GoogleNet Architecture Implementation in Keras with CIFAR-10 Dataset

Introduction

In this article, we will take a look into the GoogleNet architecture which is a Deep Learning based state-of-the-art image classification model. Then we will do an implementation of a minimalistic version of GoogleNet in Keras by using the CIFAR-10 dataset for the training purpose.

What is GoogleNet?

Developed by the Google research team, GoogleNet is a 22 layer deep, deep convolutional network for image classification. This model was the winner of ILSRVRC 2014 with an error rate of 6.67%. It achieved a top-5 93.3% accuracy on the ImageNet dataset which was significantly high at that time.

• Also Read – 7 Popular Image Classification Models in ImageNet Challenge (ILSVRC) Competition History

Architecture of GoogleNet

The model architecture is compact compared to other models like Alexnet, VGG, and Resnet. The main difference in this architecture is that it does not use multiple dense layers but instead employs pooling layers with small filters. This in turn while maintaining the depth of the neural network greatly decreases the computation required.

The new wave of deep learning image classification models has seen the usage of network-in-network modules. This essentially means the usage of a predefined collection of layers (in the form of a function) which is already tried and tested for enhancing training or making the model efficient. This is what makes the GoogleNet model so versatile. It uses a network-in-network module called the Inception Module.

The original model takes images of size 224x224x3 as input, has filters ranging from 1×1 to 5×5, ReLU activation for convolutional layers, and dropout layers for regularization.

(We will be making our own version of the original model so that it can be trained on the CIFAR-10 dataset.)

The original model representation can be seen below:

Inception Module

The general idea behind the inception module is to create an architecture where the input can be passed through different types of layers at once. In order to extract distinct features parallelly and finally concatenate them later. This is done so that the model can learn both local and abstract features which in turn enhances model performance.

In the actual model proposed in the paper, the inception module branches into four distinct paths.

1. The first path learns local features using a convolutional layer with 1×1 filters
2. The second path first applies 1×1 convolutions for dimensionality reduction. In order to prepare the input to be passed through 3×3 convolutions.
3. The third path is the same as the second one. The only difference is that this time we use 5×5 convolutions. Both the second and the third branches are tasked with learning the general features in images.
4. This is known as a pool projection branch. It applied 3×3 max pooling before learning features using a 1×1 convolutional layer.

These branches apply operations on the same input(same in value, not the same instance) parallelly and are later concatenated. In order to ensure that the concatenation of outputs can be performed, the same padding is used across the module.

CIFAR-10 Dataset

For our GoogleNet implementation in Keras, we will be using the CIFAR-10 dataset to train the model.

CIFAR-10 dataset is a famous computer vision dataset that contains 60,000 images of size 32×32. It has a total of 10 distinct classes namely airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks. It is used as a benchmark dataset for image-based computer vision machine learning models.

• Also Read – 12 Amazing Computer Vision Datasets You Should Know

GoogleNet Implementation in Keras

We will be implementing the below-optimized architecture of GoogleNet so that it can be fit to the CIFAR-10 dataset. (To view the below image properly you can right click and save it to your system and then view in full size)

i) Setting up Google Colab

Google Colab is a free GPU supported online Notebook service provided to us by Google. It is a great alternative if you don’t have GPU in your workstation.

In order to make sure Google Colab is using the GPU, go to the ‘runtime’ section –> ‘change runtime type’ and choose the hardware accelerator to GPU.

ii) Importing the Libraries

We will start by importing all the libraries mentioned below:

from keras.layers.normalization import BatchNormalization
from keras.layers.convolutional import Conv2D
from keras.layers.convolutional import AveragePooling2D
from keras.layers.convolutional import MaxPooling2D
from keras.layers.core import Activation
from keras.layers.core import Dropout
from keras.layers.core import Dense
from keras.layers import Flatten
from keras.layers import Input
from keras.models import Model
from keras.layers import concatenate
import matplotlib
matplotlib.use("Agg")
%matplotlib inline
from sklearn.preprocessing import LabelBinarizer
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import LearningRateScheduler
from keras.optimizers import SGD
from keras.datasets import cifar10
import numpy as np

iii) Convolutional Module Implementation

We construct the ‘conv_module’ which is a series of convolutional and a batch normalization layer that is ultimately passed through a Relu activation.

def conv_module(input,No_of_filters,filtersizeX,filtersizeY,stride,chanDim,padding="same"):
  input = Conv2D(No_of_filters,(filtersizeX,filtersizeY),strides=stride,padding=padding)(input)
  input = BatchNormalization(axis=chanDim)(input)
  input = Activation("relu")(input)
  return input

