Open In App
Related Articles

Residual Networks (ResNet) – Deep Learning

Like Article
Save Article
Report issue

After the first CNN-based architecture (AlexNet) that win the ImageNet 2012 competition, Every subsequent winning architecture uses more layers in a deep neural network to reduce the error rate. This works for less number of layers, but when we increase the number of layers, there is a common problem in deep learning associated with that called the Vanishing/Exploding gradient. This causes the gradient to become 0 or too large. Thus when we increases number of layers, the training and test error rate also increases. 

Comparison of 20-layer vs 56-layer architecture

In the above plot, we can observe that a 56-layer CNN gives more error rate on both training and testing dataset than a 20-layer CNN architecture. After analyzing more on error rate the authors were able to reach conclusion that it is caused by vanishing/exploding gradient. 
ResNet, which was proposed in 2015 by researchers at Microsoft Research introduced a new architecture called Residual Network. 

Residual Network: In order to solve the problem of the vanishing/exploding gradient, this architecture introduced the concept called Residual Blocks. In this network, we use a technique called skip connections. The skip connection connects activations of a  layer to further layers by skipping some layers in between. This forms a residual block. Resnets are made by stacking these residual blocks together. 
The approach behind this network is instead of layers learning the underlying mapping, we allow the network to fit the residual mapping. So, instead of say H(x), initial mapping, let the network fit, 

F(x) := H(x) - x which gives H(x) := F(x) + x

Skip (Shortcut) connection

The advantage of adding this type of skip connection is that if any layer hurt the performance of architecture then it will be skipped by regularization. So, this results in training a very deep neural network without the problems caused by vanishing/exploding gradient.  The authors of the paper experimented on 100-1000 layers of the CIFAR-10 dataset. 
There is a similar approach called “highway networks”, these networks also use skip connection. Similar to LSTM these skip connections also use parametric gates. These gates determine how much information passes through the skip connection. This architecture however has not provided accuracy better than ResNet architecture. 

Network Architecture: This network uses a 34-layer plain network architecture inspired by VGG-19 in which then the shortcut connection is added. These shortcut connections then convert the architecture into a residual network.  

ResNet -34 architecture

Implementation: Using the Tensorflow and Keras API, we can design ResNet architecture (including Residual Blocks) from scratch. Below is the implementation of different ResNet architecture. For this implementation, we use the CIFAR-10 dataset. This dataset contains 60, 000 32×32 color images in 10 different classes (airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks), etc. This dataset can be assessed from keras.datasets API function. 

Step 1: First, we import the keras module and its APIs. These APIs help in building the architecture of the ResNet model.

Code: Importing libraries 

# Import Keras modules and its important APIs
import keras
from keras.layers import Dense, Conv2D, BatchNormalization, Activation
from keras.layers import AveragePooling2D, Input, Flatten
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint, LearningRateScheduler
from keras.callbacks import ReduceLROnPlateau
from keras.preprocessing.image import ImageDataGenerator
from keras.regularizers import l2
from keras import backend as K
from keras.models import Model
from keras.datasets import cifar10
import numpy as np
import os

Step 2: Now, We set different hyper parameters that are required for ResNet architecture. We also did some preprocessing on our dataset to prepare it for training.

Code: Setting Training Hyperparameters  


# Setting Training Hyperparameters
batch_size = 32  # original ResNet paper uses batch_size = 128 for training
epochs = 200
data_augmentation = True
num_classes = 10
# Data Preprocessing 
subtract_pixel_mean = True
n = 3
# Select ResNet Version
version = 1
# Computed depth of 
if version == 1:
    depth = n * 6 + 2
elif version == 2:
    depth = n * 9 + 2
# Model name, depth and version
model_type = 'ResNet % dv % d' % (depth, version)
# Load the CIFAR-10 data.
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
# Input image dimensions.
input_shape = x_train.shape[1:]
# Normalize data.
x_train = x_train.astype('float32') / 255
x_test = x_test.astype('float32') / 255
# If subtract pixel mean is enabled
if subtract_pixel_mean:
    x_train_mean = np.mean(x_train, axis = 0)
    x_train -= x_train_mean
    x_test -= x_train_mean
# Print Training and Test Samples 
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
print('y_train shape:', y_train.shape)
# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)


Step 3: In this step, we set the learning rate according to the number of epochs. As the number of epochs the learning rate must be decreased to ensure better learning.

