-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
218 lines (188 loc) · 7.31 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
import math
import os
import sys
from pathlib import Path
import cv2
import numpy as np
import torch
import wandb
from dataset.dataset_factory import create_dataset
from dataset.loader import create_loader
from fasterrcnn import get_model
from utils import torch_utils
from utils.coco_evaluate import evaluate
from utils.load_config import load_yaml
yaml_config = "config.yaml"
hyp = load_yaml(yaml_config)
output_dir = os.path.join(os.getcwd(), "output")
sweep_config = {
"method": "random",
"metric": {"goal": "maximize", "name": "mAP_050"},
"parameters": {
"lr": {"values": [0.0005, 0.005, 0.010]},
"weight_decay": {"values": [0.0005, 0.005]},
"batch_size": {"values": [4, 8, 12]},
"input_size": {"values": [128, 224, 448]},
"num_epochs": {"values": [10, 15, 20]},
},
}
# opt.opt:
# values: ["SGD", "Adam"]
# Now initialize the sweep
sweep_id = wandb.sweep(sweep_config, project="open-images-detection")
def create_datasets_and_loaders(input_size, batch_size, transform_train_fn=None):
# input_size = 224 # input of image
# batch_size = 2
root = Path(hyp["root"])
dataset_train, dataset_val = create_dataset(root)
print(dataset_train.__len__())
print(dataset_val.__len__())
# visualize_input(dataset_train)
loader_train = create_loader(
dataset_train,
input_size,
batch_size,
interpolation=hyp["interpolation"],
fill_color=hyp["fill_color"],
num_workers=hyp["num_workers"],
is_training=True,
)
loader_val = create_loader(
dataset_val,
input_size,
batch_size,
interpolation=hyp["interpolation"],
fill_color=hyp["fill_color"],
num_workers=hyp["num_workers"],
is_training=False,
)
return loader_train, loader_val
def visualize_input(dataset):
pil_img, target = dataset.__getitem__(1)
bboxes = target["boxes"]
color = (255, 0, 0)
thickness = 2
for bbox in bboxes:
bbox = list(map(lambda x: int(x), bbox))
print(bbox)
cv_img = np.array(pil_img)
cv_img = cv2.cvtColor(cv_img, cv2.COLOR_RGB2BGR)
startp = (bbox[0], bbox[1])
endp = (bbox[2], bbox[3])
cv2.rectangle(cv_img, startp, endp, color, thickness)
cv2.imwrite("hi.jpg", cv_img)
print("img created")
def train(model, optimizer, data_loader, device, epoch, print_freq):
model.train()
metric_logger = torch_utils.MetricLogger(delimiter=" ")
metric_logger.add_meter(
"lr", torch_utils.SmoothedValue(window_size=1, fmt="{value:.6f}")
)
header = "Epoch: [{}]".format(epoch)
lr_scheduler = None
if epoch == 0:
warmup_factor = 1.0 / 1000
warmup_iters = min(1000, len(data_loader) - 1)
lr_scheduler = torch_utils.warmup_lr_scheduler(
optimizer, warmup_iters, warmup_factor
)
for batch_idx, (inputs, targets) in enumerate(
metric_logger.log_every(data_loader, print_freq, header)
):
inputs = list(img.to(device) for img in inputs)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
loss_dict = model(inputs, targets)
losses = sum(loss for loss in loss_dict.values())
loss_dict_reduced = torch_utils.reduce_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss_value = losses_reduced.item()
if batch_idx % print_freq == 0:
wandb.log({"Train loss": loss_value})
# torchvision.utils.save_image(list(inputs),
# "train%s-batch.jpg" %batch_idx, padding=0, normalize=True)
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
optimizer.zero_grad()
losses.backward()
optimizer.step()
if lr_scheduler is not None:
lr_scheduler.step()
metric_logger.update(loss=losses_reduced, **loss_dict_reduced)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
wandb.log({"lr": optimizer.param_groups[0]["lr"]})
return metric_logger, loss_value
def validate(model, optimizer, data_loader, device, print_freq):
# https://discuss.pytorch.org/t/compute-validation-loss-for-faster-rcnn/62333
model.train()
for batch_idx, (inputs, targets) in enumerate(data_loader):
inputs = list(img.to(device) for img in inputs)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
with torch.no_grad():
loss_dict = model(inputs, targets)
loss_dict_reduced = torch_utils.reduce_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss_value = losses_reduced.item()
if batch_idx % print_freq == 0:
wandb.log({"Validation loss": loss_value})
# torchvision.utils.save_image(list(inputs),
# "train%s-batch.jpg" %batch_idx, padding=0, normalize=True)
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
return loss_value
print_freq = 10
def run_training():
with wandb.init(config=hyp, entity="dmatos"):
config = wandb.config
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = get_model()
wandb.watch(model)
loader_train, loader_val = create_datasets_and_loaders(
config["input_size"], config["batch_size"]
)
model.to(device)
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(
params,
lr=config["lr"],
momentum=config["momentum"],
weight_decay=config["weight_decay"],
)
lr_scheduler = torch.optim.lr_scheduler.StepLR(
optimizer, step_size=config["step_size"], gamma=config["gamma"]
)
for epoch in range(config["num_epochs"]):
# run_url = wandb.run.get_url()
_, train_loss_epoch = train(
model, optimizer, loader_train, device, epoch, print_freq=print_freq
)
val_loss_epoch = validate(
model, optimizer, loader_val, device, print_freq=print_freq
)
wandb.log(
{
"Epoch Train Loss": train_loss_epoch,
"Epoch Validation Loss": val_loss_epoch,
}
)
# losses = [[train_loss, val_loss]]
# table = wandb.Table(data=losses, columns=["Train Loss", "Validation Loss"])
if ((epoch + 1) % 5 == 0) or epoch == 0:
torch.save(
{
"epoch": epoch,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
},
output_dir
+ "/"
+ "model_ckpt_epoch%s_sweep%s.pth" % (epoch, sweep_id),
)
_, coco_stats = evaluate(model, loader_val, device=device)
wandb.log({"mAP_050": coco_stats[1]})
lr_scheduler.step()
count = 5
wandb.agent(sweep_id, function=run_training, count=count)