solver.py 13.3 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 datetime import datetime
21
22
from utils.losses import MSELoss
from utils.data_utils import create_folder
23
from utils.data_logging_utils import LogWriter
24
from utils.early_stopping import EarlyStopping
25
from torch.optim import lr_scheduler
26
27
28

checkpoint_extension = 'path.tar'

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
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

    Returns:
54
        trained model - working on this!
55
56

    """
57

58
    def __init__(self,
59
60
61
62
                 model,
                 device,
                 number_of_classes,
                 experiment_name,
63
                 optimizer,
64
65
66
67
68
69
70
71
72
73
                 optimizer_arguments={},
                 loss_function=MSELoss(),
                 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',
74
                 logs_directory='logs',
Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
75
                 checkpoint_directory='checkpoints'
76
                 ):
77
78
79

        self.model = model
        self.device = device
80
81
82
83
84
85
86
87
88
89
90
        self.optimizer = optimizer(model.parameters(), **optimizer_arguments)

        if torch.cuda.is_available():
            self.loss_function = loss_function.cuda(device)
        else:
            self.loss_function = loss_function

        self.model_name = model_name
        self.labels = labels
        self.number_epochs = number_epochs
        self.loss_log_period = loss_log_period
91

92
93
        # 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,
94
95
96
                                                           step_size=learning_rate_scheduler_step_size,
                                                           gamma=learning_rate_scheduler_gamma)

97
98
        self.use_last_checkpoint = use_last_checkpoint

99
        experiment_directory_path = os.path.join(
100
            experiment_directory, experiment_name)
101
        self.experiment_directory_path = experiment_directory_path
102

103
104
        self.checkpoint_directory = checkpoint_directory

105
        create_folder(experiment_directory)
106
        create_folder(experiment_directory_path)
107
        create_folder(os.path.join(
108
            experiment_directory_path, self.checkpoint_directory))
109
110
111

        self.start_epoch = 1
        self.start_iteration = 1
112
113
        # self.best_mean_score = 0
        # self.best_mean_score_epoch = 0
114

115
116
117
118
119
        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)
120

121
122
123
        self.EarlyStopping = EarlyStopping()
        self.early_stop = False

124
125
126
        if use_last_checkpoint:
            self.load_checkpoint()

127
128
129
        self.MNI_152_2mm_mask = torch.from_numpy(Image('utils/MNI152_T1_2mm_brain_mask.nii.gz').data)


130
    def train(self, train_loader, validation_loader):
131
132
133
134
135
136
        """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)
137
            validation_loader (class):  Combined dataset and sampler, providing an iterable over the validationing dataset (torch.utils.data.DataLoader)
138
139

        Returns:
140
            trained model
141
142
143
        """

        model, optimizer, learning_rate_scheduler = self.model, self.optimizer, self.learning_rate_scheduler
144
        dataloaders = {'train': train_loader, 'validation': validation_loader}
145
146

        if torch.cuda.is_available():
147
148
            torch.cuda.empty_cache()  # clear memory
            model.cuda(self.device)  # Moving the model to GPU
149
150
151
152
153

        print('****************************************************************')
        print('TRAINING IS STARTING!')
        print('=====================')
        print('Model Name: {}'.format(self.model_name))
154
        if torch.cuda.is_available():
155
156
            print('Device Type: {}'.format(
                torch.cuda.get_device_name(self.device)))
157
158
        else:
            print('Device Type: {}'.format(self.device))
159
160
161
162
163
164
165
166
167
        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))

168
            for phase in ['train', 'validation']:
169
170
171
172
173
174
175
176
177
178
179
                print('-> Phase: {}'.format(phase))

                losses = []

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

                for batch_index, sampled_batch in enumerate(dataloaders[phase]):
                    X = sampled_batch[0].type(torch.FloatTensor)
180
181
182
183

                    # X = ( X - X.min() ) / ( X.max() - X.min() )
                    # X = ( X - X.mean() ) / X.std()

184
                    y = sampled_batch[1].type(torch.FloatTensor)
185

186
                    # We add an extra dimension (~ number of channels) for the 3D convolutions.
187
188
                    X = torch.unsqueeze(X, dim=1)
                    y = torch.unsqueeze(y, dim=1)
189

190
                    if model.test_if_cuda:
191
192
                        X = X.cuda(self.device, non_blocking=True)
                        y = y.cuda(self.device, non_blocking=True)
193
194
195

                    y_hat = model(X)   # Forward pass

196
197
198
199
                    ### Masking goes here
                    y_hat = torch.mul(y_hat, self.MNI_152_2mm_mask)
                    ###

