forked from NVIDIA/DeepLearningExamples
-
Notifications
You must be signed in to change notification settings - Fork 0
/
hubconf.py
220 lines (181 loc) · 9.04 KB
/
hubconf.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
import urllib.request
import torch
import os
import sys
# from https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/SpeechSynthesis/Tacotron2/inference.py
def checkpoint_from_distributed(state_dict):
"""
Checks whether checkpoint was generated by DistributedDataParallel. DDP
wraps model in additional "module.", it needs to be unwrapped for single
GPU inference.
:param state_dict: model's state dict
"""
ret = False
for key, _ in state_dict.items():
if key.find('module.') != -1:
ret = True
break
return ret
# from https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/SpeechSynthesis/Tacotron2/inference.py
def unwrap_distributed(state_dict):
"""
Unwraps model from DistributedDataParallel.
DDP wraps model in additional "module.", it needs to be removed for single
GPU inference.
:param state_dict: model's state dict
"""
new_state_dict = {}
for key, value in state_dict.items():
new_key = key.replace('module.1.', '')
new_key = new_key.replace('module.', '')
new_state_dict[new_key] = value
return new_state_dict
dependencies = ['torch']
def nvidia_ncf(pretrained=True, **kwargs):
"""Constructs an NCF model.
For detailed information on model input and output, training recipies, inference and performance
visit: github.com/NVIDIA/DeepLearningExamples and/or ngc.nvidia.com
Args:
pretrained (bool, True): If True, returns a model pretrained on ml-20m dataset.
model_math (str, 'fp32'): returns a model in given precision ('fp32' or 'fp16')
nb_users (int): number of users
nb_items (int): number of items
mf_dim (int, 64): dimension of latent space in matrix factorization
mlp_layer_sizes (list, [256,256,128,64]): sizes of layers of multi-layer-perceptron
dropout (float, 0.5): dropout
"""
from PyTorch.Recommendation.NCF import neumf as ncf
fp16 = "model_math" in kwargs and kwargs["model_math"] == "fp16"
force_reload = "force_reload" in kwargs and kwargs["force_reload"]
config = {'nb_users': None, 'nb_items': None, 'mf_dim': 64, 'mf_reg': 0.,
'mlp_layer_sizes': [256, 256, 128, 64], 'mlp_layer_regs':[0, 0, 0, 0], 'dropout': 0.5}
if pretrained:
if fp16:
checkpoint = 'https://developer.nvidia.com/joc-ncf-fp16-pyt-20190225'
else:
checkpoint = 'https://developer.nvidia.com/joc-ncf-fp32-pyt-20190225'
ckpt_file = os.path.basename(checkpoint)
if not os.path.exists(ckpt_file) or force_reload:
sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
urllib.request.urlretrieve(checkpoint, ckpt_file)
ckpt = torch.load(ckpt_file)
if checkpoint_from_distributed(ckpt):
ckpt = unwrap_distributed(ckpt)
config['nb_users'] = ckpt['mf_user_embed.weight'].shape[0]
config['nb_items'] = ckpt['mf_item_embed.weight'].shape[0]
config['mf_dim'] = ckpt['mf_item_embed.weight'].shape[1]
mlp_shapes = [ckpt[k].shape for k in ckpt.keys() if 'mlp' in k and 'weight' in k and 'embed' not in k]
config['mlp_layer_sizes'] = [mlp_shapes[0][1], mlp_shapes[1][1], mlp_shapes[2][1], mlp_shapes[2][0]]
config['mlp_layer_regs'] = [0] * len(config['mlp_layer_sizes'])
else:
if 'nb_users' not in kwargs:
raise ValueError("Missing 'nb_users' argument.")
if 'nb_items' not in kwargs:
raise ValueError("Missing 'nb_items' argument.")
for k,v in kwargs.items():
if k in config.keys():
config[k] = v
config['mlp_layer_regs'] = [0] * len(config['mlp_layer_sizes'])
m = ncf.NeuMF(**config)
if fp16:
m.half()
if pretrained:
m.load_state_dict(ckpt)
return m
def nvidia_tacotron2(pretrained=True, **kwargs):
"""Constructs a Tacotron 2 model (nn.module with additional infer(input) method).
