Further results

• cae_32x32x32_zero_pad_bin trained on 121,827 randomly selected frames from YouTube dataset.
  • num_epochs: 2, batch_size: 2, learning_rate 1e-05, save_weights_every: 500 (batches), show_out_every: 100 (batches)
• cae_32x32x32_zero_pad_bin trained on a smaller, random subset of YouTube dataset (2286 frames).
  • num_epochs: 5, batch_size: 1, learning_rate: 1e-05

Results

3k optimization steps @ epoch 1

75k optimization steps @ epoch 1

140k optimization steps @ epoch 2

212k optimization steps @ epoch 2

277k optimization steps @ epoch 3

349k optimization steps @ epoch 3

414k optimization steps @ epoch 4

486k optimization steps @ epoch 4

551k optimization steps @ epoch 5

623k optimization steps @ epoch 5

662k optimization steps @ epoch 5

683k optimization steps @ epoch 5

Note: Some slight artifacts can be seen where patches connect. This is what smoothing.py fixes.