-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_fns.py
148 lines (124 loc) · 6.18 KB
/
train_fns.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
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import utils
from utils import toggle_grad, ortho, save_weights
def create_train_fn(G, D, GD, z_, y_, ema, state_dict, config):
def train(x, y):
G.optim.zero_grad()
D.optim.zero_grad()
# How many chunks to split x and y into?
x = torch.split(x, config['batch_size'])
y = torch.split(y, config['batch_size'])
counter = 0
# Optionally toggle D and G's "require_grad"
if config['toggle_grads']:
toggle_grad(D, True)
toggle_grad(G, False)
for step_index in range(config['num_D_steps']):
# If accumulating gradients, loop multiple times before an optimizer step
D.optim.zero_grad()
for accumulation_index in range(config['num_D_accumulations']):
z_.sample_()
y_.sample_()
D_fake, D_real = GD(z_[:config['batch_size']], y_[:config['batch_size']],
x[counter], y[counter], train_G=False,
split_D=config['split_D'])
# Compute components of D's loss, average them, and divide by
# the number of gradient accumulations
D_loss_real, D_loss_fake = loss_hinge_dis(D_fake, D_real)
D_loss = (D_loss_real + D_loss_fake) / float(config['num_D_accumulations'])
D_loss.backward()
counter += 1
# Optionally apply ortho reg in D
if config['D_ortho'] > 0.0:
ortho(D, config['D_ortho'])
D.optim.step()
# Optionally toggle "requires_grad"
if config['toggle_grads']:
toggle_grad(D, False)
toggle_grad(G, True)
# Zero G's gradients by default before training G, for safety
G.optim.zero_grad()
# If accumulating gradients, loop multiple times
for accumulation_index in range(config['num_G_accumulations']):
z_.sample_()
y_.sample_()
D_fake = GD(z_, y_, train_G=True, split_D=config['split_D'])
G_loss = loss_hinge_gen(D_fake) / float(config['num_G_accumulations'])
G_loss.backward()
# Optionally apply modified ortho reg in G
if config['G_ortho'] > 0.0:
# Don't ortho reg shared, it makes no sense. Really we should blacklist any embeddings for this
ortho(G, config['G_ortho'], blacklist=[param for param in G.shared.parameters()])
G.optim.step()
# If we have an ema, update it, regardless of if we test with it or not
if config['ema']:
ema.update(state_dict['itr'])
out = {'G_loss': float(G_loss.item()),
'D_loss_real': float(D_loss_real.item()),
'D_loss_fake': float(D_loss_fake.item())}
# Return G's loss and the components of D's loss.
return out
return train
def loss_hinge_dis(dis_fake, dis_real):
loss_real = torch.mean(F.relu(1. - dis_real))
loss_fake = torch.mean(F.relu(1. + dis_fake))
return loss_real, loss_fake
def loss_hinge_gen(dis_fake):
loss = -torch.mean(dis_fake)
return loss
''' This function takes in the model, saves the weights (multiple copies if
requested), and prepares sample sheets: one consisting of samples given
a fixed noise seed (to show how the model evolves throughout training),
a set of full conditional sample sheets, and a set of interp sheets. '''
def save_and_sample(G, D, G_ema, z_, y_, fixed_z, fixed_y,
state_dict, config, experiment_name):
save_weights(G, D, state_dict, config['weights_root'],
experiment_name, None, G_ema if config['ema'] else None)
# Save an additional copy to mitigate accidental corruption if process
# is killed during a save (it's happened to me before -.-)
if config['num_save_copies'] > 0:
save_weights(G, D, state_dict, config['weights_root'],
experiment_name,
'copy%d' % state_dict['save_num'],
G_ema if config['ema'] else None)
state_dict['save_num'] = (state_dict['save_num'] + 1) % config['num_save_copies']
# Use EMA G for samples or non-EMA?
which_G = G_ema if config['ema'] and config['use_ema'] else G
# Save a random sample sheet with fixed z and y
with torch.no_grad():
if config['parallel']:
fixed_Gz = nn.parallel.data_parallel(which_G, (fixed_z, which_G.shared(fixed_y)))
else:
fixed_Gz = which_G(fixed_z, which_G.shared(fixed_y))
if not os.path.isdir('%s/%s' % (config['samples_root'], experiment_name)):
os.mkdir('%s/%s' % (config['samples_root'], experiment_name))
image_filename = '%s/%s/fixed_samples%d.jpg' % (config['samples_root'],
experiment_name,
state_dict['itr'])
torchvision.utils.save_image(fixed_Gz.float().cpu(), image_filename,
nrow=int(fixed_Gz.shape[0] ** 0.5), normalize=True)
# For now, every time we save, also save sample sheets
utils.sample_sheet(which_G,
classes_per_sheet=50,
num_classes=config['n_classes'],
samples_per_class=10, parallel=config['parallel'],
samples_root=config['samples_root'],
experiment_name=experiment_name,
folder_number=state_dict['itr'],
z_=z_)
# Also save interp sheets
for fix_z, fix_y in zip([False, False, True], [False, True, False]):
utils.interp_sheet(which_G,
num_per_sheet=16,
num_midpoints=8,
num_classes=config['n_classes'],
parallel=config['parallel'],
samples_root=config['samples_root'],
experiment_name=experiment_name,
folder_number=state_dict['itr'],
sheet_number=0,
fix_z=fix_z, fix_y=fix_y, device='cuda')