camera-trap-anomaly-detection/py/Autoencoder.py

# This is the initial autoencoder architecture.
# Convolutional with 5 conv layers + 1 dense layer per encoder and decoder.
# relu on hidden layers, tanh on output layer
# Number of latent features: 512

from torch import nn

class Autoencoder(nn.Module):
    def __init__(self):
        super(Autoencoder, self).__init__()
        self.encoder = nn.Sequential(
            nn.Conv2d(3, 128, kernel_size=7, stride=4, padding=2),
            nn.ReLU(True),
            nn.Conv2d(128, 64, kernel_size=3, stride=2, padding=1),
            nn.ReLU(True),
            nn.Conv2d(64, 32, kernel_size=3, stride=2, padding=1),
            nn.ReLU(True),
            nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1),
            nn.ReLU(True),
            nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1),
            nn.ReLU(True),
            nn.Flatten(),
            nn.Linear(2048, 512),
            nn.ReLU(True), # see https://stackoverflow.com/questions/50187127/is-it-necessary-to-use-a-linear-bottleneck-layer-for-autoencoder
        )
        self.decoder = nn.Sequential(
            nn.Linear(512, 2048),
            nn.ReLU(True),
            nn.Unflatten(1, (128, 4, 4)),
            nn.ConvTranspose2d(128, 64, kernel_size=4, stride=2, padding=1),
            nn.ReLU(True),
            nn.ConvTranspose2d(64, 32, kernel_size=4, stride=2, padding=1),
            nn.ReLU(True),
            nn.ConvTranspose2d(32, 32, kernel_size=4, stride=2, padding=1),
            nn.ReLU(True),
            nn.ConvTranspose2d(32, 32, kernel_size=4, stride=2, padding=1),
            nn.ReLU(True),
            nn.ConvTranspose2d(32, 3, kernel_size=8, stride=4, padding=2),
            nn.Tanh(),
        )
    
    def forward(self, x):
        x = self.encoder(x)
        x = self.decoder(x)
        return x