resnetv1.py

import keras
from keras.layers import *
from keras.models import Model

# A single resnet module consisting of 1 x 1 conv - 3 x 3 conv and 1 x 1 conv
def resnet_module(x, filters, pool=False):
    res = x
    stride = 1
    if pool:
        stride = 2
        res = Conv2D(filters, kernel_size=1, strides=2, padding="same")(res)
        res = BatchNormalization()(res)

    x = Conv2D(int(filters / 4), kernel_size=1, strides=stride, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = Conv2D(int(filters / 4), kernel_size=3, strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = Conv2D(filters, kernel_size=1, strides=1, padding="same")(x)
    x = BatchNormalization()(x)

    x = add([x, res])
    x = Activation("relu")(x)

    return x


def resnet_first_module(x, filters):
    res = x
    stride = 1
    res = Conv2D(filters, kernel_size=1, strides=1, padding="same")(res)
    res = BatchNormalization()(res)

    x = Conv2D(int(filters / 4), kernel_size=1, strides=stride, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = Conv2D(int(filters / 4), kernel_size=3, strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = Conv2D(filters, kernel_size=1, strides=1, padding="same")(x)
    x = BatchNormalization()(x)

    x = add([x, res])
    x = Activation("relu")(x)

    return x


# A resnet block consisting of N number of resnet modules, first layer has is pooled.
def resnet_block(x, filters, num_layers, pool_first_layer=True):
    for i in range(num_layers):
        pool = False
        if i == 0 and pool_first_layer: pool = True
        x = resnet_module(x, filters=filters, pool=pool)
    return x


# The Resnet model consisting of Conv - block1 - block2 - block3 - block 4 - FC with Softmax
def Resnet(input_shape, num_layers=50, num_classes=10):
    if num_layers not in [50, 101, 152]:
        raise ValueError("Num Layers must be either 50, 101 or 152")

    block_layers = {50: [3, 4, 6, 3],
                    101: [3, 4, 23, 3],
                    152: [3, 8, 36, 3]
                    }
    block_filters = {50: [256, 512, 1024, 2048],
                     101: [256, 512, 1024, 2048],
                     152: [256, 512, 1024, 2048]
                     }
    layers = block_layers[num_layers]
    filters = block_filters[num_layers]
    input = Input(input_shape)

    x = Conv2D(64, kernel_size=7, strides=2, padding="same")(input)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(x)

    x = resnet_first_module(x, filters[0])

    for i in range(4):
        num_filters = filters[i]
        num_layers = layers[i]

        pool_first = True
        if i == 0:
            pool_first = False
            num_layers = num_layers - 1
        x = resnet_block(x, filters=num_filters, num_layers=num_layers, pool_first_layer=pool_first)

    x = GlobalAveragePooling2D()(x)
    x = Dense(num_classes)(x)
    x = Activation("softmax")(x)

    model = Model(inputs=input, outputs=x, name="Resnet{}".format(num_layers))

    return model