solver.py 15.6 KB
Newer Older
1
2
3
4
"""Brain Mapper U-Net Solver

Description:

5
    This folder contains the Pytorch implementation of the core U-net solver, used for training the network.
6

7
Usage:
8

9
10
11
    To use this module, import it and instantiate is as you wish:

        from solver import Solver
12
13
14
15
16
"""

import os
import numpy as np
import torch
17
18
import glob

19
from fsl.data.image import Image
20
from fsl.utils.image.roi import roi
21
from datetime import datetime
22
from utils.losses import MSELoss
23
from utils.common_utils import create_folder
24
from utils.data_logging_utils import LogWriter
25
from utils.early_stopping import EarlyStopping
26
from torch.optim import lr_scheduler
27
28
29

checkpoint_extension = 'path.tar'

30

31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
class Solver():
    """Solver class for the BrainMapper U-net.

    This class contains the pytorch implementation of the U-net solver required for the BrainMapper project.

    Args:
        model (class): BrainMapper model class
        experiment_name (str): Name of the experiment
        device (int/str): Device type used for training (int - GPU id, str- CPU)
        number_of_classes (int): Number of classes
        optimizer (class): Pytorch class of desired optimizer
        optimizer_arguments (dict): Dictionary of arguments to be optimized
        loss_function (func): Function describing the desired loss function
        model_name (str): Name of the model
        labels (arr): Vector/Array of labels (if applicable)
        number_epochs (int): Number of training epochs
        loss_log_period (int): Period for writing loss value
        learning_rate_scheduler_step_size (int): Period of learning rate decay
        learning_rate_scheduler_gamma (int): Multiplicative factor of learning rate decay
        use_last_checkpoint (bool): Flag for loading the previous checkpoint
        experiment_directory (str): Experiment output directory name
        logs_directory (str): Directory for outputing training logs
53
        crop_flag (bool): Flag indicating if the volumes should be cropped from 91x109x91 to 72x90x77 to reduce storage space and speed-up training
54
55

    Returns:
56
        trained model - working on this!
57
58

    """
59

60
    def __init__(self,
61
62
63
64
                 model,
                 device,
                 number_of_classes,
                 experiment_name,
65
                 optimizer,
66
                 optimizer_arguments={},
67
68
69
                 loss_function=MSELoss(),
                #  loss_function=torch.nn.L1Loss(),
                #  loss_function=torch.nn.CosineEmbeddingLoss(),
70
71
72
73
74
75
76
77
                 model_name='BrainMapper',
                 labels=None,
                 number_epochs=10,
                 loss_log_period=5,
                 learning_rate_scheduler_step_size=5,
                 learning_rate_scheduler_gamma=0.5,
                 use_last_checkpoint=True,
                 experiment_directory='experiments',
78
                 logs_directory='logs',
79
80
                 checkpoint_directory='checkpoints',
                 save_model_directory='saved_models',
81
82
                 final_model_output_file='finetuned_alldata.pth.tar',
                 crop_flag = False
83
                 ):
84
85
86

        self.model = model
        self.device = device
87
88
89
90
        self.optimizer = optimizer(model.parameters(), **optimizer_arguments)

        if torch.cuda.is_available():
            self.loss_function = loss_function.cuda(device)
Andrei Roibu's avatar
Andrei Roibu committed
91
            self.MSE = MSELoss().cuda(device)
92
93
        else:
            self.loss_function = loss_function
Andrei Roibu's avatar
Andrei Roibu committed
94
95
            self.MSE = MSELoss()

96
97
98
99
        self.model_name = model_name
        self.labels = labels
        self.number_epochs = number_epochs
        self.loss_log_period = loss_log_period
100

101
102
        # We use a learning rate scheduler, that decays the LR of each paramter group by gamma every step_size epoch.
        self.learning_rate_scheduler = lr_scheduler.StepLR(self.optimizer,
103
104
105
                                                           step_size=learning_rate_scheduler_step_size,
                                                           gamma=learning_rate_scheduler_gamma)

