-
Notifications
You must be signed in to change notification settings - Fork 3
/
base_train.py
363 lines (282 loc) · 14.2 KB
/
base_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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
import os
import ast
import torch
import numpy as np
import pandas as pd
from PIL import Image
from typing import Union, Dict, List, Tuple
from torch import nn, optim
from torch.utils.data.dataloader import DataLoader
from torchvision import transforms
from neural_net import *
from neural_net.sampler import ShuffleSampler
from neural_net.loss import GDiceLossV2
from neural_net.utils import *
from neural_net.transform import *
from neural_net.transf_learning import *
import segmentation_models_pytorch as smp
from neural_net.unet import UNet
from neural_net.attention_unet import AttentionUnet
from neural_net.canet_parts.networks.network import Comprehensive_Atten_Unet
import pytorch_lightning as pl
import wandb
def c_binary_mean_iou(logits: torch.Tensor, targets: torch.Tensor, EPSILON = 1e-15) -> torch.Tensor:
output = (logits > 0.5).int()
intersection = (targets * output).sum()
union = targets.sum() + output.sum() - intersection
result = (intersection + EPSILON) / (union + EPSILON)
return result
class Satmodel(pl.LightningModule):
def __init__(self, hparams, opt):
super().__init__()
for key in ["binary", "log_imgs", "log_res"]:
setattr(self, key, opt.get(key, True))
self.hparams.update(hparams)
if not "SLURM_JOB_ID" in os.environ:
hparams.num_workers = 0
hparams.batch_size = 8
self.save_hyperparameters()
self.model = self.get_model()
self.drop_last = False
if self.binary:
self.criterion = eval_object(hparams.criterion)
else:
self.criterion = eval_object(hparams.regr_criterion)
self.train_set = []
self.validation_set = []
self.test_set = []
self.add_nbr = False
def forward(self, batch):
return self.model(batch)
def get_model(self):
model = eval_object(self.hparams.model)
if type(model) == UNet:
model.apply(initialize_weight)
# model = Comprehensive_Atten_Unet(in_ch=12, n_classes=1, im_size=(480, 480))
return model
def configure_optimizers(self):
optimizer = eval_object(self.hparams.optimizer, params=self.model.parameters())
self.optimizers = [optimizer]
if self.hparams.get('scheduler'):
scheduler = eval_object(self.hparams.scheduler, optimizer)
if self.hparams.scheduler.name == optim.lr_scheduler.ReduceLROnPlateau:
return {
'optimizer': optimizer,
'lr_scheduler': scheduler,
'monitor': 'val_loss'
}
return self.optimizers, [scheduler]
return self.optimizers
def setup(self, stage=0):
self.hparams.num_workers = 0
self.hparams.batch_size = 8
ordered_keys = list(self.hparams.groups.keys())
validation_fold_name = self.hparams.validation_dict[self.hparams.key]
self.validation_set = self.hparams.groups[validation_fold_name]
print(f'Test set is {self.hparams.key}, validation set is {validation_fold_name}. All the rest is training set.')
