-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_soan_wn.py
236 lines (216 loc) · 10.1 KB
/
main_soan_wn.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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
#!/usr/bin/env python
import argparse
import pathlib
import yaml
import json
from addict import Dict
import torch
import numpy as np
import os
import glob
import time
import matplotlib.pyplot as plt
from utils.data_loader import fetch_loaders
from utils.frame import Framework, IOStream
from utils.metrics import diceloss, bce_diceloss
from torch.utils.tensorboard import SummaryWriter
import utils.train as tr
def _init_():
if not os.path.exists('outputs'):
os.makedirs('outputs')
if not os.path.exists('outputs/'+args.exp_name):
os.makedirs('outputs/'+args.exp_name)
if not os.path.exists('outputs/'+args.exp_name+'/'+'models'):
os.makedirs('outputs/'+args.exp_name+'/'+'models')
os.system('cp main.py outputs'+'/'+args.exp_name+'/'+'main.py.backup')
os.system('cp models/SoAN_wn.py outputs' + '/' + args.exp_name + '/' + 'SoAN_wn.py.backup')
os.system('cp utils/data_loader.py outputs' + '/' + args.exp_name + '/' + 'data_loader.py.backup')
def train(args, io):
data_dir = args.data_dir
model_dir = args.model_dir
conf = Dict(yaml.safe_load(open(args.train_yaml, "r")))
loss_type = args.loss_type
device = torch.device("cuda" if args.cuda else "cpu")
# device = args.device
# if device is not None:
# device = torch.device(device)
args = Dict({
"batch_size": args.batch_size,
"exp_name": args.exp_name,
"epochs": args.epochs,
"save_every": args.save_every
})
loaders = fetch_loaders(data_dir, args.batch_size, shuffle=True)
# test_loader = fetch_loaders(data_dir, args.batch_size, folder='test', shuffle=False)
# if input mask dimension different than outchannels
outchannels = conf.model_opts.args.outchannels
# y_channels = [y.shape[-1] for _, y in loaders["test"]][0] ### y_channels = test loader if there isn't validation set
# if y_channels != outchannels:
# raise ValueError("Output dimension is different from model outchannels.")
# TODO: try to have less nested if/else
# get dice loss
if loss_type == "bce_dice":
loss_weight = [0.6, 0.4] # clean ice, debris, background [0.7, 0.3]
label_smoothing = 0.2
loss_fn = bce_diceloss(act=torch.nn.Softmax(dim=1), w=loss_weight,
outchannels=outchannels, label_smoothing=label_smoothing)
else:
loss_weight = [0.6, 0.4]
label_smoothing = 0.2
loss_fn = diceloss(act=torch.nn.Softmax(dim=1), w=loss_weight,
outchannels=outchannels, label_smoothing=label_smoothing)
# Try to load the models
frame = Framework(
model_opts=conf.model_opts,
optimizer_opts=conf.optim_opts,
reg_opts=conf.reg_opts,
loss_fn=loss_fn,
epochs=args.epochs,
device=device
)
# print(str(frame))
# Setup logging
writer = SummaryWriter(f"{model_dir}/{args.exp_name}/logs/")
writer.add_text("Arguments", json.dumps(vars(args)))
writer.add_text("Configuration Parameters", json.dumps(conf))
out_dir = f"{model_dir}/{args.exp_name}/models/"
mask_names = conf.log_opts.mask_names
best_test_iou = 0
for epoch in range(args.epochs):
# train loop
loss_d = {}
loss_d["train"], train_metrics = tr.train_epoch(loaders["train"], frame, conf.metrics_opts)
tr.log_metrics(writer, train_metrics, loss_d["train"], epoch, mask_names=mask_names)
if (epoch + 1) % args.save_every == 0:
tr.log_images(writer, frame, next(iter(loaders["train"])), epoch)
outstr = 'Train Epoch %d, Loss: %.6f, Train Glacial Lake IoU: %.6f, Train Background IoU: %.6f' % (
epoch, loss_d['train'], train_metrics['IoU'][0], train_metrics['IoU'][1])
io.cprint(outstr)
# Validation loop
loss_d["val"], val_metrics = tr.validate(loaders["val"], frame, conf.metrics_opts)
tr.log_metrics(writer, val_metrics, loss_d["val"], epoch, "val", mask_names=mask_names)
# if (epoch + 1) % args.save_every == 0:
# tr.log_images(writer, frame, next(iter(loaders["val"])), epoch, "val")
outstr = 'Val Epoch %d, Loss: %.6f, Val Glacial Lake IoU: %.6f, Val Background IoU: %.6f' % (
epoch, loss_d['val'], val_metrics['IoU'][0], val_metrics['IoU'][1])
io.cprint(outstr)
# Save model
writer.add_scalars("Loss", loss_d, epoch)
if (epoch + 1) % args.save_every == 0:
frame.save(out_dir, epoch)
if np.mean(val_metrics['IoU'][0].cpu().detach().numpy()) >= best_test_iou:
best_test_iou = np.mean(val_metrics['IoU'][0].cpu().detach().numpy())
frame.save(out_dir, "optimal")
tr.log_images(writer, frame, next(iter(loaders["val"])), epoch, "val")
# frame.save(out_dir, "final")
writer.close()
def test(args, io):
data_dir = args.data_dir
model_dir = args.model_dir
conf = Dict(yaml.safe_load(open(args.train_yaml, "r")))
