-
Notifications
You must be signed in to change notification settings - Fork 81
/
distiller.py
286 lines (251 loc) · 10.9 KB
/
distiller.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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
import os
# random
import random
# pytorch
import torch
import torch.nn as nn
import pytorch_lightning as pl
# dataset
from core.dataset import NoiseDataset
# teacher model
from core.model_zoo import model_zoo
from core.models.mapping_network import MappingNetwork
from core.models.synthesis_network import SynthesisNetwork
# student model
from core.models.mobile_synthesis_network import MobileSynthesisNetwork
# loss
from core.loss.distiller_loss import DistillerLoss
# evaluation network
from core.models.inception_v3 import load_inception_v3
# evaluation metric
from piq import KID
# utils
from core.utils import apply_trace_model_mode
class Distiller(pl.LightningModule):
def __init__(self, cfg, **kwargs):
super().__init__()
self.cfg = cfg.trainer
# teacher model
print("load mapping network...")
mapping_net_ckpt = model_zoo(**cfg.teacher.mapping_network)
self.mapping_net = MappingNetwork(**mapping_net_ckpt["params"]).eval()
self.mapping_net.load_state_dict(mapping_net_ckpt["ckpt"])
print("load synthesis network...")
synthesis_net_ckpt = model_zoo(**cfg.teacher.synthesis_network)
self.synthesis_net = SynthesisNetwork(**synthesis_net_ckpt["params"]).eval()
self.synthesis_net.load_state_dict(synthesis_net_ckpt["ckpt"])
# student network
self.student = MobileSynthesisNetwork(
style_dim=self.mapping_net.style_dim,
channels=synthesis_net_ckpt["params"]["channels"][:-1]
)
# dataset
self.wsize = self.student.wsize()
self.trainset = NoiseDataset(batch_size=self.cfg.batch_size, **cfg.trainset)
self.valset = NoiseDataset(batch_size=self.cfg.batch_size, **cfg.valset)
#compute style_mean
self.register_buffer(
"style_mean",
self.compute_mean_style(self.mapping_net.style_dim, wsize=self.wsize, batch_size=4096)
)
# loss
self.loss = DistillerLoss(
discriminator_size=self.synthesis_net.size,
**cfg.distillation_loss
)
# evaluator
self.kid = KID()
self.inception = load_inception_v3()
# device info
self.register_buffer("device_info", torch.tensor(1))
def _log_loss(self, loss, on_step=True, on_epoch=False, prog_bar=True, logger=True, exclude=["loss"]):
for k, v in loss.items():
if not k in exclude:
self.log(k, v, on_step=on_step, on_epoch=on_epoch, prog_bar=prog_bar, logger=logger)
def training_step(self, batch, batch_nb, optimizer_idx=0):
mode = self.opt_to_mode[optimizer_idx]
if mode == "g":
loss = self.generator_step(batch, batch_nb)
self._log_loss(loss)
elif mode == "d":
loss = self.discriminator_step(batch, batch_nb)
self._log_loss(loss)
return {"loss": loss["loss"]}
def validation_step(self, batch, batch_nb):
# compute inception_v3 features
style, gt = self.make_sample(batch)
pred = self.student(style, noise=gt["noise"])
pred_inc = self.inception(pred["img"])[0].view(style.size(0), -1)
gt_inc = self.inception(gt["img"])[0].view(style.size(0), -1)
# compute val_loss
pred = self.student(style, noise=gt["noise"])
loss = self.loss.loss_g(pred, gt)
return {"pred": pred_inc, "gt": gt_inc, "loss_val": loss["loss"]}
def validation_epoch_end(self, outputs):
# TODO: add all_gather for distributed mode
# agregate kid_val
pred, gt = [], []
for x in outputs:
pred.append(x["pred"])
gt.append(x["gt"])
pred = torch.cat(pred, axis=0)
gt = torch.cat(gt, axis=0)
kid = self.kid.compute_metric(pred, gt)
self.log("kid_val", kid, prog_bar=True)
# agregate val_loss
loss = torch.stack([x['loss_val'] for x in outputs]).mean()
self.log("loss_val", loss, prog_bar=True)
def generator_step(self, batch, batch_nb):
style, gt = self.make_sample(batch)
pred = self.student(style, noise=gt["noise"])
loss = self.loss.loss_g(pred, gt)
return loss
def discriminator_step(self, batch, batch_nb):
style, gt = self.make_sample(batch)
with torch.no_grad():
pred = self.student(style, noise=gt["noise"])
if self.global_step % self.cfg.reg_d_interval != 0:
loss = self.loss.loss_d(pred, gt)
else:
loss = self.loss.reg_d(gt)
loss["loss"] *= self.cfg.reg_d_interval
return loss
@torch.no_grad()
def make_sample(self, batch):
def make_style():
var = torch.randn(self.cfg.batch_size, self.mapping_net.style_dim).to(self.device_info.device)
style = self.mapping_net(var)
return style
coin = random.random()
if coin >= self.cfg.stylemix_p[1]:
style = self.mapping_net(batch["noise"]).unsqueeze(1).repeat(1, self.wsize, 1)
elif coin >= self.cfg.stylemix_p[0] and coin < self.cfg.stylemix_p[1]:
style_a, style_b = make_style(), make_style()
inject_index = random.randint(1, self.wsize - 1)
style_a = style_a.unsqueeze(1).repeat(1, inject_index, 1)
style_b = style_b.unsqueeze(1).repeat(1, self.wsize - inject_index, 1)
style = torch.cat([style_a, style_b], dim=1)
else:
var = torch.randn(self.wsize, self.mapping_net.style_dim).to(self.device_info.device)
style = self.mapping_net(var).view(1, self.wsize, self.mapping_net.style_dim)
if self.cfg.truncated:
style = self.style_mean + 0.5 * (style - self.style_mean)
gt = self.synthesis_net(style)
return style, gt
@torch.no_grad()
def compute_mean_style(self, style_dim, wsize=1, batch_size=4096):
style = self.mapping_net(torch.randn(4096, self.mapping_net.style_dim)).mean(0, keepdim=True)
if wsize != 1:
style = style.unsqueeze(1).repeat(1, wsize, 1)
return style
def train_dataloader(self):
return torch.utils.data.DataLoader(self.trainset, batch_size=self.trainset.batch_size, num_workers=self.cfg.num_workers, shuffle=False)
def val_dataloader(self):
return torch.utils.data.DataLoader(self.valset, batch_size=self.valset.batch_size, num_workers=self.cfg.num_workers, shuffle=False)
def configure_optimizers(self):
opts = []
self.opt_to_mode = {}
mode = self.cfg.mode.split(',')
for i, mode in enumerate(self.cfg.mode.split(',')):
if mode == "g":
print("setup generator train mode on...")
