Spaces:
Sleeping
Sleeping
unpairedelectron07
commited on
Commit
•
f2a28c0
1
Parent(s):
d61af5c
Upload losses.py
Browse files- audiocraft/adversarial/losses.py +228 -0
audiocraft/adversarial/losses.py
ADDED
@@ -0,0 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Utility module to handle adversarial losses without requiring to mess up the main training loop.
|
9 |
+
"""
|
10 |
+
|
11 |
+
import typing as tp
|
12 |
+
|
13 |
+
import flashy
|
14 |
+
import torch
|
15 |
+
import torch.nn as nn
|
16 |
+
import torch.nn.functional as F
|
17 |
+
|
18 |
+
|
19 |
+
ADVERSARIAL_LOSSES = ['mse', 'hinge', 'hinge2']
|
20 |
+
|
21 |
+
|
22 |
+
AdvLossType = tp.Union[nn.Module, tp.Callable[[torch.Tensor], torch.Tensor]]
|
23 |
+
FeatLossType = tp.Union[nn.Module, tp.Callable[[torch.Tensor, torch.Tensor], torch.Tensor]]
|
24 |
+
|
25 |
+
|
26 |
+
class AdversarialLoss(nn.Module):
|
27 |
+
"""Adversary training wrapper.
|
28 |
+
|
29 |
+
Args:
|
30 |
+
adversary (nn.Module): The adversary module will be used to estimate the logits given the fake and real samples.
|
31 |
+
We assume here the adversary output is ``Tuple[List[torch.Tensor], List[List[torch.Tensor]]]``
|
32 |
+
where the first item is a list of logits and the second item is a list of feature maps.
|
33 |
+
optimizer (torch.optim.Optimizer): Optimizer used for training the given module.
|
34 |
+
loss (AdvLossType): Loss function for generator training.
|
35 |
+
loss_real (AdvLossType): Loss function for adversarial training on logits from real samples.
|
36 |
+
loss_fake (AdvLossType): Loss function for adversarial training on logits from fake samples.
|
37 |
+
loss_feat (FeatLossType): Feature matching loss function for generator training.
|
38 |
+
normalize (bool): Whether to normalize by number of sub-discriminators.
|
39 |
+
|
40 |
+
Example of usage:
|
41 |
+
adv_loss = AdversarialLoss(adversaries, optimizer, loss, loss_real, loss_fake)
|
42 |
+
for real in loader:
|
43 |
+
noise = torch.randn(...)
|
44 |
+
fake = model(noise)
|
45 |
+
adv_loss.train_adv(fake, real)
|
46 |
+
loss, _ = adv_loss(fake, real)
|
47 |
+
loss.backward()
|
48 |
+
"""
|
49 |
+
def __init__(self,
|
50 |
+
adversary: nn.Module,
|
51 |
+
optimizer: torch.optim.Optimizer,
|
52 |
+
loss: AdvLossType,
|
53 |
+
loss_real: AdvLossType,
|
54 |
+
loss_fake: AdvLossType,
|
55 |
+
loss_feat: tp.Optional[FeatLossType] = None,
|
56 |
+
normalize: bool = True):
|
57 |
+
super().__init__()
|
58 |
+
self.adversary: nn.Module = adversary
|
59 |
+
flashy.distrib.broadcast_model(self.adversary)
|
60 |
+
self.optimizer = optimizer
|
61 |
+
self.loss = loss
|
62 |
+
self.loss_real = loss_real
|
63 |
+
self.loss_fake = loss_fake
|
64 |
+
self.loss_feat = loss_feat
|
65 |
+
self.normalize = normalize
|
66 |
+
|
67 |
+
def _save_to_state_dict(self, destination, prefix, keep_vars):
|
68 |
+
# Add the optimizer state dict inside our own.
|
69 |
+
super()._save_to_state_dict(destination, prefix, keep_vars)
|
70 |
+
destination[prefix + 'optimizer'] = self.optimizer.state_dict()
|
71 |
+
return destination
|
72 |
+
|
73 |
+
def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
|
74 |
+
# Load optimizer state.
|
75 |
+
self.optimizer.load_state_dict(state_dict.pop(prefix + 'optimizer'))
|
76 |
+
super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
|
77 |
+
|
78 |
+
def get_adversary_pred(self, x):
|
79 |
+
"""Run adversary model, validating expected output format."""
|
80 |
+
logits, fmaps = self.adversary(x)
|
81 |
+
assert isinstance(logits, list) and all([isinstance(t, torch.Tensor) for t in logits]), \
|
82 |
+
f'Expecting a list of tensors as logits but {type(logits)} found.'
