A Tensorflow implementation of CapsNet based on Geoffrey Hinton's paper Dynamic Routing Between Capsules
Status:
- The capsule of MNIST version is finished. Now we have two version: a) tag v0.1, not work well due to Issue #8; b) the current version with
test accuracy 99.57
, see details in theResults
section
Daily task
- multi-GPU support
- Improving the reusability of
capsLayer.py
, what you need isimport capsLayer.fully_connected
orimport capsLayer.conv2d
in your code
Others
- Here(η₯δΉ) is an answer explaining my understanding of Section 4 of the paper (the core part of CapsNet). It may be helpful in understanding the code.
- If you find out any problems, please let me know. I will try my best to 'kill' it ASAP.
- Python
- NumPy
- Tensorflow (I'm using 1.3.0, not yet tested for older version)
- tqdm (for displaying training progress info)
- scipy (for saving images)
Step 1.
Clone this repository with git
.
$ git clone https://github.com/naturomics/CapsNet-Tensorflow.git
$ cd CapsNet-Tensorflow
Step 2.
Download the MNIST dataset, mv
and extract it 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 the training:
$ pip install tqdm # install it if you haven't installed yet
$ python main.py
The default parameters of batch size is 128, and epoch is 50. You may need to modify the
config.py
file or use command line parameters to suit your case. In my case, I runpython main.py --test_sum_freq=200 --batch_size=48
for my 4G GPU(~10min/epoch)
- training loss
- test accuracy(the best result is 99.57%)
My simple comments for capsule
- A new version neural unit(vector in vector out, not scalar in scalar out)
- The routing algorithm is similar to attention mechanism
- Anyway, a great potential work, a lot to be built upon
-
Finish the MNIST version of capsNet (progress:90%)
-
Do some different experiments for capsNet:
- Try Using other datasets
- Adjusting the model structure
-
There is another new paper about capsules(submitted to ICLR 2018), a follow-up of the CapsNet paper.
- XifengGuo/CapsNet-Keras: referred for code optimization