ZeroPadding in the vailla_encoder

[bobvo23]

It seems to me that before many convolution layers, the Zeropadding 2d is applied to keep the original dimension of the input. I think one leaner way to implement is to strip the ZeroPadding layers and leveraging "same" padding param of the Conv2D layers.

It is not only easier to read, but also more concise. Please let me know your thoughts?

I can make a PR for this issue if you think it is possible.

From this

    x = (ZeroPadding2D((pad, pad), data_format=IMAGE_ORDERING))(x)
    x = (Conv2D(filter_size, (kernel, kernel),
                data_format=IMAGE_ORDERING, padding='valid'))(x)

to this

    x = (Conv2D(filter_size, (kernel, kernel),
                data_format=IMAGE_ORDERING, padding='same'))(x)

def vanilla_encoder(input_height=224,  input_width=224):

    kernel = 3
    filter_size = 64
    pad = 1
    pool_size = 2

    if IMAGE_ORDERING == 'channels_first':
        img_input = Input(shape=(3, input_height, input_width))
    elif IMAGE_ORDERING == 'channels_last':
        img_input = Input(shape=(input_height, input_width, 3))

    x = img_input
    levels = []
    # from 224 to 226, padding 1 zero at each dimention
    x = (ZeroPadding2D((pad, pad), data_format=IMAGE_ORDERING))(x)
    x = (Conv2D(filter_size, (kernel, kernel),
                data_format=IMAGE_ORDERING, padding='valid'))(x)
    x = (BatchNormalization())(x)
    x = (Activation('relu'))(x)
    x = (MaxPooling2D((pool_size, pool_size), data_format=IMAGE_ORDERING))(x)
    levels.append(x)

    x = (ZeroPadding2D((pad, pad), data_format=IMAGE_ORDERING))(x)
    x = (Conv2D(128, (kernel, kernel), data_format=IMAGE_ORDERING,
                padding='valid'))(x)
    x = (BatchNormalization())(x)
    x = (Activation('relu'))(x)
    x = (MaxPooling2D((pool_size, pool_size), data_format=IMAGE_ORDERING))(x)
    levels.append(x)

    for _ in range(3):
        x = (ZeroPadding2D((pad, pad), data_format=IMAGE_ORDERING))(x)
        x = (Conv2D(256, (kernel, kernel),
                    data_format=IMAGE_ORDERING, padding='valid'))(x)
        x = (BatchNormalization())(x)
        x = (Activation('relu'))(x)
        x = (MaxPooling2D((pool_size, pool_size),
                          data_format=IMAGE_ORDERING))(x)
        levels.append(x)

    return img_input, levels