forked from Hippogriff/3DCSGNet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
208 lines (176 loc) · 8.78 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
"""
Train the network using mixture of programs.
"""
import sys
import numpy as np
import torch
import torch.optim as optim
from tensorboard_logger import configure, log_value
from torch.autograd.variable import Variable
from src.Utils import read_config
from src.Generator.generator import Generator
from src.Models.loss import losses_joint
from src.Models.models import CsgNet, ParseModelOutput
from src.Utils.learn_utils import LearningRate
from src.Utils.train_utils import prepare_input_op, Callbacks
if len(sys.argv) > 1:
config = read_config.Config(sys.argv[1])
else:
config = read_config.Config("config.yml")
model_name = config.model_path.format(config.proportion,
config.top_k,
config.hidden_size,
config.batch_size,
config.optim, config.lr,
config.weight_decay,
config.dropout,
"mix",
config.mode)
print(config.config)
config.write_config("log/configs/{}_config.json".format(model_name))
configure("log/tensorboard/{}".format(model_name), flush_secs=5)
callback = Callbacks(config.batch_size, "log/db/{}".format(model_name))
callback.add_element(["train_loss", "test_loss", "train_mse", "test_mse"])
data_labels_paths = {3: "data/one_op/expressions.txt",
5: "data/two_ops/expressions.txt",
7: "data/three_ops/expressions.txt"}
proportion = config.proportion # proportion is in percentage. vary from [1, 100].
# First is training size and second is validation size per program length
dataset_sizes = {3: [proportion * 1000, proportion * 250],
5: [proportion * 2000, proportion * 500],
7: [proportion * 4000, proportion * 100]}
config.train_size = sum(dataset_sizes[k][0] for k in dataset_sizes.keys())
config.test_size = sum(dataset_sizes[k][1] for k in dataset_sizes.keys())
types_prog = len(dataset_sizes)
generator = Generator(data_labels_paths=data_labels_paths,
batch_size=config.batch_size,
time_steps=max(data_labels_paths.keys()),
stack_size=max(data_labels_paths.keys()) // 2 + 1)
imitate_net = CsgNet(grid_shape=[64, 64, 64], dropout=config.dropout,
mode=config.mode, timesteps=max(data_labels_paths.keys()),
num_draws=len(generator.unique_draw),
in_sz=config.input_size,
hd_sz=config.hidden_size,
stack_len=config.top_k)
# If you want to use multiple GPUs for training.
cuda_devices = torch.cuda.device_count()
if torch.cuda.device_count() > 1:
imitate_net.cuda_devices = torch.cuda.device_count()
print("using multi gpus", flush=True)
imitate_net = torch.nn.DataParallel(imitate_net, device_ids=[0, 1], dim=0)
imitate_net.cuda()
if config.preload_model:
imitate_net.load_state_dict(torch.load(config.pretrain_modelpath))
for param in imitate_net.parameters():
param.requires_grad = True
if config.optim == "sgd":
optimizer = optim.SGD(
[para for para in imitate_net.parameters() if para.requires_grad],
weight_decay=config.weight_decay,
momentum=0.9, lr=config.lr, nesterov=False)
elif config.optim == "adam":
optimizer = optim.Adam(
[para for para in imitate_net.parameters() if para.requires_grad],
weight_decay=config.weight_decay, lr=config.lr)
reduce_plat = LearningRate(optimizer, init_lr=config.lr, lr_dacay_fact=0.2,
lr_decay_epoch=3, patience=config.patience)
train_gen_objs = {}
test_gen_objs = {}
# Prefetching minibatches
for k in data_labels_paths.keys():
