-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
200 lines (160 loc) · 6.96 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
import argparse
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
from torch.utils.tensorboard import SummaryWriter
import os
import random
import math
import numpy as np
from utils.loss import CombinedLoss
from utils.config import load_config
from utils.draw import draw_overlay
from utils.schedulers import CosineAnnealingLR, LinearLR
from torch.optim.lr_scheduler import LambdaLR#CosineAnnealingLR
import utils.augmentations as A
from datasets.loaders import get_loader
from models.model import Model
random.seed(0)
torch.manual_seed(0)
np.random.seed(0)
parser = argparse.ArgumentParser()
parser.add_argument('-cfg', '--config', type=str, help="Path to training config", required=True)
args = parser.parse_args()
np.testing.suppress_warnings()
cfg = load_config(args.config)
batch_size = cfg.batch_size
learning_rate = cfg.learning_rate
epochs=cfg.epochs
transforms = A.Compose([
A.RandomGaussianBlur(cfg.blur_p, cfg.blur_ks),
A.RandomHFlip(cfg.flip_p),
A.RandomRotate(cfg.rotate_p),
A.RandomCropToAspect(cfg.img_shape),
A.AutoContrast(),
A.ColorJitter(),
A.Occlusion(),
#A.RandomCrop(cfg.min_scale),
A.Resize(cfg.img_shape)
])
transforms_val = A.Compose([A.RandomCropToAspect(cfg.img_shape),A.Resize(cfg.img_shape)])
trainloader = get_loader(cfg.dataset, "train", cfg.dataset_dir, cfg.batch_size, transforms=transforms, num_workers=cfg.num_workers, pin_memory=True, shuffle=True)
valloader = get_loader(cfg.dataset, "val", cfg.dataset_dir, 1, transforms=transforms_val, num_workers=cfg.num_workers, pin_memory=True)
model = Model(cfg.num_classes, cfg.anchors, cfg.strides, cfg.reduction)
model.train()
model.cuda()
if cfg.pretrained is not None:
#state_dict = torch.load(cfg.pretrained, map_location="cpu")
#model.load_state_dict(state_dict)
pretrained_dict = torch.load(cfg.pretrained, map_location="cpu")
model_dict = model.state_dict()
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict and model_dict[k].shape == pretrained_dict[k].shape}
print(f"Following weight not loaded: {[k for k in model_dict.keys() if k not in pretrained_dict]}")
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
if cfg.reset_bias:
print("Resetting bias on pretrained model!")
for m in model.det_head.detect.m:
fan_in, _ = torch.nn.init._calculate_fan_in_and_fan_out(m.weight)
bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
torch.nn.init.uniform_(m.bias, -bound, bound)
if not cfg.backbone:
print("Freezing backbone.")
for k, v in model.backbone.named_parameters():
v.requires_grad = False
if not cfg.decoder:
print("Freezing decoder.")
for k, v in model.decoder.named_parameters():
v.requires_grad = False
if not cfg.head_seg:
print("Freezing seg head.")
for k, v in model.seg_head.named_parameters():
v.requires_grad = False
cfg.w_seg, cfg.w_iou = 0, 0
if not cfg.head_det:
print("Freezing det head.")
for k, v in model.det_head.named_parameters():
v.requires_grad = False
cfg.w_cls, cfg.w_obj, cfg.w_box = 0, 0, 0
criterion = CombinedLoss(cfg, cfg.num_classes, model.gr, anchors = cfg.anchors, stride=cfg.strides, device = next(model.parameters()).device)
if cfg.optimizer.lower() == "adam":
optimizer = optim.Adam(model.parameters(), lr=learning_rate, betas=(cfg.momentum, 0.999), weight_decay = cfg.weight_decay)
elif cfg.optimizer.lower() == "sgd":
optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=cfg.momentum, nesterov=True)
else:
raise RuntimeError(f"Optimizer {cfg.optimizer} not supported.")
iters_per_epoch = len(trainloader)
#scheduler = CosineAnnealingLR(optimizer,
# iters_per_epoch * cfg.epochs, eta_min = 0, warmup = cfg.warmup, warmup_iters = cfg.warmup_iters)
scheduler = LinearLR(optimizer,
iters_per_epoch * cfg.epochs, eta_min = 0, warmup = cfg.warmup, warmup_iters = cfg.warmup_iters)#scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
writer = SummaryWriter(cfg.save_dir)
n_iter = 0
val_loss_best = torch.inf
for epoch in range(epochs):
# train
model.train()
with tqdm(total=len(trainloader.dataset), desc ='Training - Epoch: '+str(epoch)+"/"+str(epochs), unit='chunks') as prog_bar:
for i, data in enumerate(trainloader):
inputs, labels = data
inputs = inputs.cuda()
optimizer.zero_grad()
outputs = model(inputs)
labels[0] = labels[0].cuda()
labels[1] = labels[1].cuda()
lseg, liou, lbox, lobj, lcls = criterion(outputs,labels)
loss = lseg + liou + lbox + lobj + lcls
loss.backward()
scheduler.step(n_iter)
logs = {
"train/l_combined":loss.item(),
"train/l_seg":lseg.item(),
"train/l_iou":liou.item(),
"train/l_box":lbox.item(),
"train/l_obj":lobj.item(),
"train/l_cls":lcls.item(),
"train/lr":scheduler.optimizer.param_groups[0]['lr']
}
for loss_name, loss_val in logs.items():
writer.add_scalar(loss_name, loss_val, n_iter)
n_iter += 1
optimizer.step()
prog_bar.set_postfix(**{'run:': "model_name",
**{key.split("/")[1] : value for key,value in logs.items()}})
prog_bar.update(batch_size)
# validate
model.eval()
idx_draw = random.sample(range(len(valloader)), min(len(valloader),cfg.val_plot_num))
losses = []
imgs = []
ious = []
with torch.no_grad():
for i, data in tqdm(enumerate(valloader)):
inputs, labels = data
inputs = inputs.cuda()
optimizer.zero_grad()
outputs = model(inputs)
labels[0] = labels[0].cuda()
labels[1] = labels[1].cuda()
# compute loss
lseg, liou, lbox, lobj, lcls = criterion([outputs[0], outputs[1][1]],labels)
loss = lseg + liou + lbox + lobj + lcls
losses.append(loss.item())
ious.append(1 - liou.item())
if i in idx_draw:
img_out = inputs.detach().cpu()
boxes_out = outputs[1][0].detach().cpu()
seg_out = outputs[0].detach().cpu()
img = draw_overlay(img_out[0], boxes_out[0].unsqueeze(0), seg_out[0][1], cfg.thr_conf, cfg.thr_iou, show=False)
imgs.append(img)
val_loss = np.mean(losses)
val_iou = np.mean(ious)
print(f"Val loss: {val_loss}, Val mIoU: {val_iou}")
writer.add_scalar("val/l_combined", val_loss, epoch)
writer.add_scalar("val/mIoU", val_iou, epoch)
writer.add_images("val/images", torch.stack(imgs), epoch)
torch.save(model.state_dict(), os.path.join(cfg.save_dir, "last.pt"))
if val_loss < val_loss_best:
torch.save(model.state_dict(), os.path.join(cfg.save_dir, "best.pt"))