-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_boundary.py
201 lines (168 loc) · 7.22 KB
/
train_boundary.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
_# -*- coding:utf-8 -*-
import os
import sys
sys.path.append(os.path.join(os.path.dirname(__file__), 'networks'))
sys.path.append(os.path.join(os.path.dirname(__file__), 'datasets'))
import torch
from datasets.wflw import WFLW
from datasets.newWflw import newWflw
import torch.nn.functional as F
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
import torch.optim as optim
from tqdm import tqdm
from networks.net import BoundaryHeatmapEstimator, BoundaryHeatmapEstimatorwithMPL, \
DiscriL2
from networks.newNet import FAN
from lab_args import parse_args
from shutil import rmtree
import horovod.torch as hvd
from torch.backends import cudnn
from torch.utils.data.distributed import DistributedSampler
from lab_utils import lrSchedule, figPicture
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
# Ignore warnings
import warnings
def heatmapLoss(pred, target, loss_type):
loss = 0
if loss_type == 0:
# [B, 13, 64, 64] x 4
for i in range(len(pred) - 1):
loss += F.mse_loss(pred[i], target)
loss += 2 * F.mse_loss(pred[-1], target)
elif loss_type == 1:
for i in range(len(pred)):
loss += model_dirsc(target, pred[i])
return loss / (len(pred) + 1)
def save_checkpoint(path, epoch):
path, time = path.split(',')
state = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict()
}
# torch.save(net, path + 'net.pkl')
torch.save(state, path + f'/boundaries{time}.pt')
state = {
'model': model_dirsc.state_dict(),
'optimizer': optim_dirsc.state_dict()
}
# torch.save(net, 'net.pkl')
torch.save(state, path + f'/dirscLoss{time}.pt')
def adjustLr(iter, iters, per_epoch, epoch):
lr = lrSchedule(args.lr_base[0], iter, iters, epoch=epoch, time=args.lr_period, target_lr=args.lr_target[0],
mode='cycleCosine', per_epoch=per_epoch)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return lr
def main():
per_epoch = len(train_loader)
final_iters = per_epoch * args.lr_epoch
for epoch in range(args.epochs):
model.train()
if hvd.rank() == 0:
pbar = tqdm(train_loader)
else:
pbar = train_loader
for iter_idx, (data, heatmap, landmarks) in enumerate(pbar):
# Setup
global_iters = epoch * per_epoch + iter_idx
if epoch < args.lr_epoch:
lr = adjustLr(global_iters, final_iters, per_epoch, epoch)
else:
lr = args.lr_target[0]
lam = np.random.beta(args.mixup_alpha, args.mixup_alpha)
# Clean model grad, otherwise append
optimizer.zero_grad()
if args.loss_type == 1:
optim_dirsc.zero_grad()
# Put the data to GPUs
data, heatmap = data.cuda(), heatmap.cuda()
# Model forward
# mixup data mixup可以当作数据增强遮挡
if epoch < args.mixup_epoch:
# Can't use [::-1]
inv_idx = torch.arange(data.shape[0] - 1, -1, -1).long().cuda()
mixup_data = lam * data + (1 - lam) * data.index_select(0, inv_idx)
mixup_heatmap = lam * heatmap + (1 - lam) * heatmap.index_select(0, inv_idx)
else:
mixup_data = data
mixup_heatmap = heatmap
pred_heatmap = model(mixup_data)
loss_heatmap = heatmapLoss(pred_heatmap, mixup_heatmap, args.loss_type)
# loss_heatmap = model_dirsc(mixup_heatmap, pred_heatmap)
# Calc loss and Get the model grad (range from 0 to 1)
loss_heatmap.backward()
# Setup grad_scale
# torch.nn.utils.clip_grad_value_(model.parameters(), args.grad_clip)
# Update model
optimizer.step()
if args.loss_type == 1:
optim_dirsc.step()
# The 8 cards average output
loss_tb = hvd.allreduce(loss_heatmap, True, name='loss_heatmap')
# Others
if hvd.rank() == 0:
pic = figPicture(mixup_data[0].cpu().detach().numpy(), mixup_heatmap[0].cpu().detach().numpy(),
pred_heatmap[3][0].cpu().detach().numpy())
plt.imsave(f'/home/ubuntu/pic/par1/debug.png', pic)
pbar.set_description(f'Epoch {epoch} Loss: {loss_tb}')
writer.add_scalar('Net/loss_heatmap', loss_tb, global_iters)
writer.add_scalar('LR/lr', lr, global_iters)
if hvd.rank() == 0:
pbar.close()
if epoch % 4 is 3:
pic = figPicture(data[0].cpu().detach().numpy(), heatmap[0].cpu().detach().numpy(),
pred_heatmap[3][0].cpu().detach().numpy())
plt.imsave(f'/home/ubuntu/pic/par1/{epoch}.png', pic)
if epoch % (args.epochs // args.lr_period) == args.epochs // args.lr_period - 1 and not torch.isnan(
loss_heatmap):
path, time = args.save_params_path.split(',')
# save_checkpoint(args.save_params_path, epoch)
torch.save(model, path + f"/{time}-model-{epoch+1}.pkl")
if args.loss_type == 1:
torch.save(model_dirsc, path + f"/model_dirsc-{time}.pkl")
print('Saved...')
if hvd.rank() == 0:
# Verification per epoch
writer.close()
if __name__ == '__main__':
# Init horovod and torch.cuda
warnings.filterwarnings("ignore")
hvd.init()
torch.cuda.set_device(hvd.local_rank())
# Setup
args = parse_args()
# This flag allows you to enable the inbuilt cudnn
# auto-tuner to find the best algorithm to use for your hardware.
cudnn.benchmark = True
if hvd.rank() == 0:
# Announce
print(args)
# Init tensorboard
rmtree(args.tensorboard_path, ignore_errors=True)
writer = SummaryWriter(args.tensorboard_path)
# DataLoader
train_dataset = newWflw(mode='train')
train_sampler = DistributedSampler(dataset=train_dataset, num_replicas=hvd.size(), rank=hvd.rank())
train_loader = DataLoader(dataset=train_dataset, batch_size=args.per_batch, sampler=train_sampler)
# Model
Estimator = BoundaryHeatmapEstimatorwithMPL if args.mpl else BoundaryHeatmapEstimator
model = FAN(4).cuda()
# Load pretrained Model
if hvd.rank() == 0:
model = torch.load('/home/ubuntu/param/0-model-200.pkl').cuda()
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
# Optimizer
# optimizer = optim.SGD(model.parameters(), lr=args.lr_base[0], momentum=0.9, weight_decay=args.wd)
optimizer = optim.Adam(model.parameters(), lr=args.lr_base[0], betas=args.beta, weight_decay=args.wd)
optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters())
if args.loss_type == 1:
model_dirsc = DiscriL2(args.boundary).cuda()
optim_dirsc = optim.SGD(model_dirsc.parameters(), lr=args.lr_target[2], momentum=0.9, weight_decay=args.wd)
optim_dirsc = hvd.DistributedOptimizer(optim_dirsc, model_dirsc.named_parameters())
# Main function
if hvd.rank() == 0:
print('Training......')
main()