-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
270 lines (212 loc) · 11.6 KB
/
train.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Wed Nov 20 17:24:29 2019
@author: Samiul Arshad <[email protected]>
Training script for PCGAN.
Part of the code was taken from
-https://github.com/seowok/TreeGAN
"""
import torch
import torch.optim as optim
import json
import logging
import time
from datetime import datetime
import os
from shutil import rmtree
from torch.utils.tensorboard import SummaryWriter
from data.dataset_benchmark import BenchmarkDataset
from model.gan_network import Generator, Discriminator
from model.gradient_penalty import GradientPenalty
from arguments import Arguments
from utils import find_latest_epoch_and_step, setup_logging, writer_histogram, \
one_hot
# class : lable -> {'airplane': 0, 'chair': 1, 'motorcycle': 2, 'sofa': 3, 'table': 4}
class TreeGAN():
def __init__(self, args):
self.args = args
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# ------------------------------ Logger ---------------------------- #
logger = logging.getLogger()
if not len(logger.handlers):
setup_logging(self.args.result_root)
self.log = logging.getLogger(__name__)
self.log.debug('Using '+str(torch.cuda.device_count()) + ' GPUs!')
self.prepare_results_dir()
# Save args in file
if not os.path.exists(os.path.join(self.args.result_root, 'config.json')):
with open(os.path.join(self.args.result_root, 'config.json'), mode='w') as f:
json.dump(self.args.__dict__, f)
# Set device
self.args.device = torch.device('cuda:'+str(self.args.gpu) if torch.cuda.is_available() else 'cpu')
torch.cuda.set_device(self.args.device)
# ------------------------------ Dataset ---------------------------- #
self.data = BenchmarkDataset(root=args.dataset_path,
npoints=args.point_num,
class_choice=args.class_choice)
self.dataLoader = torch.utils.data.DataLoader(self.data,
batch_size=args.batch_size,
shuffle=True, pin_memory=True,
num_workers=16, drop_last=True)
self.log.debug(f'Training Dataset : {len(self.data)} prepared.')
# ------------------------------ Module ---------------------------- #
self.G = Generator(args).to(args.device)
self.G = torch.nn.DataParallel(self.G)
self.G = self.G.to(self.args.device)
self.optimizerG = optim.Adam(self.G.parameters(), lr=args.lr, betas=(0, 0.95))
self.D = Discriminator(args).to(args.device)
self.D = torch.nn.DataParallel(self.D)
self.D = self.D.to(self.args.device)
self.optimizerD = optim.Adam(self.D.parameters(), lr=args.lr, betas=(0, 0.95))
self.GP = GradientPenalty(args.lambdaGP, gamma=1, device=args.device)
self.log.debug(f"Network prepared.")
def prepare_results_dir(self):
# Clean previous files in case of retrain.
if self.args.retrain and os.path.isdir(self.args.result_root):
self.log.warning('Attention! Cleaning results directory in 10 seconds!')
time.sleep(10)
rmtree(self.args.result_root)
# Create appropriate result dirs.
os.makedirs(self.args.result_root, exist_ok=True)
os.makedirs(os.path.join(self.args.result_root,
'checkpoints'), exist_ok=True)
self.args.checkpoint_path = os.path.join(self.args.result_root,
'checkpoints')
os.makedirs(os.path.join(self.args.result_root,
'samples'), exist_ok=True)
self.args.samples_path = os.path.join(self.args.result_root, 'samples')
os.makedirs(os.path.join(self.args.result_root,
'runs'), exist_ok=True)
self.args.logs_path = os.path.join(self.args.result_root,
'runs', str(datetime.now()))
def save_results(self, step, epoch):
fake_pointclouds = torch.Tensor([])
for i in range(10): # For batch_size*10 samples
z = torch.randn(self.args.batch_size, 1, 64).to(self.args.device)
label = torch.randint(0, len(self.args.class_choice), (self.args.batch_size, 1))
label_oh = one_hot(label, len(self.args.class_choice))
with torch.no_grad():
sample = self.G([z, label_oh], step).cpu()
label = label.type(torch.FloatTensor).unsqueeze(1).repeat(1,sample.shape[1],1)
sample = torch.cat((sample, label), dim=2)
fake_pointclouds = torch.cat((fake_pointclouds, sample), dim=0)
path = os.path.join(self.args.samples_path, f'step_{step}_sample_{epoch:05}.pt')
torch.save(fake_pointclouds, path)
del fake_pointclouds
self.save_checkpoints(step, epoch)
def save_checkpoints(self, step, epoch):
# ---------------------- Save checkpoint --------------------- #
torch.save({
'D_state_dict': self.D.state_dict(),
'G_state_dict': self.G.state_dict(),
}, os.path.join(self.args.checkpoint_path,
f'step_{step}_epoch_{epoch:05}.pt'))
self.log.debug(f'Checkpoint is saved.')
