Auto-Encoders

Deep generative models especially Auto Encoders and VAEs in both TensorFlow and PyTorch.

Auto Encoders Linear

Status:

Pytorch model - Done.
Tensorflow model - In progress.

• MNIST dataset.
• Flattened images.
• Refer auto_encoder_linear_torch.ipynb for Pytorch code.
• Refer auto_encoder_linear_TF.py for Tensorflow code.

Model architecture

Insights

• Linear layers in an encoder decoder setup.
• Good for non image data.
• The number of latent space dimensions drastically changes the accuracy of the model.

Results

• Flattened images of MNIST dataset were used.
• A Semi-supervised learning method was demonstrated in the code mentioned above.
• Accuracy could not get beyond 80 percent when 5 - 10 percent of the test data was labeled.
• Convolutional neural networks may work better than these linear networks.
  

---

Convolutional Auto Encoders

Status:

Pytorch Model - Done.
Tensorflow Model - In progress.

• Refer auto_encoders_cnn_torch.py for Pytorch code.
• Refer auto_encoders_cnn_TF.py for Tensorflow code.

Model architecture