Interpretable CNN

This code implements a tensorflow version of the paper Interpretable Convolutional Neural Networks Quanshi Zhang, Ying Nian Wu, Song-Chun Zhu (https://arxiv.org/abs/1710.00935). And the work is done with Quanshi Zhang while I was an exchange student at UCLA, 2017 fall.

Haiwen's approach

11.20 updates:

• upload the latest file our_cnn_mnist2.py, which includes forward and back propagation.
• upload our_mnist.py, where I modified the label extraction function.

11.4 updates:

• uploaded the demo framework: mask_cnn_mnist.py;
• modified mask_op.py, with forward pass successfully run in the "framework".
• NEXT: combine forward with corresponding bprop function.
• Also: added a numpy tutorial file, in which the "broadcast" part is most useful(starting from In[63])

10.31 updates:

uploaded mask_op, which has the forward pass function changed to be the mask function.

brief intro to the approach:

My approach is to create a new op with its bprop method and put it into the tensorflow, which will enable auto-training and gradient-computing.

I've already succeeded in creating an relu op with its bprop method. In the our_cnn_mnist.py, I changed the activation function of conv1 to our customized tf_relu, and now it can run and get gradients successfully.

Under this approach, our next step is to modify the our_py_func.py, change the demo function into our customized mask and its bprop method. Hopefully we could get it done soon!