# if using multi gpu training, train and test batch size should be multiple of
# number of GPU edvices.
train_batch_size = config.batch_size // types_prog
test_batch_size = config.batch_size // types_prog
train_gen_objs[k] = generator.get_train_data(train_batch_size,
k,
num_train_images=dataset_sizes[k][0],
if_primitives=True,
if_jitter=False)
test_gen_objs[k] = generator.get_test_data(test_batch_size,
k,
num_train_images=dataset_sizes[k][0],
num_test_images=dataset_sizes[k][1],
if_primitives=True,
if_jitter=False)
prev_test_loss = 1e20
prev_test_reward = 0
for epoch in range(0, config.epochs):
train_loss = 0
Accuracies = []
imitate_net.train()
# Number of times to accumulate gradients
num_accums = config.num_traj
for batch_idx in range(config.train_size // (config.batch_size * config.num_traj)):
optimizer.zero_grad()
loss_sum = Variable(torch.zeros(1)).cuda().data
for _ in range(num_accums):
for k in data_labels_paths.keys():
data, labels = next(train_gen_objs[k])
data = data[:, :, 0:config.top_k + 1, :, :, :]
one_hot_labels = prepare_input_op(labels, len(generator.unique_draw))
one_hot_labels = Variable(torch.from_numpy(one_hot_labels)).cuda()
data = Variable(torch.from_numpy(data)).cuda()
labels = Variable(torch.from_numpy(labels)).cuda()
data = data.permute(1, 0, 2, 3, 4, 5)
# forward pass
outputs = imitate_net([data, one_hot_labels, k])
loss = losses_joint(outputs, labels, time_steps=k + 1) / types_prog / \
num_accums
loss.backward()
loss_sum += loss.data
# Clip the gradient to fixed value to stabilize training.
torch.nn.utils.clip_grad_norm(imitate_net.parameters(), 20)
optimizer.step()
l = loss_sum
train_loss += l
log_value('train_loss_batch', l.cpu().numpy(), epoch * (
config.train_size //
(config.batch_size * num_accums)) + batch_idx)
mean_train_loss = train_loss / (config.train_size // (config.batch_size * num_accums))
log_value('train_loss', mean_train_loss.cpu().numpy(), epoch)
del data, loss, loss_sum, train_loss, outputs
test_losses = 0
imitate_net.eval()
test_reward = 0
for batch_idx in range(config.test_size // config.batch_size):
for k in data_labels_paths.keys():
parser = ParseModelOutput(generator.unique_draw,
stack_size=(k + 1) // 2 + 1,
steps=k,
canvas_shape=[64, 64, 64],
primitives=generator.primitives)
data_, labels = next(test_gen_objs[k])
one_hot_labels = prepare_input_op(labels, len(generator.unique_draw))
one_hot_labels = Variable(torch.from_numpy(one_hot_labels)).cuda()
data = Variable(torch.from_numpy(data_[:, :, 0:config.top_k + 1, :, :, :]), volatile=True).cuda()
data = data.permute(1, 0, 2, 3, 4, 5)
labels = Variable(torch.from_numpy(labels)).cuda()
test_output = imitate_net([data, one_hot_labels, k])
l = losses_joint(test_output, labels, time_steps=k + 1).data
test_losses += l
if cuda_devices > 1:
test_output = imitate_net.module.test([data, one_hot_labels, k])
else:
test_output = imitate_net.test([data, one_hot_labels, k])
stack, _, _ = parser.get_final_canvas(test_output, if_pred_images=True,
if_just_expressions=False)
data_ = data_[-1, :, 0, :, :, :]
R = np.sum(np.logical_and(stack, data_), (1, 2, 3)) / (
np.sum(np.logical_or(stack, data_), (1, 2, 3)) + 1)
test_reward += np.mean(R)
test_reward = test_reward / types_prog
test_loss = test_losses.cpu().numpy() / (config.test_size // config.batch_size) / types_prog
log_value('test_loss', test_loss, epoch)
log_value('test_IOU', test_reward / (config.test_size // config.batch_size), epoch)
callback.add_value({
"test_loss": test_loss,
})
if config.if_schedule:
reduce_plat.reduce_on_plateu(-test_reward)
del test_losses, test_output
if test_reward > prev_test_reward:
torch.save(imitate_net.state_dict(),
"trained_models/{}.pth".format(model_name))
prev_test_reward = test_reward
callback.dump_all()