-
Notifications
You must be signed in to change notification settings - Fork 77
/
train_deep_globe.py
executable file
·183 lines (154 loc) · 9.97 KB
/
train_deep_globe.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
#!/usr/bin/env python
# coding: utf-8
from __future__ import absolute_import, division, print_function
import os
import numpy as np
import torch
import torch.nn as nn
from torchvision import transforms
from tqdm import tqdm
from dataset.deep_globe import DeepGlobe, classToRGB, is_image_file
from utils.loss import CrossEntropyLoss2d, SoftCrossEntropyLoss2d, FocalLoss
from utils.lovasz_losses import lovasz_softmax
from utils.lr_scheduler import LR_Scheduler
from tensorboardX import SummaryWriter
from helper import create_model_load_weights, get_optimizer, Trainer, Evaluator, collate, collate_test
from option import Options
args = Options().parse()
n_class = args.n_class
# torch.cuda.synchronize()
# torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = True
data_path = args.data_path
model_path = args.model_path
if not os.path.isdir(model_path): os.mkdir(model_path)
log_path = args.log_path
if not os.path.isdir(log_path): os.mkdir(log_path)
task_name = args.task_name
print(task_name)
###################################
mode = args.mode # 1: train global; 2: train local from global; 3: train global from local
evaluation = args.evaluation
test = evaluation and False
print("mode:", mode, "evaluation:", evaluation, "test:", test)
###################################
print("preparing datasets and dataloaders......")
batch_size = args.batch_size
ids_train = [image_name for image_name in os.listdir(os.path.join(data_path, "train", "Sat")) if is_image_file(image_name)]
ids_val = [image_name for image_name in os.listdir(os.path.join(data_path, "crossvali", "Sat")) if is_image_file(image_name)]
ids_test = [image_name for image_name in os.listdir(os.path.join(data_path, "offical_crossvali", "Sat")) if is_image_file(image_name)]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dataset_train = DeepGlobe(os.path.join(data_path, "train"), ids_train, label=True, transform=True)
dataloader_train = torch.utils.data.DataLoader(dataset=dataset_train, batch_size=batch_size, num_workers=10, collate_fn=collate, shuffle=True, pin_memory=True)
dataset_val = DeepGlobe(os.path.join(data_path, "crossvali"), ids_val, label=True)
dataloader_val = torch.utils.data.DataLoader(dataset=dataset_val, batch_size=batch_size, num_workers=10, collate_fn=collate, shuffle=False, pin_memory=True)
dataset_test = DeepGlobe(os.path.join(data_path, "offical_crossvali"), ids_test, label=False)
dataloader_test = torch.utils.data.DataLoader(dataset=dataset_test, batch_size=batch_size, num_workers=10, collate_fn=collate_test, shuffle=False, pin_memory=True)
##### sizes are (w, h) ##############################
# make sure margin / 32 is over 1.5 AND size_g is divisible by 4
size_g = (args.size_g, args.size_g) # resized global image
size_p = (args.size_p, args.size_p) # cropped local patch size
sub_batch_size = args.sub_batch_size # batch size for train local patches
###################################
print("creating models......")
path_g = os.path.join(model_path, args.path_g)
path_g2l = os.path.join(model_path, args.path_g2l)
path_l2g = os.path.join(model_path, args.path_l2g)
model, global_fixed = create_model_load_weights(n_class, mode, evaluation, path_g=path_g, path_g2l=path_g2l, path_l2g=path_l2g)
###################################
num_epochs = args.num_epochs
learning_rate = args.lr
lamb_fmreg = args.lamb_fmreg
optimizer = get_optimizer(model, mode, learning_rate=learning_rate)
scheduler = LR_Scheduler('poly', learning_rate, num_epochs, len(dataloader_train))
##################################
criterion1 = FocalLoss(gamma=3)
criterion2 = nn.CrossEntropyLoss()
criterion3 = lovasz_softmax
criterion = lambda x,y: criterion1(x, y)
# criterion = lambda x,y: 0.5*criterion1(x, y) + 0.5*criterion3(x, y)
mse = nn.MSELoss()
if not evaluation:
writer = SummaryWriter(log_dir=log_path + task_name)
f_log = open(log_path + task_name + ".log", 'w')
trainer = Trainer(criterion, optimizer, n_class, size_g, size_p, sub_batch_size, mode, lamb_fmreg)
evaluator = Evaluator(n_class, size_g, size_p, sub_batch_size, mode, test)
best_pred = 0.0
print("start training......")
for epoch in range(num_epochs):
trainer.set_train(model)
optimizer.zero_grad()
tbar = tqdm(dataloader_train); train_loss = 0
for i_batch, sample_batched in enumerate(tbar):
if evaluation: break
scheduler(optimizer, i_batch, epoch, best_pred)
loss = trainer.train(sample_batched, model, global_fixed)
train_loss += loss.item()
score_train, score_train_global, score_train_local = trainer.get_scores()
if mode == 1: tbar.set_description('Train loss: %.3f; global mIoU: %.3f' % (train_loss / (i_batch + 1), np.mean(np.nan_to_num(score_train_global["iou"]))))
else: tbar.set_description('Train loss: %.3f; agg mIoU: %.3f' % (train_loss / (i_batch + 1), np.mean(np.nan_to_num(score_train["iou"]))))
score_train, score_train_global, score_train_local = trainer.get_scores()
trainer.reset_metrics()
# torch.cuda.empty_cache()
if epoch % 1 == 0:
with torch.no_grad():
model.eval()
print("evaluating...")