opts.append(torch.optim.Adam(self.student.parameters(), lr=self.cfg.lr_student))
self.opt_to_mode[i] = "g"
elif mode == "d":
print("setup discriminator train mode on...")
opts.append(torch.optim.Adam(self.loss.gan_loss.parameters(), lr=self.cfg.lr_gan))
self.opt_to_mode[i] = "d"
return opts, []
def forward(self, var, truncated=False, generator="student"):
var = var.to(self.device_info.device)
style = self.mapping_net(var)
if truncated:
style = self.style_mean + 0.5 * (style - self.style_mean)
if generator == "student":
img = self.student(style)["img"]
else:
img = self.synthesis_net(style)["img"]
return img
def simultaneous_forward(self, var, truncated=False):
var = var.to(self.device_info.device)
style = self.mapping_net(var)
if truncated:
style = self.style_mean + 0.5 * (style - self.style_mean)
out_t = self.synthesis_net(style)
img_s = self.student(style, noise=out_t["noise"])["img"]
return img_s, out_t["img"]
def to_onnx(self, output_dir, w_plus=False):
class Wrapper(nn.Module):
def __init__(
self,
synthesis_network,
style_tmp
):
super().__init__()
self.m = synthesis_network
self.noise = self.m(style_tmp)["noise"]
def forward(self, style):
return self.m(style, noise=self.noise)["img"]
print("prepare style...")
if not w_plus:
var = torch.randn(1, self.mapping_net.style_dim).to(self.device_info.device)
style = self.mapping_net(var)
else:
var = torch.randn(self.wsize, self.mapping_net.style_dim).to(self.device_info.device)
style = self.mapping_net(var)
style = style.view(1, self.wsize, -1)
print("convert mapping network...")
self.mapping_net.apply(apply_trace_model_mode(True))
torch.onnx.export(
self.mapping_net,
(var,),
os.path.join(output_dir, "MappingNetwork.onnx"),
input_names = ['var'],
output_names = ['style'],
verbose=True
)
print("convert synthesis network...")
self.student.apply(apply_trace_model_mode(True))
torch.onnx.export(
Wrapper(self.student, style),
(style,),
os.path.join(output_dir, "SynthesisNetwork.onnx"),
input_names = ['style'],
output_names = ['img'],
verbose=True
)
def to_coreml(self, output_dir, w_plus=False):
import coremltools as ct
print("prepare style...")
if not w_plus:
var = torch.randn(1, self.mapping_net.style_dim).to(self.device_info.device)
style = self.mapping_net(var)
else:
var = torch.randn(self.wsize, self.mapping_net.style_dim).to(self.device_info.device)
style = self.mapping_net(var)
style = style.view(1, self.wsize, -1)
print("convert mapping network...")
self.mapping_net.apply(apply_trace_model_mode(True))
mapping_net_trace = torch.jit.trace(self.mapping_net, var)
mapping_net_coreml = ct.convert(
mapping_net_trace,
inputs=[ct.TensorType(name="var", shape=var.shape)]
)
mapping_net_coreml.save(os.path.join(output_dir, "MappingNetwork.mlmodel"))
print("convert synthesis network...")
self.student.apply(apply_trace_model_mode(True))
# initialize noise buffers
self.student(style)
class Wrapper(nn.Module):
def __init__(self, m):
super().__init__()
self.m = m
def forward(self, style):
return self.m(style)["img"]
synthesis_net = Wrapper(self.student)
synthesis_net_trace = torch.jit.trace(synthesis_net, style)
synthesis_net_coreml = ct.convert(
synthesis_net_trace,
inputs=[ct.TensorType(name="style", shape=style.shape)]
)
synthesis_net_coreml.save(os.path.join(output_dir, "SynthesisNetwork.mlmodel"))