I am pretraining an encoder within a VAE architecture. My goal is to transfer the encoder’s weights and fine-tune it for a classification task. Because my final goal is a classification problem, not generation, is coding “dumb” decoder a general practice?

For example:

        # Dumb decoder
        self.decoder = nn.Sequential(
            nn.Linear(latent_size, 5000),
            nn.Sigmoid()
        )

        # Make the decoder not trainable
        for param in self.decoder.parameters():
            param.requires_grad = False