106
107
        self.use_last_checkpoint = use_last_checkpoint

108
        experiment_directory_path = os.path.join(
109
            experiment_directory, experiment_name)
110
        self.experiment_directory_path = experiment_directory_path
111

112
113
        self.checkpoint_directory = checkpoint_directory

114
        create_folder(experiment_directory)
115
        create_folder(experiment_directory_path)
116
        create_folder(os.path.join(
117
            experiment_directory_path, self.checkpoint_directory))
118
119
120
121

        self.start_epoch = 1
        self.start_iteration = 1

122
123
124
125
126
        self.LogWriter = LogWriter(number_of_classes=number_of_classes,
                                   logs_directory=logs_directory,
                                   experiment_name=experiment_name,
                                   use_last_checkpoint=use_last_checkpoint,
                                   labels=labels)
127

128
        self.EarlyStopping = EarlyStopping(patience=10, min_delta=0)
129
130
        self.early_stop = False

131
132
133
134
        if crop_flag == False:
            self.MNI152_T1_2mm_brain_mask = torch.from_numpy(Image('utils/MNI152_T1_2mm_brain_mask.nii.gz').data)
        elif crop_flag == True:
            self.MNI152_T1_2mm_brain_mask = torch.from_numpy(roi(Image('utils/MNI152_T1_2mm_brain_mask.nii.gz'),((9,81),(10,100),(0,77))).data)
135

136
137
138
        self.save_model_directory = save_model_directory
        self.final_model_output_file = final_model_output_file

139
140
141
        if use_last_checkpoint:
            self.load_checkpoint()

142
    def train(self, train_loader, validation_loader):
143
144
145
146
147
148
        """Training Function

        This function trains a given model using the provided training data.

        Args:
            train_loader (class): Combined dataset and sampler, providing an iterable over the training dataset (torch.utils.data.DataLoader)
149
            validation_loader (class):  Combined dataset and sampler, providing an iterable over the validationing dataset (torch.utils.data.DataLoader)
150
151

        Returns:
152
            trained model
153
154
155
        """

        model, optimizer, learning_rate_scheduler = self.model, self.optimizer, self.learning_rate_scheduler
156
        dataloaders = {'train': train_loader, 'validation': validation_loader}
157
158

        if torch.cuda.is_available():
159
160
            torch.cuda.empty_cache()  # clear memory
            model.cuda(self.device)  # Moving the model to GPU
161

162
        previous_checkpoint = None
163
        previous_loss = None
Andrei Roibu's avatar
Andrei Roibu committed
164
        previous_MSE = None
165

166
167
168
169
        print('****************************************************************')
        print('TRAINING IS STARTING!')
        print('=====================')
        print('Model Name: {}'.format(self.model_name))
170
        if torch.cuda.is_available():
171
172
            print('Device Type: {}'.format(
                torch.cuda.get_device_name(self.device)))
173
174
        else:
            print('Device Type: {}'.format(self.device))
175
176
177
178
179
180
181
182
183
        start_time = datetime.now()
        print('Started At: {}'.format(start_time))
        print('----------------------------------------')

        iteration = self.start_iteration

        for epoch in range(self.start_epoch, self.number_epochs+1):
            print("Epoch {}/{}".format(epoch, self.number_epochs))

184
            for phase in ['train', 'validation']:
185
186
187
                print('-> Phase: {}'.format(phase))

                losses = []
Andrei Roibu's avatar
Andrei Roibu committed
188
                MSEs = []
189
190
191
192
193
194
195
196

                if phase == 'train':
                    model.train()
                else:
                    model.eval()

                for batch_index, sampled_batch in enumerate(dataloaders[phase]):
                    X = sampled_batch[0].type(torch.FloatTensor)
197
                    y = sampled_batch[1].type(torch.FloatTensor)
198

199
                    # We add an extra dimension (~ number of channels) for the 3D convolutions.
200
201
                    X = torch.unsqueeze(X, dim=1)
                    y = torch.unsqueeze(y, dim=1)
Andrei Roibu's avatar
Andrei Roibu committed
202
203
204