Code: Setting LR for different numbers of Epochs 


# Setting LR for different number of Epochs
def lr_schedule(epoch):
    lr = 1e-3
    if epoch > 180:
        lr *= 0.5e-3
    elif epoch > 160:
        lr *= 1e-3
    elif epoch > 120:
        lr *= 1e-2
    elif epoch > 80:
        lr *= 1e-1
    print('Learning rate: ', lr)
    return lr


Step 4: Define basic ResNet building block that can be used for defining the ResNet V1 and V2 architecture.

Code: Basic ResNet Building Block 


# Basic ResNet Building Block
def resnet_layer(inputs,
    if conv_first:
        x = conv(x)
        if batch_normalization:
            x = BatchNormalization()(x)
        if activation is not None:
            x = Activation(activation)(x)
        if batch_normalization:
            x = BatchNormalization()(x)
        if activation is not None:
            x = Activation(activation)(x)
        x = conv(x)
    return x


Step 5: Define ResNet V1 architecture that is based on the ResNet building block we defined above:

Code: ResNet V1 architecture 


def resnet_v1(input_shape, depth, num_classes=10):
    if (depth - 2) % 6 != 0:
        raise ValueError('depth should be 6n + 2 (eg 20, 32, 44 in [a])')
    # Start model definition.
    num_filters = 16
    num_res_blocks = int((depth - 2) / 6)
    inputs = Input(shape=input_shape)
    x = resnet_layer(inputs=inputs)
    # Instantiate the stack of residual units
    for stack in range(3):
        for res_block in range(num_res_blocks):
            strides = 1
            if stack & gt
            0 and res_block == 0# first layer but not first stack
                strides = 2  # downsample
            y = resnet_layer(inputs=x,
            y = resnet_layer(inputs=y,
            if stack & gt
            0 and res_block == 0# first layer but not first stack
                # linear projection residual shortcut connection to match
                # changed dims
                x = resnet_layer(inputs=x,
            x = keras.layers.add([x, y])
            x = Activation('relu')(x)
        num_filters *= 2
    # Add classifier on top.
    # v1 does not use BN after last shortcut connection-ReLU
    x = AveragePooling2D(pool_size=8)(x)
    y = Flatten()(x)
    outputs = Dense(num_classes,
    # Instantiate model.
    model = Model(inputs=inputs, outputs=outputs)
    return model


Step 6: Define ResNet V2 architecture that is based on the ResNet building block we defined above:

Code: ResNet V2 architecture 


# ResNet V2 architecture
def resnet_v2(input_shape, depth, num_classes=10):
    if (depth - 2) % 9 != 0:
        raise ValueError('depth should be 9n + 2 (eg 56 or 110 in [b])')
    # Start model definition.
    num_filters_in = 16
    num_res_blocks = int((depth - 2) / 9)
    inputs = Input(shape=input_shape)
    # v2 performs Conv2D with BN-ReLU on input before splitting into 2 paths
    x = resnet_layer(inputs=inputs,
    # Instantiate the stack of residual units
    for stage in range(3):
        for res_block in range(num_res_blocks):
            activation = 'relu'
            batch_normalization = True
            strides = 1
            if stage == 0:
                num_filters_out = num_filters_in * 4
                if res_block == 0# first layer and first stage
                    activation = None
                    batch_normalization = False
                num_filters_out = num_filters_in * 2
                if res_block == 0# first layer but not first stage
                    strides = 2    # downsample
            # bottleneck residual unit
            y = resnet_layer(inputs=x,
            y = resnet_layer(inputs=y,
            y = resnet_layer(inputs=y,
            if res_block == 0:
                # linear projection residual shortcut connection to match
                # changed dims
                x = resnet_layer(inputs=x,
            x = keras.layers.add([x, y])
        num_filters_in = num_filters_out
    # Add classifier on top.
    # v2 has BN-ReLU before Pooling
    x = BatchNormalization()(x)
    x = Activation('relu')(x)
    x = AveragePooling2D(pool_size=8)(x)
    y = Flatten()(x)
    outputs = Dense(num_classes,
    # Instantiate model.
    model = Model(inputs=inputs, outputs=outputs)
    return model


