README.md

metadata

license: mit
datasets:
  - PolyAI/banking77
language:
  - en
tags:
  - autoencoder

VAE trained on Banking 77 Open Intent Classification Dataset

This is a Variational Autoencoder (VAE) trained on the PolyAI/banking77 dataset.

Architecture

• input_dim: 768
• hidden_dim: 256
• latent_dim: 64

Encoder

The encoder maps the input to a latent space distribution.

encoder = nn.Sequential(
            nn.Linear(input_dim, hidden_dim),
            nn.ReLU()
        )

mu = nn.Linear(hidden_dim, latent_dim)
logvar = nn.Linear(hidden_dim, latent_dim)

Decoder

The decoder reconstructs the input from a sample of the latent space.

decoder = nn.Sequential(
            nn.Linear(latent_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, input_dim)
        )

Metrics

The model was trained and evaluated using the following metrics:

1. Training set: VAE Loss
   • 50% reconstruction loss between original input vs reconstructed output
   • 50% KL divergence between Latent Z vs standard normal distribution
2. Validation set: 100% reconstruction loss -> used to find the best model (with the lowest reconstruction loss)