for grp in self.hparams.groups:
if grp == validation_fold_name or grp == self.hparams.key:
continue
else:
self.train_set.extend(self.hparams.groups[grp])
self.test_set = self.hparams.groups[self.hparams.key]
# if self.binary:
# print(f'Training set ({len(self.train_set)}): {str(self.train_set)}')
# print(f'Validation set ({len(self.validation_set)}): {str(self.validation_set)}')
def train_dataloader(self) -> DataLoader:
train_dataset = SatelliteDataset(folder_list=self.train_set,
transform=self.hparams.train_transform,
**self.hparams.dataset_specs)
train_sampler = ShuffleSampler(train_dataset, self.hparams.seed)
result = DataLoader(train_dataset,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.num_workers,
sampler=train_sampler,
pin_memory=True,
drop_last=self.drop_last
)
print(f'Train set dim: {len(train_dataset)} ({len(result)} batches)')
return result
def val_dataloader(self):
validation_dataset = SatelliteDataset(folder_list=self.validation_set,
transform=self.hparams.test_transform,
**self.hparams.dataset_specs)
validation_sampler = ShuffleSampler(validation_dataset, self.hparams.seed)
result = DataLoader(validation_dataset,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.num_workers,
sampler=validation_sampler,
pin_memory=True,
drop_last=self.drop_last
)
print(f'Validation set dim: {len(validation_dataset)} ({len(result)} batches)')
return result
def test_dataloader(self):
test_dataset = SatelliteDataset(folder_list=self.test_set,
transform=self.hparams.test_transform,
**self.hparams.dataset_specs)
result = DataLoader(test_dataset,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.num_workers,
pin_memory=True,
drop_last=False
)
print(f'Validation set dim: {len(test_dataset)} ({len(result)} batches)')
return result
def training_step(self, batch, batch_idx):
images = batch["image"]
masks = batch["mask"]
if self.binary: masks = (masks > .5).type_as(masks)
logits = self.forward(images)
loss = self.criterion(logits, masks)
if wandb.run is not None:
if self.binary:
self.log('train_iou', binary_mean_iou(logits, masks))
else:
sq_err, counters = compute_squared_errors(logits, masks, len(self.hparams.dataset_specs.mask_intervals))
mse = np.true_divide(sq_err, counters, np.full(sq_err.shape, np.nan), where=counters != 0)
self.log('train_rmse', np.sqrt(mse[-1]))
self.log('lr', self._get_current_lr())
self.log('loss', loss)
return loss
def validation_step(self, batch, batch_idx):
images = batch["image"]
masks = batch["mask"]
logits = self.forward(images)
if self.binary: masks = (masks > .5).type_as(masks)
loss = self.criterion(logits, masks)
if self.log_res: self.log("val_loss", loss)
if self.binary:
val_kpi = binary_mean_iou(logits, masks)
if wandb.run is not None:
self.log('val_iou', val_kpi)
else:
sq_err, counters = compute_squared_errors(logits, masks, len(self.hparams.dataset_specs.mask_intervals))
mse = np.true_divide(sq_err, counters, np.full(sq_err.shape, np.nan), where=counters != 0)
val_kpi = np.sqrt(mse[-1])
if wandb.run is not None:
if self.log_res: self.log('val_rmse', val_kpi)
if self.log_imgs and self.log_res:
self.log_images(images, logits, masks)
return {'val_loss': loss}
def log_images(self, images, logits, masks, log_dist=3):
if self.binary:
class_labels = {0: "background", 1: "fire"}
logits_ = (torch.sigmoid(logits) > 0.5).cpu().detach().numpy().astype("float")
masks_ = (masks > 0.5).cpu().detach().numpy().astype("float")
else:
class_labels = {0: "unburnt", 1: "1", 2: "2", 3: "3", 4: "4"}
logits_ = logits.cpu().detach().numpy().astype("float")
masks_ = masks.cpu().detach().numpy().astype("float")
if self.trainer.current_epoch % log_dist == 0 and wandb.run is not None:
for i in range(images.shape[0]):
mask_img = wandb.Image(
images[i, [3,2,1], :, :]*2.5,
masks={
"predictions": {
"mask_data": logits_[i, 0, :, :],
"class_labels": class_labels,
},
"groud_truth": {
"mask_data": masks_[i, 0, :, :],
"class_labels": class_labels,
},
},
)
self.logger.experiment.log({"val_images": [mask_img]}, commit=False)
def _get_current_lr(self):
