Categorical VAE (using Gumbel-Softmax approximation) in Tensorflow

(Adapted version) Semi-supervised learning part of the Categorical Reparameterization with Gumbel-Softmax
Modifications are list as follows:

1. Batch Norm
2. ConvNet specifications
3. alpha value
4. temperature:
   Eric's: tau = max(0.5, exp(-r*t)), t is step, r = {1e-5, 1e-4}
   Mine: tau = tau0 + (1 - tau0) exp(-r*t), t is epoch, r ~ 2.7e-4
5. (more?)

Semi-supervised learning for MNIST dataset

Classification results on the test set Error rate: ~ 3 - 8% (depending on the configs)
Confusion matrix:

	0	1	2	3	4	5	6	7	8	9
0	969	0	1	0	0	0	3	1	5	1
1	0	1101	10	1	1	3	2	4	12	1
2	1	0	991	3	3	1	2	14	14	3
3	0	0	3	957	0	34	0	4	10	2
4	0	0	2	0	940	0	6	1	1	32
5	3	0	0	1	0	869	2	0	14	3
6	3	1	1	0	2	7	923	0	21	0
7	0	0	7	1	1	0	0	997	3	19
8	3	0	2	4	0	2	2	3	950	8
9	2	2	1	8	1	3	1	5	12	974

Accuracy: 9671/10000 = 96.71%

Labeled images (from the Training set)

Ten images per class.

Style Change (Testing set)

Row: same style across 10 classes. Style is obtained from the diagonal image.

Reconstruction (Testing set)

Reconstruction using the inferred class label. - Left: Input (Ground-truth) - Middle: using inferred dense label. - Right: using inferred discrete label.

Environment and Dependency

• Linux Ubuntu 16.04
• Python 3.5
• Tensorflow 1.3 (before 2017-0912, it was compatible to 0.12)
• Matplotlib
• sklearn
• Numpy
• Json

Usage

pip install -r requirements.txt
git clone https://github.com/JeremyCCHsu/Gumbel-Softmax-VAE-in-tensorflow.git
cd Gumbel-Softmax-VAE-in-tensorflow
mkdir dataset
cd dataset
curl http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz -o train-images-idx3-ubyte.gz
gzip -d train-images-idx3-ubyte.gz
curl http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz -o train-labels-idx1-ubyte.gz
gzip -d train-labels-idx1-ubyte.gz
curl http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz -o t10k-images-idx3-ubyte.gz
gzip -d t10k-images-idx3-ubyte.gz
curl http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz -o t10k-labels-idx1-ubyte.gz
gzip -d t10k-labels-idx1-ubyte.gz
cd ..
python train.py

The outputs will be in ./tmp

Or equivalently, git clone this repo.
Download and unzip MNIST to a sub-folder dataset
Specify your configurations in architecture.json, and execute python train.py