CapsNet-Tensorflow

A Tensorflow implementation of CapsNet in Hinton's paper Dynamic Routing Between Capsules

States: 1. The code is runnable, but still some different from the paper, not a logical error, but misunstand the capsule, including The routing "for" loop, margin loss(issue #8, thanks for reminding, it's my carelessness) 2. some results of the 'wrong' version has been pasted out, but not effective as the results in the paper,

Daily task 1. Update the code of routing algorithm 2. Adjust margin loss 3. Improve the eval pipeline

Others 1. Here(知乎) is my understanding of the section 4 of the paper (the core part of CapsNet), it might be helpful for understanding the code. 2. If you find out any problems, please let me know. I will try my best to 'kill' it as quickly as possible.

In the day of waiting, be patient: Merry days will come, believe. ---- Alexander PuskinIf :blush:

Chat group:

WeChat: Gitter: my weChat:

• We have a lot of interesting discussion in the WeChat group, welcome to join us. But gitter & English first, please. Anyway, we will release the discussion results in the name of this group(pointing out the contribution of any contributors)

• If you find out that the Wechat group QR is invalid, add my personal account.

Requirements

• Python
• NumPy
• Tensorflow (I'm using 1.3.0, others should work, too)
• tqdm (for showing training progress info)

Usage

Training

Step 1. Clone this repository with git.

$ git clone https://github.com/naturomics/CapsNet-Tensorflow.git
$ cd CapsNet-Tensorflow

Step 2. Download MNIST dataset, mv and extract them into data/mnist directory.(Be careful the backslash appeared around the curly braces when you copy the wget command to your terminal, remove it)

$ mkdir -p data/mnist
$ wget -c -P data/mnist http://yann.lecun.com/exdb/mnist/{train-images-idx3-ubyte.gz,train-labels-idx1-ubyte.gz,t10k-images-idx3-ubyte.gz,t10k-labels-idx1-ubyte.gz}
$ gunzip data/mnist/*.gz

Step 3. Start training with command line:

$ pip install tqdm  # install it if you haven't installed yet
$ python train.py

the tqdm package is not necessary, just a tool for showing the training progress. If you don't want it, change the loop for in step ... to for step in range(num_batch) in train.py

Evaluation

$ python eval.py --is_training False

Results

Results for the 'wrong' version(Issues #8):

• training loss

• test acc |Epoch|49|51| |:----:|:----:|:--:| |test acc|94.69|94.71|

Results after fix Issues #8:

My simple comments for capsule 1. A new version neural unit(vector in vector out, not scalar in scalar out) 2. The routing algorithm is similar to attention mechanism 3. Anyway, a work with great potential, we can do a lot of work on it

TODO:

• Finish the MNIST version of capsNet (progress:90%)
• Do some different experiments for capsNet:
• Using other datasets
• Adjusting model structure
• There is another new paper about capsules(submitted to ICLR 2018), follow-up.