if test: tbar = tqdm(dataloader_test)
else: tbar = tqdm(dataloader_val)
for i_batch, sample_batched in enumerate(tbar):
predictions, predictions_global, predictions_local = evaluator.eval_test(sample_batched, model, global_fixed)
score_val, score_val_global, score_val_local = evaluator.get_scores()
# use [1:] since class0 is not considered in deep_globe metric
if mode == 1: tbar.set_description('global mIoU: %.3f' % (np.mean(np.nan_to_num(score_val_global["iou"])[1:])))
else: tbar.set_description('agg mIoU: %.3f' % (np.mean(np.nan_to_num(score_val["iou"])[1:])))
images = sample_batched['image']
if not test:
labels = sample_batched['label'] # PIL images
if test:
if not os.path.isdir("./prediction/"): os.mkdir("./prediction/")
for i in range(len(images)):
if mode == 1:
transforms.functional.to_pil_image(classToRGB(predictions_global[i]) * 255.).save("./prediction/" + sample_batched['id'][i] + "_mask.png")
else:
transforms.functional.to_pil_image(classToRGB(predictions[i]) * 255.).save("./prediction/" + sample_batched['id'][i] + "_mask.png")
if not evaluation and not test:
if i_batch * batch_size + len(images) > (epoch % len(dataloader_val)) and i_batch * batch_size <= (epoch % len(dataloader_val)):
writer.add_image('image', transforms.ToTensor()(images[(epoch % len(dataloader_val)) - i_batch * batch_size]), epoch)
if not test:
writer.add_image('mask', classToRGB(np.array(labels[(epoch % len(dataloader_val)) - i_batch * batch_size])) * 255., epoch)
if mode == 2 or mode == 3:
writer.add_image('prediction', classToRGB(predictions[(epoch % len(dataloader_val)) - i_batch * batch_size]) * 255., epoch)
writer.add_image('prediction_local', classToRGB(predictions_local[(epoch % len(dataloader_val)) - i_batch * batch_size]) * 255., epoch)
writer.add_image('prediction_global', classToRGB(predictions_global[(epoch % len(dataloader_val)) - i_batch * batch_size]) * 255., epoch)
# torch.cuda.empty_cache()
# if not (test or evaluation): torch.save(model.state_dict(), "./saved_models/" + task_name + ".epoch" + str(epoch) + ".pth")
if not (test or evaluation): torch.save(model.state_dict(), "./saved_models/" + task_name + ".pth")
if test: break
else:
score_val, score_val_global, score_val_local = evaluator.get_scores()
evaluator.reset_metrics()
if mode == 1:
if np.mean(np.nan_to_num(score_val_global["iou"][1:])) > best_pred: best_pred = np.mean(np.nan_to_num(score_val_global["iou"][1:]))
else:
if np.mean(np.nan_to_num(score_val["iou"][1:])) > best_pred: best_pred = np.mean(np.nan_to_num(score_val["iou"][1:]))
log = ""
log = log + 'epoch [{}/{}] IoU: train = {:.4f}, val = {:.4f}'.format(epoch+1, num_epochs, np.mean(np.nan_to_num(score_train["iou"][1:])), np.mean(np.nan_to_num(score_val["iou"][1:]))) + "\n"
log = log + 'epoch [{}/{}] Local -- IoU: train = {:.4f}, val = {:.4f}'.format(epoch+1, num_epochs, np.mean(np.nan_to_num(score_train_local["iou"][1:])), np.mean(np.nan_to_num(score_val_local["iou"][1:]))) + "\n"
log = log + 'epoch [{}/{}] Global -- IoU: train = {:.4f}, val = {:.4f}'.format(epoch+1, num_epochs, np.mean(np.nan_to_num(score_train_global["iou"][1:])), np.mean(np.nan_to_num(score_val_global["iou"][1:]))) + "\n"
log = log + "train: " + str(score_train["iou"]) + "\n"
log = log + "val:" + str(score_val["iou"]) + "\n"
log = log + "Local train:" + str(score_train_local["iou"]) + "\n"
log = log + "Local val:" + str(score_val_local["iou"]) + "\n"
log = log + "Global train:" + str(score_train_global["iou"]) + "\n"
log = log + "Global val:" + str(score_val_global["iou"]) + "\n"
log += "================================\n"
print(log)
if evaluation: break
f_log.write(log)
f_log.flush()
if mode == 1:
writer.add_scalars('IoU', {'train iou': np.mean(np.nan_to_num(score_train_global["iou"][1:])), 'validation iou': np.mean(np.nan_to_num(score_val_global["iou"][1:]))}, epoch)
else:
writer.add_scalars('IoU', {'train iou': np.mean(np.nan_to_num(score_train["iou"][1:])), 'validation iou': np.mean(np.nan_to_num(score_val["iou"][1:]))}, epoch)
if not evaluation: f_log.close()