def load_checkpoint(self, step, epoch):
# Expand model for step. No change needed for first step.
if step > 0:
self.change_params_for_step(current=0, target=step, epoch=epoch)
# Load checkpoints
path = os.path.join(self.args.checkpoint_path,
f'step_{step}_epoch_{epoch:05}.pt')
self.log.debug(f'Loading Checkpoint: {path}')
checkpoint = torch.load(path, map_location=self.args.device)
self.D.load_state_dict(checkpoint['D_state_dict'])
self.G.load_state_dict(checkpoint['G_state_dict'])
self.log.debug(f'Checkpoint loaded.')
def change_params_for_step(self, current, target, epoch=-1):
self.log.debug(f'Changing Params for step: {target} and epoch: {epoch}')
for i in range(current, target):
self.G.module.expand(i+1)
self.G = self.G.to(self.args.device)
self.optimizerG = optim.Adam(self.G.parameters(), lr=args.lr, betas=(0, 0.95))
self.D.module.k = 20 + target*10
self.D = self.D.to(self.args.device)
self.optimizerD = optim.Adam(self.D.parameters(), lr=args.lr, betas=(0, 0.95))
# increasing points number & decreasing batch size
self.data.npoints = self.args.point_num * pow(2,target)
self.args.batch_size = self.args.batch_size // pow(2,target)
del self.dataLoader
self.dataLoader = torch.utils.data.DataLoader(self.data,
batch_size=self.args.batch_size,
shuffle=True, pin_memory=False,
num_workers=16, drop_last=True)
self.log.debug(f'Current num points: {self.data.npoints} '
f'batch size: {self.args.batch_size}')
def run(self, save_ckpt=None, load_ckpt=None, result_path=None):
writer = SummaryWriter(self.args.logs_path)
writer.add_text('args', str(self.args), 0)
# load most current step and epoch. returns 0 for step, -1 for epoch if nothing was found.
starting_step, starting_epoch = find_latest_epoch_and_step(self.args.checkpoint_path)
if not self.args.retrain and starting_step+1 and starting_epoch+1:
self.load_checkpoint(starting_step, starting_epoch)
for step in range(starting_step, self.args.total_step):
for epoch in range(starting_epoch+1, self.args.epochs):
start_time = time.time()
G_loss, D_loss = 0., 0.
for _iter, data in enumerate(self.dataLoader):
point, y = data[0], data[1]
point = point.to(self.args.device)
y_onehot = torch.Tensor(one_hot(y,
len(self.args.class_choice))).to(self.args.device)
# ---------------------- Discriminator -------------------- #
for d_iter in range(self.args.D_iter):
self.D.zero_grad()
z = torch.randn(self.args.batch_size, 1, 64).\
to(self.args.device)
with torch.no_grad():
fake_point = self.G([z, y_onehot], step)
D_real = self.D([point, y_onehot])
D_realm = D_real.mean()
D_fake = self.D([fake_point, y_onehot])
D_fakem = D_fake.mean()
gp_loss = self.GP(self.D, point.data, fake_point.data, y_onehot)
d_loss = -D_realm + D_fakem
d_loss_gp = d_loss + gp_loss
d_loss_gp.backward()
self.optimizerD.step()
# ---------------------- Generator ---------------------- #
self.G.zero_grad()
z = torch.randn(self.args.batch_size, 1, 64).to(self.args.device)
fake_point = self.G([z, y_onehot], step)
G_fake = self.D([fake_point, y_onehot],)
G_fakem = G_fake.mean()
g_loss = -G_fakem
g_loss.backward()
self.optimizerG.step()
G_loss += g_loss.item()
D_loss += d_loss.item()
# --------------------- Visualization -------------------- #
self.log.debug(f' Step: {step: 3d}'
f' Epoch/Iter: {epoch:3d} / {_iter:3d} '
f' G_Loss: {g_loss: 7.5f} '
f' D_Loss: {d_loss: 7.6f} '
f' Time: {(time.time()-start_time):4.2f}s')
mean_D = D_loss/len(self.dataLoader)
mean_G = G_loss/len(self.dataLoader)
self.log.debug(f' Step: {step: 3d}'
f' Epoch: {epoch:3d} '
f' G_mean {mean_G:7.6f}'
f' D_mean: {mean_D:7.6f} '
f' Time: {(time.time()-start_time):4.2f}s')
# ------------------ Summery Writer ----------------- #
writer.add_scalar('Loss/D', mean_D, self.args.epochs*step+epoch)
writer.add_scalar('Loss/G', mean_G, self.args.epochs*step+epoch)
writer_histogram(writer, self.D, self.args.epochs*step+epoch)
writer_histogram(writer, self.G, self.args.epochs*step+epoch)
# ---------------- Save Generated Pointcloud --------------- #
if epoch > 0 and epoch % self.args.saving_freq == 0:
self.save_results(step, epoch)
if step+1 < self.args.total_step:
# update params for most current step
self.change_params_for_step(step, step+1)
starting_epoch = -1
writer.close()
del self.data, self.dataLoader
if __name__ == '__main__':
args = Arguments().parser().parse_args()
args.result_root = os.path.join(args.result_root, args.experiment)
model = TreeGAN(args)
model.run()