos_groups'],
conv_pos_use_same_pad=encoder_extra_args['conv_pos_use_same_pad'],
pos_conv_depth=encoder_extra_args['pos_conv_depth'],
dropout_rate=encoder_extra_args['dropout_rate'],
attention_dropout_rate=encoder_extra_args['attention_dropout_rate'],
activation_dropout_rate=encoder_extra_args['activation_dropout_rate'],
activation_fn=encoder_extra_args['activation_fn'],
layer_norm_first=encoder_extra_args['layer_norm_first'],
required_seq_len_multiple=encoder_extra_args['required_seq_len_multiple'],
encoder_layer_drop_rate=encoder_extra_args['encoder_layerdrop'],
compute_type=mindspore.float16)
model = HuBERTTraining(hubert_extra_args['input_dim'],
hubert_extra_args['output_dim'],
feature_extractor=feature_extractor,
encoder=encoder,
n_classes=hubert_extra_args['n_classes'],
final_dim=hubert_extra_args['final_dim'],
logit_temp=hubert_extra_args['logit_temp'],
dropout_input=hubert_extra_args['dropout_input'],
dropout_features=hubert_extra_args['dropout_features'],
pred_masked_weight=hubert_extra_args['pred_masked_weight'],
pred_unmasked_weight=hubert_extra_args['pred_unmasked_weight'],
feature_pen_weight=hubert_extra_args['feature_pen_weight'],
encoder_type=config['encoder'],
compute_type=mindspore.float16,
feature_grad_mult=extractor_extra_args['feature_grad_mult'],
stride_list=extractor_extra_args['stride_list'],
label_rate=label_rate,
mask_conf=mask_conf)
return model