outchannels = conf.model_opts.args.outchannels
loaders = fetch_loaders(data_dir, args.batch_size, shuffle=True)
device = torch.device("cuda" if args.cuda else "cpu")
# TODO: try to have less nested if/else
loss_type = args.loss_type
# get dice loss
if loss_type == "bce_dice":
loss_weight = [0.6, 0.4] # clean ice, debris, background [0.7, 0.3]
label_smoothing = 0.2
loss_fn = bce_diceloss(act=torch.nn.Softmax(dim=1), w=loss_weight,
outchannels=outchannels, label_smoothing=label_smoothing)
else:
loss_weight = [0.6, 0.4]
label_smoothing = 0.2
loss_fn = diceloss(act=torch.nn.Softmax(dim=1), w=loss_weight,
outchannels=outchannels, label_smoothing=label_smoothing)
# Try to load the models
frame = Framework(
model_opts=conf.model_opts,
optimizer_opts=conf.optim_opts,
reg_opts=conf.reg_opts,
loss_fn=loss_fn,
device=device
)
frame.model.load_state_dict(torch.load(args.saved_model_dir))
model = frame.model.to(device)
model = model.eval()
t_metrics = {}
# channel_first = lambda x: x.permute(0, 3, 1, 2)
slices_dir = f"{data_dir}/test/*img*"
pred_dir = f"{model_dir}/{args.exp_name}/preds/"
if not os.path.exists(pred_dir):
os.makedirs(pred_dir)
slices = glob.glob(slices_dir)
total_inference_time = 0
for s in slices:
filename = s.split("/")[-1].replace("npy", "png")
inp_np = np.load(s)
start = time.time()
nan_mask = np.isnan(inp_np[:, :, :]).any(axis=2)
inp_tensor = torch.from_numpy(np.expand_dims(np.transpose(inp_np, (2, 0, 1)), axis=0))
inp_tensor = inp_tensor.to(device)
output = model(inp_tensor)
output_np = output.detach().cpu().numpy()
output_np = np.transpose(output_np[0], (1, 2, 0))
output_np = np.argmax(output_np, axis=2)
output_np[nan_mask] = 3
total_inference_time += (time.time() - start)
average_inference_time = total_inference_time/len(slices)
# plt.imsave(f"{pred_dir}{filename}", output_np, vmin=0, vmax=3)
plt.imsave(f"{pred_dir}{filename}", 1 - output_np, cmap='gray')
print(f"Total Inference Time : {total_inference_time}")
print(f"Average Inference Time for an Image: {average_inference_time}")
for x, y in loaders["test"]:
x = x.permute(0, 3, 1, 2).to(device)
y_hat = model(x).permute(0, 2, 3, 1)
y_hat = frame.segment(y_hat)
metrics_ = frame.metrics(y_hat, y, conf.metrics_opts)
tr.update_metrics(t_metrics, metrics_)
tr.agg_metrics(t_metrics)
outstr = 'Test Glacial Lake IoU: %.6f, Test Background IoU: %.6f, \n' \
'Test Glacial Lake Pixel_Acc: %.6f, Test Background Pixel_Acc: %.6f, \n' \
'Test Glacial Lake Precision: %.6f, Test Background Precision: %.6f, \n' \
'Test Glacial Lake Recall: %.6f, Test Background Recall: %.6f, \n' \
'Test Glacial Lake Dice Score: %.6f, Test Background Dice Score: %.6f, \n' \
'Test Glacial Lake F1 Score: %.6f, Test Background F1 Score: %.6f, \n' \
'Test Glacial Lake F2 Score: %.6f, Test Background F2 Score: %.6f,' \
% (t_metrics['IoU'][0], t_metrics['IoU'][1],
t_metrics['pixel_acc'][0], t_metrics['pixel_acc'][1],
t_metrics['precision'][0], t_metrics['precision'][1],
t_metrics['recall'][0], t_metrics['recall'][1],
t_metrics['dice'][0], t_metrics['dice'][1],
t_metrics['f1_score'][0], t_metrics['f1_score'][1],
t_metrics['f2_score'][0], t_metrics['f2_score'][1])
io.cprint(outstr)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Preprocess raw tiffs into slices")
parser.add_argument("--data_dir", type=str, default="patches/splits/")
parser.add_argument("--model_dir", type=str, default="outputs/")
parser.add_argument("--saved_model_dir", type=str, default="outputs/SoAN_wn/models/model_optimal.pt")
parser.add_argument("--train_yaml", type=str, default="conf/train_soan_wn.yaml")
parser.add_argument("--batch_size", type=int, default=32)
parser.add_argument("--exp_name", type=str, default="SoAN_wn")
parser.add_argument("--epochs", type=int, default=300)
parser.add_argument("--save_every", type=int, default=10)
parser.add_argument("--loss_type", type=str, default="dice")
parser.add_argument('--seed', type=int, default=1, metavar='S', help='random seed (default: 1)')
parser.add_argument('--eval', type=bool, default=False, help='evaluate the model')
parser.add_argument("--device", type=str, default='cuda')
args = parser.parse_args()
_init_()
io = IOStream('outputs/' + args.exp_name + '/run.log')
io.cprint(str(args))
args.cuda = torch.cuda.is_available()
torch.manual_seed(args.seed)
if args.cuda:
io.cprint(
'Using GPU : ' + str(torch.cuda.current_device()) + ' from ' + str(torch.cuda.device_count()) + ' devices')
torch.cuda.manual_seed(args.seed)
else:
io.cprint('Using CPU')
# test(args, io)
if not args.eval:
train(args, io)
else:
test(args, io)