Step 7: The code below is used to train and test the ResNet v1 and v2 architecture we defined above:

Code: Main function 


# Main function 
if version == 2:
    model = resnet_v2(input_shape = input_shape, depth = depth)
    model = resnet_v1(input_shape = input_shape, depth = depth)
model.compile(loss ='categorical_crossentropy',
              optimizer = Adam(learning_rate = lr_schedule(0)),
              metrics =['accuracy'])
# Prepare model saving directory.
save_dir = os.path.join(os.getcwd(), 'saved_models')
model_name = 'cifar10_% s_model.{epoch:03d}.h5' % model_type
if not os.path.isdir(save_dir):
filepath = os.path.join(save_dir, model_name)
# Prepare callbacks for model saving and for learning rate adjustment.
checkpoint = ModelCheckpoint(filepath = filepath,
                             monitor ='val_acc',
                             verbose = 1,
                             save_best_only = True)
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(factor = np.sqrt(0.1),
                               cooldown = 0,
                               patience = 5,
                               min_lr = 0.5e-6)
callbacks = [checkpoint, lr_reducer, lr_scheduler]
# Run training, with or without data augmentation.
if not data_augmentation:
    print('Not using data augmentation.'), y_train,
              batch_size = batch_size,
              epochs = epochs,
              validation_data =(x_test, y_test),
              shuffle = True,
              callbacks = callbacks)
    print('Using real-time data augmentation.')
    # This will do preprocessing and realtime data augmentation:
    datagen = ImageDataGenerator(
        # set input mean to 0 over the dataset
        featurewise_center = False,
        # set each sample mean to 0
        samplewise_center = False,
        # divide inputs by std of dataset
        featurewise_std_normalization = False,
        # divide each input by its std
        samplewise_std_normalization = False,
        # apply ZCA whitening
        zca_whitening = False,
        # epsilon for ZCA whitening
        zca_epsilon = 1e-06,
        # randomly rotate images in the range (deg 0 to 180)
        rotation_range = 0,
        # randomly shift images horizontally
        width_shift_range = 0.1,
        # randomly shift images vertically
        height_shift_range = 0.1,
        # set range for random shear
        shear_range = 0.,
        # set range for random zoom
        zoom_range = 0.,
        # set range for random channel shifts
        channel_shift_range = 0.,
        # set mode for filling points outside the input boundaries
        fill_mode ='nearest',
        # value used for fill_mode = "constant"
        cval = 0.,
        # randomly flip images
        horizontal_flip = True,
        # randomly flip images
        vertical_flip = False,
        # set rescaling factor (applied before any other transformation)
        rescale = None,
        # set function that will be applied on each input
        preprocessing_function = None,
        # image data format, either "channels_first" or "channels_last"
        data_format = None,
        # fraction of images reserved for validation (strictly between 0 and 1)
        validation_split = 0.0)
    # Compute quantities required for featurewise normalization
    # (std, mean, and principal components if ZCA whitening is applied).
    # Fit the model on the batches generated by datagen.flow().
    model.fit_generator(datagen.flow(x_train, y_train, batch_size = batch_size),
                        validation_data =(x_test, y_test),
                        epochs = epochs, verbose = 1, workers = 4,
                        callbacks = callbacks)
# Score trained model.
scores = model.evaluate(x_test, y_test, verbose = 1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])


Results & Conclusion: 
On the ImageNet dataset,  the authors uses a 152-layers ResNet, which is 8 times more deep than VGG19 but still have less parameters. An ensemble of these ResNets generated an error of only 3.7% on ImageNet test set, the result which won ILSVRC 2015 competition. On COCO object detection dataset, it also generates a 28% relative improvement due to its very deep representation. 

Error-rate on ResNet Architecture

  • The result above shows that shortcut connections would be able to solve the problem caused by increasing the layers because as we increase layers from 18 to 34 the error rate on ImageNet Validation Set also decreases unlike the plain network.


top-1 and top-5 Error rate on ImageNet Validation Set.

  • Below are the results on ImageNet Test Set. The 3.57% top-5 error rate of ResNet was the lowest and thus ResNet architecture came first in ImageNet classification challenge in 2015.

Last Updated : 10 Jan, 2023
Like Article
Save Article
Share your thoughts in the comments
Similar Reads