Arguments:

1. Input: Input to be processed
2. No of filters: No of filters that should be in the Conv2D layer.
3. FilterX and FilterY: Size of the filters.
4. Stride
5. Channel dimension
6. Padding: Predefined to ‘same’ for the whole model.

iv) Inception Module Implementation

We define our modified inception module by using two conv_modules. The first module is initialized with:

1. 1×1 filters(used to learn local features in images).
2. The second and third with 3×3 and 5×5 respectively(responsible for learning general features).
3. Define the pool projection layer using a global pooling layer.
4. After that, we concatenate the layer outputs along the channel dimension(chanDim).

def inception_module(input,numK1x1,numK3x3,numk5x5,numPoolProj,chanDim):
                                 #Step 1
  conv_1x1 = conv_module(input, numK1x1, 1, 1,(1, 1), chanDim) 
                                 #Step 2
  conv_3x3 = conv_module(input, numK3x3, 3, 3,(1, 1), chanDim)
  conv_5x5 = conv_module(input, numk5x5, 5, 5,(1, 1), chanDim)
                                 #Step 3
  pool_proj = MaxPooling2D((3, 3), strides=(1, 1), padding='same')(input)
  pool_proj = Conv2D(numPoolProj, (1, 1), padding='same', activation='relu')(pool_proj)
                                 #Step 4
  input = concatenate([conv_1x1, conv_3x3, conv_5x5, pool_proj], axis=chanDim)
  return input

v) Downsample Module Implementation

As we could see in the architecture which is supposed to be built above the model is very deep. Thus it requires downsampling so that the no of trainable parameters can be controlled. For that function, we use a downsample_module. It is basically the concatenation of the output of a max-pooling layer and a conv_module(with 3×3 filters).

def downsample_module(input,No_of_filters,chanDim):
  conv_3x3=conv_module(input,No_of_filters,3,3,(2,2),chanDim,padding="valid")
  pool = MaxPooling2D((3,3),strides=(2,2))(input)
  input = concatenate([conv_3x3,pool],axis=chanDim)
  return input

vi) Model Implementation

1. Define an input layer with the width, height, and depth parameters of the function.
2. Use two inception modules along with a downsampling module.
3. Use Five inception modules along with a downsampling module.
4. Use two inception modules with pooling and dropout.
5. Flatten layer, with a dense layer(with no of units equal to no of classes) along with an activation layer with softmax classifier.

def MiniGoogleNet(width,height,depth,classes):
  inputShape=(height,width,depth)
  chanDim=-1

  # (Step 1) Define the model input
  inputs = Input(shape=inputShape)

  # First CONV module
  x = conv_module(inputs, 96, 3, 3, (1, 1),chanDim)

  # (Step 2) Two Inception modules followed by a downsample module
  x = inception_module(x, 32, 32,32,32,chanDim)
  x = inception_module(x, 32, 48, 48,32,chanDim)
  x = downsample_module(x, 80, chanDim)
  
  # (Step 3) Five Inception modules followed by a downsample module
  x = inception_module(x, 112, 48, 32, 48,chanDim)
  x = inception_module(x, 96, 64, 32,32,chanDim)
  x = inception_module(x, 80, 80, 32,32,chanDim)
  x = inception_module(x, 48, 96, 32,32,chanDim)
  x = inception_module(x, 112, 48, 32, 48,chanDim)
  x = downsample_module(x, 96, chanDim)

  # (Step 4) Two Inception modules followed
  x = inception_module(x, 176, 160,96,96, chanDim)
  x = inception_module(x, 176, 160, 96,96,chanDim)
  
  # Global POOL and dropout
  x = AveragePooling2D((7, 7))(x)
  x = Dropout(0.5)(x)

  # (Step 5) Softmax classifier
  x = Flatten()(x)
  x = Dense(classes)(x)
  x = Activation("softmax")(x)

  # Create the model
  model = Model(inputs, x, name="googlenet")
  return model

You can look at the model architecture by creating a model instance and using the summary function with it.