200
                    loss = self.loss_function(y_hat, y)  # Loss computation
201
202

                    if phase == 'train':
203
204
                        optimizer.zero_grad()  # Zero the parameter gradients
                        loss.backward()  # Backward propagation
205
206
207
208
                        optimizer.step()

                        if batch_index % self.loss_log_period == 0:

209
210
                            self.LogWriter.loss_per_iteration(
                                loss.item(), batch_index, iteration)
211

212
213
                        iteration += 1

Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
214
                    losses.append(loss.item())
215
216
217
218
219
220

                    # Clear the memory

                    del X, y, y_hat, loss
                    torch.cuda.empty_cache()

Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
221
                    if phase == 'validation':
222

223
224
225
226
227
228
229
                        if batch_index != len(dataloaders[phase]) - 1:
                            print("#", end='', flush=True)
                        else:
                            print("100%", flush=True)

                with torch.no_grad():

230
231
                    self.LogWriter.loss_per_epoch(losses, phase, epoch)

232
                    if phase == 'validation':
Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
233
234
                        early_stop, save_checkpoint = self.EarlyStopping(
                            np.mean(losses))
235
236
                        self.early_stop = early_stop
                        if save_checkpoint == True:
237
                            validation_loss = np.mean(losses)
238
239
240
241
242
243
244
                            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()
                                                        },
Andrei-Claudiu Roibu's avatar
Andrei-Claudiu Roibu committed
245
246
247
                                                 filename=os.path.join(self.experiment_directory_path, self.checkpoint_directory,
                                                                       'checkpoint_epoch_' + str(epoch) + '.' + checkpoint_extension)
                                                 )
248
249
250
                            # if epoch != self.start_epoch:
                            #     os.remove(os.path.join(self.experiment_directory_path, self.checkpoint_directory,
                            #                            'checkpoint_epoch_' + str(epoch-1) + '.' + checkpoint_extension))
251

252
253
254
                if phase == 'train':
                    learning_rate_scheduler.step()

255
            print("Epoch {}/{} DONE!".format(epoch, self.number_epochs))
256

257
258
259
260
261
262
263
            # Early Stop Condition

            if self.early_stop == True:
                print("ATTENTION!: Training stopped early to prevent overfitting!")
                break
            else:
                continue
264

265
        self.LogWriter.close()
266
267
268
269
270
271
272
273
274

        print('----------------------------------------')
        print('TRAINING IS COMPLETE!')
        print('=====================')
        end_time = datetime.now()
        print('Completed At: {}'.format(end_time))
        print('Training Duration: {}'.format(end_time - start_time))
        print('****************************************************************')

275
276
        return validation_loss

277
278
279
280
281
282
283
284
285
286
    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)
287

288
    def load_checkpoint(self, epoch=None):
289
290
291
292
293
294
295
        """General Checkpoint Loader

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

        Args:
            epoch (int): Current epoch value
        """
296

297
        if epoch is not None:
298
            checkpoint_file_path = os.path.join(
299
                self.experiment_directory_path, self.checkpoint_directory, 'checkpoint_epoch_' + str(epoch) + '.' + checkpoint_extension)
300
301
            self._checkpoint_reader(checkpoint_file_path)
        else:
302
            universal_path = os.path.join(
303
                self.experiment_directory_path, self.checkpoint_directory, '*.' + checkpoint_extension)
304
305
306
307
308
            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:
309
310
                checkpoint_file_path = max(
                    files_in_universal_path, key=os.path.getatime)
311
312
313
                self._checkpoint_reader(checkpoint_file_path)

            else:
314
                self.LogWriter.log("No Checkpoint found at {}".format(
315
                    os.path.join(self.experiment_directory_path, self.checkpoint_directory)))
316

317
318
319
320
321
322
323
324
325
    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
        """

326
327
        self.LogWriter.log(
            "Loading Checkpoint {}".format(checkpoint_file_path))
328
329
330
331

        checkpoint = torch.load(checkpoint_file_path)
        self.start_epoch = checkpoint['epoch']
        self.start_iteration = checkpoint['start_iteration']
332
        # We are not loading the model_name as we might want to pre-train a model and then use it.
333
334
        self.model.load_state_dict = checkpoint['state_dict']
        self.optimizer.load_state_dict = checkpoint['optimizer']
335
        self.learning_rate_scheduler.load_state_dict = checkpoint['scheduler']
336
337

        for state in self.optimizer.state.values():
338
            for key, value in state.items():
339
340
341
                if torch.is_tensor(value):
                    state[key] = value.to(self.device)

342
343
        self.LogWriter.log(
            "Checkpoint Loaded {} - epoch {}".format(checkpoint_file_path, checkpoint['epoch']))