Pytorch implementation of Variational Autoencoder with convolutional encoder/decoder.

https://github.com/Lyusungwon/generative_models_pytorch

Note

  • Divided encoder and decoder

Results

Config

  • model: 180823222128_cvae_5000_200_1e-05_28_28_1_64_16_1
  • epochs 5000 batch-size 200 lr 1e-5 filter-num 64 (32, 128) latent-size 16 L 1

Train

image

  • 32: KLD 27.08 Recon 74.22 Train loss 101.2
  • 64: KLD 26.76 Recon 69.90 Test loss 96.66
  • 128: KLD 27.03 Recon 66.04 Train loss 93.08
  • VAE: KLD 24.13 Recon 76.11 Train loss 100.2

Test

image

  • 32: KLD 26.96 Recon 67.90 Test loss 94.81
  • 64: KLD 26.6 Recon 65.98 Test loss 92.58
  • 128: KLD 26.59 Recon 66.33 Test loss 92.92
  • VAE: KLD 24.08 Recon 69.89 Test loss 93.87

Reconstruction

image

Sampling

image