solver.py 11.6 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 datetime import datetime
20
21
from utils.losses import MSELoss
from utils.data_utils import create_folder
22
from utils.data_logging_utils import LogWriter
23
from torch.optim import lr_scheduler
24
25
26
27

checkpoint_directory = 'checkpoints'
checkpoint_extension = 'path.tar'

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
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:
53
        trained model - working on this!
54
55

    """
56

57
    def __init__(self,
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
                 model,
                 device,
                 number_of_classes,
                 experiment_name,
                 optimizer=torch.optim.Adam,
                 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',
                 logs_directory='logs'
                 ):
75
76
77

        self.model = model
        self.device = device
78
79
80
81
82
83
84
85
86
87
88
        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
89

90
91
        # 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,
92
93
94
                                                           step_size=learning_rate_scheduler_step_size,
                                                           gamma=learning_rate_scheduler_gamma)

95
96
        self.use_last_checkpoint = use_last_checkpoint

97
        experiment_directory_path = os.path.join(
98
            experiment_directory, experiment_name)
99
        self.experiment_directory_path = experiment_directory_path
100

101
        create_folder(experiment_directory)
102
        create_folder(experiment_directory_path)
103
        create_folder(os.path.join(
104
            experiment_directory_path, checkpoint_directory))
105
106
107

        self.start_epoch = 1
        self.start_iteration = 1
108
109
        # self.best_mean_score = 0
        # self.best_mean_score_epoch = 0
110

111
112
113
114
115
        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)
116

117
118
119
        if use_last_checkpoint:
            self.load_checkpoint()

120
121
122
123
124
125
126
127
128
129
    def train(self, train_loader, test_loader):
        """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)
            test_loader (class):  Combined dataset and sampler, providing an iterable over the testing dataset (torch.utils.data.DataLoader)

        Returns:
130
            trained model
131
132
133
134
135
136
        """

        model, optimizer, learning_rate_scheduler = self.model, self.optimizer, self.learning_rate_scheduler
        dataloaders = {'train': train_loader, 'test': test_loader}

        if torch.cuda.is_available():
137
138
            torch.cuda.empty_cache()  # clear memory
            model.cuda(self.device)  # Moving the model to GPU
139
140
141
142
143

        print('****************************************************************')
        print('TRAINING IS STARTING!')
        print('=====================')
        print('Model Name: {}'.format(self.model_name))
144
        if torch.cuda.is_available():
145
146
            print('Device Type: {}'.format(
                torch.cuda.get_device_name(self.device)))
147
148
        else:
            print('Device Type: {}'.format(self.device))
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
        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))

            for phase in ['train', 'test']:
                print('-> Phase: {}'.format(phase))

                losses = []

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

                for batch_index, sampled_batch in enumerate(dataloaders[phase]):
                    X = sampled_batch[0].type(torch.FloatTensor)
171
                    y = sampled_batch[1].type(torch.FloatTensor)
172

173
                    # We add an extra dimension (~ number of channels) for the 3D convolutions.
174
175
                    X = torch.unsqueeze(X, dim=1)
                    y = torch.unsqueeze(y, dim=1)
176

177
                    if model.test_if_cuda:
178
179
                        X = X.cuda(self.device, non_blocking=True)
                        y = y.cuda(self.device, non_blocking=True)
180
181
182

                    y_hat = model(X)   # Forward pass

183
                    loss = self.loss_function(y_hat, y)  # Loss computation
184
185

                    if phase == 'train':
186
187
                        optimizer.zero_grad()  # Zero the parameter gradients
                        loss.backward()  # Backward propagation
188
189
190
191
                        optimizer.step()

                        if batch_index % self.loss_log_period == 0:

192
193
                            self.LogWriter.loss_per_iteration(
                                loss.item(), batch_index, iteration)
194

195
196
197
                        iteration += 1

                    losses.append(loss.item())
198
199
200
201
202
203
204
205
206
207
208
209
210
211

                    # Clear the memory

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

                    if phase == 'test':
                        if batch_index != len(dataloaders[phase]) - 1:
                            print("#", end='', flush=True)
                        else:
                            print("100%", flush=True)

                with torch.no_grad():

212
213
                    self.LogWriter.loss_per_epoch(losses, phase, epoch)

214
            print("Epoch {}/{} DONE!".format(epoch, self.number_epochs))
215

216
217
218
219
220
221
222
            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()
                                        },
223
224
225
                                 filename=os.path.join(self.experiment_directory_path, checkpoint_directory,
                                                       'checkpoint_epoch_' + str(epoch) + '.' + checkpoint_extension)
                                 )
226

227
        self.LogWriter.close()
228
229
230
231
232
233
234
235
236

        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('****************************************************************')

237
238
239
240
241
242
243
244
245
246
    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)
247

248
    def load_checkpoint(self, epoch=None):
249
250
251
252
253
254
255
        """General Checkpoint Loader

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

        Args:
            epoch (int): Current epoch value
        """
256

257
        if epoch is not None:
258
259
            checkpoint_file_path = os.path.join(
                self.experiment_directory_path, checkpoint_directory, 'checkpoint_epoch_' + str(epoch) + '.' + checkpoint_extension)
260
261
            self._checkpoint_reader(checkpoint_file_path)
        else:
262
263
            universal_path = os.path.join(
                self.experiment_directory_path, checkpoint_directory, '*.' + checkpoint_extension)
264
265
266
267
268
            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:
269
270
                checkpoint_file_path = max(
                    files_in_universal_path, key=os.path.getatime)
271
272
273
                self._checkpoint_reader(checkpoint_file_path)

            else:
274
275
                self.LogWriter.log("No Checkpoint found at {}".format(
                    os.path.join(self.experiment_directory_path, checkpoint_directory)))
276

277
278
279
280
281
282
283
284
285
    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
        """

286
287
        self.LogWriter.log(
            "Loading Checkpoint {}".format(checkpoint_file_path))
288
289
290
291

        checkpoint = torch.load(checkpoint_file_path)
        self.start_epoch = checkpoint['epoch']
        self.start_iteration = checkpoint['start_iteration']
292
        # We are not loading the model_name as we might want to pre-train a model and then use it.
293
294
        self.model.load_state_dict = checkpoint['state_dict']
        self.optimizer.load_state_dict = checkpoint['optimizer']
295
        self.learning_rate_scheduler.load_state_dict = checkpoint['scheduler']
296
297

        for state in self.optimizer.state.values():
298
            for key, value in state.items():
299
300
301
                if torch.is_tensor(value):
                    state[key] = value.to(self.device)

302
303
        self.LogWriter.log(
            "Checkpoint Loaded {} - epoch {}".format(checkpoint_file_path, checkpoint['epoch']))