|
83 |
+
assert isinstance(fmaps, list), f'Expecting a list of features maps but {type(fmaps)} found.'
|
84 |
+
for fmap in fmaps:
|
85 |
+
assert isinstance(fmap, list) and all([isinstance(f, torch.Tensor) for f in fmap]), \
|
86 |
+
f'Expecting a list of tensors as feature maps but {type(fmap)} found.'
|
87 |
+
return logits, fmaps
|
88 |
+
|
89 |
+
def train_adv(self, fake: torch.Tensor, real: torch.Tensor) -> torch.Tensor:
|
90 |
+
"""Train the adversary with the given fake and real example.
|
91 |
+
|
92 |
+
We assume the adversary output is the following format: Tuple[List[torch.Tensor], List[List[torch.Tensor]]].
|
93 |
+
The first item being the logits and second item being a list of feature maps for each sub-discriminator.
|
94 |
+
|
95 |
+
This will automatically synchronize gradients (with `flashy.distrib.eager_sync_model`)
|
96 |
+
and call the optimizer.
|
97 |
+
"""
|
98 |
+
loss = torch.tensor(0., device=fake.device)
|
99 |
+
all_logits_fake_is_fake, _ = self.get_adversary_pred(fake.detach())
|
100 |
+
all_logits_real_is_fake, _ = self.get_adversary_pred(real.detach())
|
101 |
+
n_sub_adversaries = len(all_logits_fake_is_fake)
|
102 |
+
for logit_fake_is_fake, logit_real_is_fake in zip(all_logits_fake_is_fake, all_logits_real_is_fake):
|
103 |
+
loss += self.loss_fake(logit_fake_is_fake) + self.loss_real(logit_real_is_fake)
|
104 |
+
|
105 |
+
if self.normalize:
|
106 |
+
loss /= n_sub_adversaries
|
107 |
+
|
108 |
+
self.optimizer.zero_grad()
|
109 |
+
with flashy.distrib.eager_sync_model(self.adversary):
|
110 |
+
loss.backward()
|
111 |
+
self.optimizer.step()
|
112 |
+
|
113 |
+
return loss
|
114 |
+
|
115 |
+
def forward(self, fake: torch.Tensor, real: torch.Tensor) -> tp.Tuple[torch.Tensor, torch.Tensor]:
|
116 |
+
"""Return the loss for the generator, i.e. trying to fool the adversary,
|
117 |
+
and feature matching loss if provided.
|
118 |
+
"""
|
119 |
+
adv = torch.tensor(0., device=fake.device)
|
120 |
+
feat = torch.tensor(0., device=fake.device)
|
121 |
+
with flashy.utils.readonly(self.adversary):
|
122 |
+
all_logits_fake_is_fake, all_fmap_fake = self.get_adversary_pred(fake)
|
123 |
+
all_logits_real_is_fake, all_fmap_real = self.get_adversary_pred(real)
|
124 |
+
n_sub_adversaries = len(all_logits_fake_is_fake)
|
125 |
+
for logit_fake_is_fake in all_logits_fake_is_fake:
|
126 |
+
adv += self.loss(logit_fake_is_fake)
|
127 |
+
if self.loss_feat:
|
128 |
+
for fmap_fake, fmap_real in zip(all_fmap_fake, all_fmap_real):
|
129 |
+
feat += self.loss_feat(fmap_fake, fmap_real)
|
130 |
+
|
131 |
+
if self.normalize:
|
132 |
+
adv /= n_sub_adversaries
|
133 |
+
feat /= n_sub_adversaries
|
134 |
+
|
135 |
+
return adv, feat
|
136 |
+
|
137 |
+
|
138 |
+
def get_adv_criterion(loss_type: str) -> tp.Callable:
|
139 |
+
assert loss_type in ADVERSARIAL_LOSSES
|
140 |
+
if loss_type == 'mse':
|
141 |
+
return mse_loss
|
142 |
+
elif loss_type == 'hinge':
|
143 |
+
return hinge_loss
|
144 |
+
elif loss_type == 'hinge2':
|
145 |
+
return hinge2_loss
|
146 |
+
raise ValueError('Unsupported loss')
|
147 |
+
|
148 |
+
|
149 |
+
def get_fake_criterion(loss_type: str) -> tp.Callable:
|
150 |