                    MNI152_T1_2mm_brain_mask = torch.unsqueeze(
                        torch.unsqueeze(self.MNI152_T1_2mm_brain_mask, dim=0), dim=0)
205

206
                    if model.test_if_cuda:
207
208
                        X = X.cuda(self.device, non_blocking=True)
                        y = y.cuda(self.device, non_blocking=True)
Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
209
210
                        MNI152_T1_2mm_brain_mask = MNI152_T1_2mm_brain_mask.cuda(
                            self.device, non_blocking=True)
211

212
                    y_hat = model(X)   # Forward pass & Masking
213

214
                    y_hat = torch.mul(y_hat, MNI152_T1_2mm_brain_mask)
215

216
217
                    loss = self.loss_function(y_hat, y)  # Loss computation
                    # loss = self.loss_function(y_hat+1e-4, y+1e-4, torch.tensor(1.0).cuda(self.device, non_blocking=True))
Andrei Roibu's avatar
Andrei Roibu committed
218
219
220
221

                    # We also calculate a separate MSE for cost function comparison!
                    MSE = self.MSE(y_hat, y)
                    MSEs.append(MSE.item())
222
223

                    if phase == 'train':
224
225
                        optimizer.zero_grad()  # Zero the parameter gradients
                        loss.backward()  # Backward propagation
226
227
228
229
                        optimizer.step()

                        if batch_index % self.loss_log_period == 0:

230
231
                            self.LogWriter.loss_per_iteration(
                                loss.item(), batch_index, iteration)
232

233
234
                        iteration += 1

Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
235
                    losses.append(loss.item())
236
237
238

                    # Clear the memory

Andrei Roibu's avatar
Andrei Roibu committed
239
                    del X, y, y_hat, loss, MNI152_T1_2mm_brain_mask, MSE
240
241
                    torch.cuda.empty_cache()

Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
242
                    if phase == 'validation':
243

244
245
246
247
248
249
250
                        if batch_index != len(dataloaders[phase]) - 1:
                            print("#", end='', flush=True)
                        else:
                            print("100%", flush=True)

                with torch.no_grad():

251
252
                    if phase == 'train':
                        self.LogWriter.loss_per_epoch(losses, phase, epoch)
Andrei Roibu's avatar
Andrei Roibu committed
253
                        self.LogWriter.MSE_per_epoch(MSEs, phase, epoch)
254
                    elif phase == 'validation':
Andrei Roibu's avatar
Andrei Roibu committed
255
256
                        self.LogWriter.loss_per_epoch(
                            losses, phase, epoch, previous_loss=previous_loss)
257
                        previous_loss = np.mean(losses)
Andrei Roibu's avatar
Andrei Roibu committed
258
259
260
                        self.LogWriter.MSE_per_epoch(
                            MSEs, phase, epoch, previous_loss=previous_MSE)
                        previous_MSE = np.mean(MSEs)
261

262
                    if phase == 'validation':
Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
263
264
                        early_stop, save_checkpoint = self.EarlyStopping(
                            np.mean(losses))
265
266
                        self.early_stop = early_stop
                        if save_checkpoint == True:
267
                            validation_loss = np.mean(losses)
268
269
                            checkpoint_name = os.path.join(
                                self.experiment_directory_path, self.checkpoint_directory, 'checkpoint_epoch_' + str(epoch) + '.' + checkpoint_extension)
270
271
272
273
274
275
276
                            self.save_checkpoint(state={'epoch': epoch + 1,
                                                        'start_iteration': iteration + 1,
                                                        'arch': self.model_name,
                                                        'state_dict': model.state_dict(),
                                                        'optimizer': optimizer.state_dict(),
                                                        'scheduler': learning_rate_scheduler.state_dict()
                                                        },
277
                                                 filename=checkpoint_name
Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
278
                                                 )
279
280
281
282
283
284

