How to update the parameters βn and vn? about accurate-binary-convolution-network HOT 4 CLOSED

Yes, the β (called betas in the code) and ν (called shift_parameters in the code) are trainable. Both of these are defined as variables and hence are trained when training the complete network. These along with BatchNorm parameters are trained when train_step is computed. The initial values are chosen to be 1 for betas and 0 for shift_parameters (just like it is done in the BatchNorm layer) in order to retain the actual information and these are subsequently trained by the network.

from accurate-binary-convolution-network.

`
def ABCLayer(input_tensor, name=None):

     with tf.name_scope(name, "ABCLayer"):

            # Reshaping betas to match the input tensor
            reshaped_betas = tf.reshape(betas,
                                        shape=[no_ApproxConvLayers] + [1] * len(input_tensor.get_shape()),
                                        name="reshaped_betas")

            # Calculating ApproxConv for each shifted input
            ApproxConv_layers = []
            for index in range(no_ApproxConvLayers):
                # Shifting and binarizing input
                shifted_input = tf.clip_by_value(input_tensor + shift_parameters[index], 0., 1.,
                                                 name="shifted_input_" + str(index))
                binarized_activation = tf.sign(shifted_input - 0.5)
                
                # Passing through the ApproxConv layer
                ApproxConv_layers.append(ApproxConvLayer(binarized_activation))
            ApproxConv_output = tf.convert_to_tensor(ApproxConv_layers, dtype=tf.float32,
                                                     name="ApproxConv_output")
            
            # Taking the weighted sum using the betas
            ABC_output = tf.reduce_sum(tf.multiply(ApproxConv_output, reshaped_betas), axis=0)
            return ABC_output`

As is shown above, shift_parameters and reshaped_betas are computed in the same way with the same data, I wonder if all shift_parameters will be trained to the same value. Or I miss something?

from accurate-binary-convolution-network.

You are probably right. This might be a fatal bug in the code, that I have overlooked. Nonetheless, I would have to test this whether all shift_parameters and betas are learning the same values. If you have already tested it, do share your results.

If it actually is a bug, an easy fix would be to initialize shift_parameters and betas from a gaussian (Glorot initialization should be better) rather than fixed values.

P.S. - Thanks for pointing this out

from accurate-binary-convolution-network.

I'm sorry that I haven't tested it, but I find that the paper gives some advise to initializing shift_parameters and betas in supplementary material. Thank you again for your generous share.

from accurate-binary-convolution-network.