70cf20f6cf5e21c3b6165067483b925db61c161c,deepmedic/neuralnet/ops.py,,applyPrelu,#Any#Any#Any#Any#,112

Before Change

    //input is a tensor of shape (batchSize, FMs, r, c, z)
    aPreluValues = np.ones( (numberOfInputChannels), dtype = "float32" )*0.01 //"Delving deep into rectifiers" initializes it like this. LeakyRelus are at 0.01
    aPrelu = theano.shared(value=aPreluValues, borrow=True) //One separate a (activation) per feature map.
    aPreluBroadCastedForMultiplWithChannels = aPrelu.dimshuffle("x", 0, "x", "x", "x")
    
    posTrain = T.maximum(0, inputTrain)
    negTrain = aPreluBroadCastedForMultiplWithChannels * (inputTrain - abs(inputTrain)) * 0.5
    outputTrain = posTrain + negTrain

After Change

    //input is a tensor of shape (batchSize, FMs, r, c, z)
    aPreluValues = np.ones( (numberOfInputChannels), dtype = "float32" ) * 0.01 //"Delving deep into rectifiers" initializes it like this. LeakyRelus are at 0.01
    aPrelu = tf.Variable(aPreluValues, name="aPrelu") //One separate a (activation) per feature map.
    aPrelu5D = tf.reshape(aPrelu, shape=[1, numberOfInputChannels, 1, 1, 1] )
    
    posTrain = tf.maximum(0., inputTrain)
    negTrain = aPrelu5D * (inputTrain - abs(inputTrain)) * 0.5
    outputTrain = posTrain + negTrain

Instances