For detailed information on model input and output, training recipies, inference and performance
visit: github.com/NVIDIA/DeepLearningExamples and/or ngc.nvidia.com
Args (type[, default value]):
pretrained (bool, True): If True, returns a model pretrained on LJ Speech dataset.
model_math (str, 'fp32'): returns a model in given precision ('fp32' or 'fp16')
n_symbols (int, 148): Number of symbols used in a sequence passed to the prenet, see
https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/SpeechSynthesis/Tacotron2/tacotron2/text/symbols.py
p_attention_dropout (float, 0.1): dropout probability on attention LSTM (1st LSTM layer in decoder)
p_decoder_dropout (float, 0.1): dropout probability on decoder LSTM (2nd LSTM layer in decoder)
max_decoder_steps (int, 1000): maximum number of generated mel spectrograms during inference
"""
from PyTorch.SpeechSynthesis.Tacotron2.tacotron2 import model as tacotron2
from PyTorch.SpeechSynthesis.Tacotron2.models import lstmcell_to_float, batchnorm_to_float
fp16 = "model_math" in kwargs and kwargs["model_math"] == "fp16"
force_reload = "force_reload" in kwargs and kwargs["force_reload"]
if pretrained:
if fp16:
checkpoint = 'https://developer.nvidia.com/joc-tacotron2-fp16-pyt-20190306'
else:
checkpoint = 'https://developer.nvidia.com/joc-tacotron2-fp32-pyt-20190306'
ckpt_file = os.path.basename(checkpoint)
if not os.path.exists(ckpt_file) or force_reload:
sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
urllib.request.urlretrieve(checkpoint, ckpt_file)
ckpt = torch.load(ckpt_file)
state_dict = ckpt['state_dict']
if checkpoint_from_distributed(state_dict):
state_dict = unwrap_distributed(state_dict)
config = ckpt['config']
else:
config = {'mask_padding': False, 'n_mel_channels': 80, 'n_symbols': 148,
'symbols_embedding_dim': 512, 'encoder_kernel_size': 5,
'encoder_n_convolutions': 3, 'encoder_embedding_dim': 512,
'attention_rnn_dim': 1024, 'attention_dim': 128,
'attention_location_n_filters': 32,
'attention_location_kernel_size': 31, 'n_frames_per_step': 1,
'decoder_rnn_dim': 1024, 'prenet_dim': 256,
'max_decoder_steps': 1000, 'gate_threshold': 0.5,
'p_attention_dropout': 0.1, 'p_decoder_dropout': 0.1,
'postnet_embedding_dim': 512, 'postnet_kernel_size': 5,
'postnet_n_convolutions': 5, 'decoder_no_early_stopping': False}
for k,v in kwargs.items():
if k in config.keys():
config[k] = v
m = tacotron2.Tacotron2(**config)
if fp16:
m = batchnorm_to_float(m.half())
m = lstmcell_to_float(m)
if pretrained:
m.load_state_dict(state_dict)
return m
def nvidia_waveglow(pretrained=True, **kwargs):
"""Constructs a WaveGlow model (nn.module with additional infer(input) method).
For detailed information on model input and output, training recipies, inference and performance
visit: github.com/NVIDIA/DeepLearningExamples and/or ngc.nvidia.com
Args:
pretrained (bool): If True, returns a model pretrained on LJ Speech dataset.
model_math (str, 'fp32'): returns a model in given precision ('fp32' or 'fp16')
"""
from PyTorch.SpeechSynthesis.Tacotron2.waveglow import model as waveglow
from PyTorch.SpeechSynthesis.Tacotron2.models import batchnorm_to_float
fp16 = "model_math" in kwargs and kwargs["model_math"] == "fp16"
force_reload = "force_reload" in kwargs and kwargs["force_reload"]
if pretrained:
if fp16:
checkpoint = 'https://developer.nvidia.com/joc-waveglow-fp16-pyt-20190306'
else:
checkpoint = 'https://developer.nvidia.com/joc-waveglow-fp32-pyt-20190306'
ckpt_file = os.path.basename(checkpoint)
if not os.path.exists(ckpt_file) or force_reload:
sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
urllib.request.urlretrieve(checkpoint, ckpt_file)
ckpt = torch.load(ckpt_file)
state_dict = ckpt['state_dict']
if checkpoint_from_distributed(state_dict):
state_dict = unwrap_distributed(state_dict)
config = ckpt['config']
else:
config = {'n_mel_channels': 80, 'n_flows': 12, 'n_group': 8,
'n_early_every': 4, 'n_early_size': 2,
'WN_config': {'n_layers': 8, 'kernel_size': 3,
'n_channels': 512}}
for k,v in kwargs.items():
if k in config.keys():
config[k] = v
elif k in config['WN_config'].keys():
config['WN_config'][k] = v
m = waveglow.WaveGlow(**config)
if fp16:
m = batchnorm_to_float(m.half())
for mat in m.convinv:
mat.float()
if pretrained:
m.load_state_dict(state_dict)
return m