MrDavidYu / TF_Convolutional_Autoencoder

Convolutional autoencoder for encoding/decoding RGB images in TensorFlow with high compression ratio

TF_Convolutional_Autoencoder

Convolutional autoencoder for encoding/decoding RGB images in TensorFlow with high compression

This is a sample template adapted from Arash Saber Tehrani's Deep-Convolutional-AutoEncoder tutorial https://github.com/arashsaber/Deep-Convolutional-AutoEncoder for encoding/decoding 3-channel images. The template has been fully commented. I have tested this implementation on rescaled samples from the CelebA dataset from CUHK http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html to produce reasonably decent results from a short period of training. The compression ratio of this implementation is 108. That is, for an input tensor of shape [-1, 48, 48, 3], the bottleneck layer has been reduced to a tensor of shape [-1, 64].

Add on features:

• Takes 3-channel images as input instead of MNIST
• Training now performs checkpoint saves and restores
• Both inputs to the encoder and outputs from the decoder are available for viewing in TensorBoard
• Input autorescaling
• ReLU activation replaced by LeakyReLU to resolve dying ReLU

Caveats:

• It is highly recommended to perform training on the GPU (Took ~40 min to train 20,000 steps on a Tesla K80 for celebG).
• The input size can be increased, but during testing OOM errors occured on the K80 for the input size of 84x84. This will be fixed in a later update. For now if you get any OOM errors in tensor allocation, try to reduce the input size.
• Sample output is currently visibly undersaturated owing to noise introduced by dropout and high model bias. Problems should go away with further training.

Outputs

N.B. The input images are 48x48, hence the blurriness. Additionally these outputs are from setting n_epochs to 1000, which could be increased for even better results (note the cost function trend).

Inputs:
Outputs:

How to run

1. Make sure to create directory ./logs/run1/ to save TensorBoard output. For pushing multiple runs to TensorBoard, simply save additional logs as ./logs/run2/, ./logs/run3/ etc.

2. Unzip ./celebG.tar.gz and save jpegs in ./data/celebG/

3. Either use provided image set or your own. If using your own dataset, I recommend ImageMagick for resizing: https://www.imagemagick.org/script/download.php

4. If using ImageMagick, start Bash in ./data/<your_dir>/:

for file in $PWD/*.jpg
do
convert $file -resize 42x42 $file
done

5. In root dir, python ConvAutoencoder.py

Debug

Here is a list of common problems:

1. The error(cost) is very high (in the thousands or millions): Check that the input images are fetched properly when transforming batch_files to batch_images etc. This high an error is typical of very large natural differences in MSE of input/output and is not caused by a large number of model parameters.

Additional References

Reference https://github.com/carpedm20/DCGAN-tensorflow/blob/master/utils.py for several dynamic image resize functions I have incorporated into my implementation.