This repository contains a TensorFlow implementation of the CNN-MIL combination described in Classifying and Segmenting Microscopy Images with Deep Multiple Instance Learning from Brendan Frey's lab. The unique contribution of this paper is the noisyAND module. The implemented model uses the MNIST dataset for classification.
The architecture presented in the paper involves a CNN, the noisyAND layer, followed by an additional fully connected layer with a joint cross entropy loss.
The joint cross entropy loss is summed over all classes and compares the
output of the MIL layer, P_i, and the output of the additional fully connected layer, y_i,
to the class layer for the entire image, t_i. The authors claim that the
additional fully connect layer will learn dependencies between classes.
- Python 3.5 or greater
- Tensorflow 1.0 or greater
- Latest commit of TensorBase as of this writing: d7be3da.
To run the demo, clone the repo onto your local machine, navigate into the repository,
using the command line/terminal, and type python Conv_MIL_Mnist.py {argparser flags}.
Replace {argparser flags} with the appropriate choice of options as described in
the section below. You will also have to clone TensorBase and place it in
the same folder as this repo.
The following options are available for running the model:
-n, run_num, default = 0, saves all model files under/save_directory/model_directory/Model[n]-e, epochs, default = 1, number of epochs for which to train the model-r, restore, default = 0, binary value indicating whether to restore from a model.-m, model_restore, default = 1, restores from/save_directory/model_directory/Model[n]-f, file_epoch, default = 1, restores model with the following checkpoint:part_[f].ckpt.meta-t, train, default=1, # Binary to train model. 0 = No train.-v, eval, default=1) # Binary to evalulate model. 0 = No eval.-l, learn_rate, default=1e-3, learning rate-g, gpu, default = 0, accepts a single integer or a string "all" to use all available.