/RCNN

Recurrent Convolutional Neural Network (RCNN) implementation using python-tensorflow.

Primary LanguagePython

RCNN

Recurrent Convolutional Neural Network (RCNN) implementation using python-tensorflow.

Library versions

python 2.7
numpy 1.12
tensorflow 1.12

Usage

  1. Import models.py

    import models

  2. Create an RCNN object

    rcnn = models.RCNN(batch_size=128, 
                  width=1, # input shape \[batch, width, height, channel\]
                  height=1024, 
                  channel = 126, 
                  conv_num=4, #number of convolution layer
                  rrcl_iter=\[2,2,2,2\], #number of recurrents at each convolution layer. basic convolution if '0'.
                  filters=\[256,256,256,256\], # number of filters at each convolution
                  pool=['p', 'p', 'p', 'p'], #'p' for maxpooling, 'c' for convolution, 'n' for skipping pooling.
                  w_filter_size=[1,9], #convolution filter size
                  p_filter_size=[1,4], #pooling filter size
                  forward_layers=[200, 3], #feedforward layer after conv_num number of convolution layers.
                  keep_probs=[0.5,0.5,0.5,0.5,0.5],  #keep probability of dropout. Use None if you don't want to use dropout.
                  use_batchnorm=True, # batch normalization
                  nonlinearity=tf.nn.relu, 
                  std=0.01, 
                  l_rate=0.01, 
                  l_decay=0.95, #learning rate decay after l_step training.
                  l_step=100000, 
                  scale=1, offset=0.01, epsilon=0.01, 
                  decay=0.9, 
                  momentum=0.9)

  3. Training

    model.train(data=train_input, target=input_target)

train_input should be in shape [batch_size, width, height, channel] and input_target in [batch_size, forward_layers[-1]].
train function returns loss value.

  1. Test
model.test(data=test_input, target=test_target)

train_input should be in shape [batch_size, width, height, channel] and input_target in [batch_size, forward_layers[-1]].
test function returns test input's loss value without updating parameters.

  1. Reconstruct

    model.reconstrut(data=test_input)

This returns the output values.

  1. Save and load the model.
model.save(save_path)
model.load(load_path)
  1. Terminate the model.
model.terminate()

terminate() function closes the session and calls tf.reset_default_graph()