+
assert loss_type in ADVERSARIAL_LOSSES
|
151 |
+
if loss_type == 'mse':
|
152 |
+
return mse_fake_loss
|
153 |
+
elif loss_type in ['hinge', 'hinge2']:
|
154 |
+
return hinge_fake_loss
|
155 |
+
raise ValueError('Unsupported loss')
|
156 |
+
|
157 |
+
|
158 |
+
def get_real_criterion(loss_type: str) -> tp.Callable:
|
159 |
+
assert loss_type in ADVERSARIAL_LOSSES
|
160 |
+
if loss_type == 'mse':
|
161 |
+
return mse_real_loss
|
162 |
+
elif loss_type in ['hinge', 'hinge2']:
|
163 |
+
return hinge_real_loss
|
164 |
+
raise ValueError('Unsupported loss')
|
165 |
+
|
166 |
+
|
167 |
+
def mse_real_loss(x: torch.Tensor) -> torch.Tensor:
|
168 |
+
return F.mse_loss(x, torch.tensor(1., device=x.device).expand_as(x))
|
169 |
+
|
170 |
+
|
171 |
+
def mse_fake_loss(x: torch.Tensor) -> torch.Tensor:
|
172 |
+
return F.mse_loss(x, torch.tensor(0., device=x.device).expand_as(x))
|
173 |
+
|
174 |
+
|
175 |
+
def hinge_real_loss(x: torch.Tensor) -> torch.Tensor:
|
176 |
+
return -torch.mean(torch.min(x - 1, torch.tensor(0., device=x.device).expand_as(x)))
|
177 |
+
|
178 |
+
|
179 |
+
def hinge_fake_loss(x: torch.Tensor) -> torch.Tensor:
|
180 |
+
return -torch.mean(torch.min(-x - 1, torch.tensor(0., device=x.device).expand_as(x)))
|
181 |
+
|
182 |
+
|
183 |
+
def mse_loss(x: torch.Tensor) -> torch.Tensor:
|
184 |
+
if x.numel() == 0:
|
185 |
+
return torch.tensor([0.0], device=x.device)
|
186 |
+
return F.mse_loss(x, torch.tensor(1., device=x.device).expand_as(x))
|
187 |
+
|
188 |
+
|
189 |
+
def hinge_loss(x: torch.Tensor) -> torch.Tensor:
|
190 |
+
if x.numel() == 0:
|
191 |
+
return torch.tensor([0.0], device=x.device)
|
192 |
+
return -x.mean()
|
193 |
+
|
194 |
+
|
195 |
+
def hinge2_loss(x: torch.Tensor) -> torch.Tensor:
|
196 |
+
if x.numel() == 0:
|
197 |
+
return torch.tensor([0.0])
|
198 |
+
return -torch.mean(torch.min(x - 1, torch.tensor(0., device=x.device).expand_as(x)))
|
199 |
+
|
200 |
+
|
201 |
+
class FeatureMatchingLoss(nn.Module):
|
202 |
+
"""Feature matching loss for adversarial training.
|
203 |
+
|
204 |
+
Args:
|
205 |
+
loss (nn.Module): Loss to use for feature matching (default=torch.nn.L1).
|
206 |
+
normalize (bool): Whether to normalize the loss.
|
207 |
+
by number of feature maps.
|
208 |
+
"""
|
209 |
+
def __init__(self, loss: nn.Module = torch.nn.L1Loss(), normalize: bool = True):
|
210 |
+
super().__init__()
|
211 |
+
self.loss = loss
|
212 |
+
self.normalize = normalize
|
213 |
+
|
214 |
+
def forward(self, fmap_fake: tp.List[torch.Tensor], fmap_real: tp.List[torch.Tensor]) -> torch.Tensor:
|
215 |
+
assert len(fmap_fake) == len(fmap_real) and len(fmap_fake) > 0
|
216 |
+
feat_loss = torch.tensor(0., device=fmap_fake[0].device)
|
217 |
+
feat_scale = torch.tensor(0., device=fmap_fake[0].device)
|
218 |
+
n_fmaps = 0
|
219 |
+
for (feat_fake, feat_real) in zip(fmap_fake, fmap_real):
|
220 |
+
assert feat_fake.shape == feat_real.shape
|
221 |
+
n_fmaps += 1
|
222 |
+
feat_loss += self.loss(feat_fake, feat_real)
|
223 |
+
feat_scale += torch.mean(torch.abs(feat_real))
|
224 |
+
|
225 |
+
if self.normalize:
|
226 |
+
feat_loss /= n_fmaps
|
227 |
+
|
228 |
+
return feat_loss
|