#

MNIST Experiment for Batch Normalization

12import torch.nn as nn
13import torch.nn.functional as F
14import torch.utils.data
15
16from labml import experiment
17from labml.configs import option
18from labml_helpers.module import Module
19from labml_nn.experiments.mnist import MNISTConfigs
20from labml_nn.normalization.batch_norm import BatchNorm

#

Model definition

23class Model(Module):

#

28    def __init__(self):
29        super().__init__()

#

Note that we omit the bias parameter

31        self.conv1 = nn.Conv2d(1, 20, 5, 1, bias=False)

#

Batch normalization with 20 channels (output of convolution layer). The input to this layer will have shape [batch_size, 20, height(24), width(24)]

34        self.bn1 = BatchNorm(20)

#

36        self.conv2 = nn.Conv2d(20, 50, 5, 1, bias=False)

#

Batch normalization with 50 channels. The input to this layer will have shape [batch_size, 50, height(8), width(8)]

39        self.bn2 = BatchNorm(50)

#

41        self.fc1 = nn.Linear(4 * 4 * 50, 500, bias=False)

#

Batch normalization with 500 channels (output of fully connected layer). The input to this layer will have shape [batch_size, 500]

44        self.bn3 = BatchNorm(500)

#

46        self.fc2 = nn.Linear(500, 10)

#

48    def forward(self, x: torch.Tensor):
49        x = F.relu(self.bn1(self.conv1(x)))
50        x = F.max_pool2d(x, 2, 2)
51        x = F.relu(self.bn2(self.conv2(x)))
52        x = F.max_pool2d(x, 2, 2)
53        x = x.view(-1, 4 * 4 * 50)
54        x = F.relu(self.bn3(self.fc1(x)))
55        return self.fc2(x)

#

Create model

We use MNISTConfigs configurations and set a new function to calculate the model.

58@option(MNISTConfigs.model)
59def model(c: MNISTConfigs):

#

66    return Model().to(c.device)

#

69def main():

#

Create experiment

71    experiment.create(name='mnist_batch_norm')

#

Create configurations

73    conf = MNISTConfigs()

#

Load configurations

75    experiment.configs(conf, {
76        'optimizer.optimizer': 'Adam',
77        'optimizer.learning_rate': 0.001,
78    })

#

Start the experiment and run the training loop

80    with experiment.start():
81        conf.run()

#

85if __name__ == '__main__':
86    main()