lr = [x["lr"] for x in self.optimizers[0].param_groups][0] # type: ignore
if torch.cuda.is_available(): return torch.Tensor([lr])[0].cuda()
return torch.Tensor([lr])[0]
def validation_end(self, outputs):
# OPTIONAL
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
tensorboard_logs = {'val_loss': avg_loss}
return {'val_loss': avg_loss}
# def validation_epoch_end(self, outputs):
# self.log("epoch", self.trainer.current_epoch)
# if self.binary:
# avg_val_iou = find_average(outputs, "val_iou")
# self.log("val_iou", avg_val_iou)
# else:
# avg_val_rmse = find_average(outputs, "val_rmse")
# self.log("val_rmse", avg_val_rmse)
# return
class Double_Satmodel(Satmodel):
def __init__(self, hparams, opt):
super().__init__(hparams, opt)
if self.binary:
print(f'\n Iteration step 1/2 - Binary network training... (test on {self.hparams.key} fold)\n')
else:
print(f'\n Iteration step 2/2 - Regression network training... (test on {self.hparams.key} fold)\n')
self.set_model()
def set_model(self):
self.hparams.backbone = self.model.backbone
if self.binary:
if type(self.model.binary_unet) == UNet: self.model.apply(initialize_weight)
self.model.unfreeze_binary_unet()
self.model.freeze_regression_unet()
else:
self.model.freeze_binary_unet()
self.model.unfreeze_regression_unet()
def forward(self, batch):
if self.binary:
self.model.regression_unet.eval()
self.model.freeze_regression_unet()
return self.model(batch)[0]
else:
self.model.binary_unet.eval()
self.model.freeze_binary_unet()
return self.model(batch)[1]
# mytype = torch.float32
def configure_optimizers(self):
if self.binary:
optimizer = eval_object(self.hparams.optimizer, params=self.model.binary_unet.parameters())
else:
optimizer = eval_object(self.hparams.optimizer, params=self.model.parameters())
self.optimizers = [optimizer]
if self.hparams.get('scheduler'):
scheduler = eval(self.hparams.scheduler.pop('name'))(optimizer, **self.hparams.scheduler)
if self.hparams.scheduler.name == optim.lr_scheduler.ReduceLROnPlateau:
return {
'optimizer': optimizer,
'lr_scheduler': scheduler,
'monitor': 'val_loss'
}
return self.optimizers, [scheduler]
return self.optimizers
def test_step(self, batch, batch_idx):
images = batch["image"]
masks = batch["mask"]
bin_pred, regr_pred = self.model(images)
bin_pred = torch.sigmoid(bin_pred).squeeze().cpu().detach().numpy() > 0.5
intersection, union = binary_mean_iou(masks.cpu().detach().numpy(), bin_pred, average=False)
intersection = intersection.sum()
union = union.sum()
regr_pred = regr_pred.clamp(0, max=(len(self.hparams.dataset_specs.mask_intervals) - 1))
tmp_sq_err, tmp_counters = compute_squared_errors(regr_pred, masks,
len(self.hparams.dataset_specs.mask_intervals)
)
severity_pred = regr_pred.round().squeeze().cpu().detach().numpy().astype("float")
severity_pred *= (255.0/(len(self.hparams.dataset_specs.mask_intervals) - 1))
out_path = self.hparams.checkpoint.dirpath / "predictions"
out_path.mkdir(parents=True, exist_ok=True)
for i in range(images.shape[0]):
Image.fromarray(255*bin_pred[i].astype(float)).convert('RGB').save(out_path / f"bin_{batch_idx*self.hparams.batch_size + i}.png")
Image.fromarray(severity_pred[i]).convert('RGB').save(out_path / f"sev_{batch_idx*self.hparams.batch_size + i}.png")
return {'regr': [tmp_sq_err, tmp_counters], 'bin': [intersection, union]}
def test_epoch_end(self, outputs):
r_outputs = np.array([out['regr'] for out in outputs])
bin_outputs = np.array([out['bin'] for out in outputs])
sqe = r_outputs[:, 0, :].sum(axis=0)
counters = r_outputs[:, 1, :].sum(axis=0)
intersections, unions = bin_outputs.sum(0)
ious = (intersections + 1e-6) / (unions + 1e-6)
pd.DataFrame([intersections, unions, ious],
index=['intersection', 'union', 'iou'],
columns=['value'],
).to_csv(self.hparams.checkpoint.dirpath / "bin_kpi.csv")
mse = np.true_divide(sqe, counters, np.full(sqe.shape, np.nan), where=counters != 0)
pd.DataFrame([sqe, counters, np.sqrt(mse)],
index=['sqe', 'count', 'rmse'],
).rename(columns={5: 'all'}).to_csv(self.hparams.checkpoint.dirpath / "regr_kpi.csv")