ResNet downsampling @ ConvNets.ipynb

[raytroop]

Thanks for sharing. But it seems that downsampling should be performed by stride-2 convolutions in the 3×3 layer of the first block in each stage. This way, number of parameter may be reduced.

def resnet(input_shape, n_classes):

    def conv_bn_rl(x, f, k=1, s=1, p='same'):
        x = Conv2D(f, k, strides=s, padding=p)(x)
        x = BatchNormalization()(x)
        x = ReLU()(x)
        return x

    def conv_block(tensor, f1, f2, s):
#         x = conv_bn_rl(tensor, f1, s=s)
#         x = conv_bn_rl(x, f1, 3)
        x = conv_bn_rl(tensor, f1)
        x = conv_bn_rl(x, f1, 3, s=s)
        x = Conv2D(f2, 1)(x)
        x = BatchNormalization()(x)

        shortcut = Conv2D(f2, 1, strides=s, padding='same')(tensor)
        shortcut = BatchNormalization()(shortcut)

        x = add([shortcut, x])
        output = ReLU()(x)

        return output

according to Aggregated Residual Transformations for Deep Neural Networks

[dkatsios]

Thanks for pointing this out. I couldn't find this information at the original paper but indeed it makes more sense like this.