                            if previous_checkpoint != None:
                                os.remove(previous_checkpoint)
                                previous_checkpoint = checkpoint_name
                            else:
                                previous_checkpoint = checkpoint_name
285

286
287
288
                if phase == 'train':
                    learning_rate_scheduler.step()

289
            print("Epoch {}/{} DONE!".format(epoch, self.number_epochs))
290

291
292
293
294
            # Early Stop Condition

            if self.early_stop == True:
                print("ATTENTION!: Training stopped early to prevent overfitting!")
295
                self.load_checkpoint()
296
297
298
                break
            else:
                continue
299

300
301
302
303
304
305
306
        model_output_path = os.path.join(
            self.save_model_directory, self.final_model_output_file)

        create_folder(self.save_model_directory)

        model.save(model_output_path)

307
        self.LogWriter.close()
308
309
310
311
312
313
314

        print('----------------------------------------')
        print('TRAINING IS COMPLETE!')
        print('=====================')
        end_time = datetime.now()
        print('Completed At: {}'.format(end_time))
        print('Training Duration: {}'.format(end_time - start_time))
315
        print('Final Model Saved in: {}'.format(model_output_path))
316
317
        print('****************************************************************')

318
319
320
        if self.start_epoch >= self.number_epochs+1:
            validation_loss = None

321
322
        return validation_loss

323
324
325
326
327
328
329
330
331
332
    def save_checkpoint(self, state, filename):
        """General Checkpoint Save

        This function saves a general checkpoint for inference and/or resuming training

        Args:
            state (dict): Dictionary of all the relevant model components
        """

        torch.save(state, filename)
333

334
    def load_checkpoint(self, epoch=None):
335
336
337
338
339
340
341
        """General Checkpoint Loader

        This function loads a previous checkpoint for inference and/or resuming training

        Args:
            epoch (int): Current epoch value
        """
342

343
        if epoch is not None:
344
            checkpoint_file_path = os.path.join(
345
                self.experiment_directory_path, self.checkpoint_directory, 'checkpoint_epoch_' + str(epoch) + '.' + checkpoint_extension)
346
347
            self._checkpoint_reader(checkpoint_file_path)
        else:
348
            universal_path = os.path.join(
349
                self.experiment_directory_path, self.checkpoint_directory, '*.' + checkpoint_extension)
350
351
352
353
354
            files_in_universal_path = glob.glob(universal_path)

            # We will sort through all the files in path to see which one is most recent

            if len(files_in_universal_path) > 0:
355
356
                checkpoint_file_path = max(
                    files_in_universal_path, key=os.path.getatime)
357
358
359
                self._checkpoint_reader(checkpoint_file_path)

            else:
360
                self.LogWriter.log("No Checkpoint found at {}".format(
361
                    os.path.join(self.experiment_directory_path, self.checkpoint_directory)))
362

363
364
365
366
367
368
369
370
371
    def _checkpoint_reader(self, checkpoint_file_path):
        """Checkpoint Reader

        This private function reads a checkpoint file and then loads the relevant variables

        Args:
            checkpoint_file_path (str): path to checkpoint file
        """

372
373
        self.LogWriter.log(
            "Loading Checkpoint {}".format(checkpoint_file_path))
374
375
376
377

        checkpoint = torch.load(checkpoint_file_path)
        self.start_epoch = checkpoint['epoch']
        self.start_iteration = checkpoint['start_iteration']
378
        # We are not loading the model_name as we might want to pre-train a model and then use it.
379
380
381
382
        self.model.load_state_dict = checkpoint['state_dict']
        self.optimizer.load_state_dict = checkpoint['optimizer']

        for state in self.optimizer.state.values():
383
            for key, value in state.items():
384
385
386
                if torch.is_tensor(value):
                    state[key] = value.to(self.device)

387
        self.learning_rate_scheduler.load_state_dict = checkpoint['scheduler']
388
389
        self.LogWriter.log(
            "Checkpoint Loaded {} - epoch {}".format(checkpoint_file_path, checkpoint['epoch']))