Output:

Model: "googlenet" 
__________________________________________________________________________________________________ 
Layer (type)                    Output Shape         Param #  Connected to 
================================================================================================== 
input_1 (InputLayer)            [(None, 32, 32, 3)]  0 
__________________________________________________________________________________________________ 
conv2d (Conv2D)                 (None, 32, 32, 96)   2688     input_1[0][0] 
__________________________________________________________________________________________________ 
batch_normalization (BatchNorma (None, 32, 32, 96)   384      conv2d[0][0] 
__________________________________________________________________________________________________ 
activation (Activation)         (None, 32, 32, 96)   0        batch_normalization[0][0] 
__________________________________________________________________________________________________ 
conv2d_1 (Conv2D)               (None, 32, 32, 32)   3104     activation[0][0] 
__________________________________________________________________________________________________ 
conv2d_2 (Conv2D)               (None, 32, 32, 32)   27680    activation[0][0] 
__________________________________________________________________________________________________ 
conv2d_3 (Conv2D)               (None, 32, 32, 32)   76832    activation[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_1 (BatchNor (None, 32, 32, 32)   128      conv2d_1[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_2 (BatchNor (None, 32, 32, 32)   128      conv2d_2[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_3 (BatchNor (None, 32, 32, 32)   128      conv2d_3[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d (MaxPooling2D)    (None, 32, 32, 96)   0        activation[0][0] 
__________________________________________________________________________________________________ 
activation_1 (Activation)       (None, 32, 32, 32)   0        batch_normalization_1[0][0] 
__________________________________________________________________________________________________ 
activation_2 (Activation)       (None, 32, 32, 32)   0        batch_normalization_2[0][0] 
__________________________________________________________________________________________________ 
activation_3 (Activation)       (None, 32, 32, 32)   0        batch_normalization_3[0][0] 
__________________________________________________________________________________________________ 
conv2d_4 (Conv2D)               (None, 32, 32, 32)   3104     max_pooling2d[0][0] 
__________________________________________________________________________________________________ 
concatenate (Concatenate)       (None, 32, 32, 128)  0        activation_1[0][0] activation_2[0][0] activation_3[0][0] conv2d_4[0][0] 
__________________________________________________________________________________________________ 
conv2d_5 (Conv2D)               (None, 32, 32, 32)   4128     concatenate[0][0] 
__________________________________________________________________________________________________ 
conv2d_6 (Conv2D)               (None, 32, 32, 48)   55344    concatenate[0][0] 
__________________________________________________________________________________________________ 
conv2d_7 (Conv2D)               (None, 32, 32, 48)   153648   concatenate[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_4 (BatchNor (None, 32, 32, 32)   128      conv2d_5[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_5 (BatchNor (None, 32, 32, 48)   192      conv2d_6[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_6 (BatchNor (None, 32, 32, 48)   192      conv2d_7[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_1 (MaxPooling2D)  (None, 32, 32, 128)  0        concatenate[0][0] 
__________________________________________________________________________________________________ 
activation_4 (Activation)       (None, 32, 32, 32)   0        batch_normalization_4[0][0] 
__________________________________________________________________________________________________ 
activation_5 (Activation)       (None, 32, 32, 48)   0        batch_normalization_5[0][0] 
__________________________________________________________________________________________________ 
activation_6 (Activation)       (None, 32, 32, 48)   0        batch_normalization_6[0][0] 
__________________________________________________________________________________________________ 
conv2d_8 (Conv2D)               (None, 32, 32, 32)   4128     max_pooling2d_1[0][0] 
__________________________________________________________________________________________________ 
concatenate_1 (Concatenate)     (None, 32, 32, 160)  0        activation_4[0][0] activation_5[0][0] activation_6[0][0] conv2d_8[0][0] 
__________________________________________________________________________________________________ 
conv2d_9 (Conv2D)               (None, 15, 15, 80)   115280   concatenate_1[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_7 (BatchNor (None, 15, 15, 80)   320      conv2d_9[0][0] 
__________________________________________________________________________________________________ 
activation_7 (Activation)       (None, 15, 15, 80)   0        batch_normalization_7[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_2 (MaxPooling2D)  (None, 15, 15, 160)  0        concatenate_1[0][0] 
__________________________________________________________________________________________________ 
concatenate_2 (Concatenate)     (None, 15, 15, 240)  0        activation_7[0][0] max_pooling2d_2[0][0] 
__________________________________________________________________________________________________ 
conv2d_10 (Conv2D)              (None, 15, 15, 112)  26992    concatenate_2[0][0] 
__________________________________________________________________________________________________ 
conv2d_11 (Conv2D)              (None, 15, 15, 48)   103728   concatenate_2[0][0] 
__________________________________________________________________________________________________ 
conv2d_12 (Conv2D)              (None, 15, 15, 32)   192032   concatenate_2[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_8 (BatchNor (None, 15, 15, 112)  448      conv2d_10[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_9 (BatchNor (None, 15, 15, 48)   192      conv2d_11[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_10 (BatchNo (None, 15, 15, 32)   128      conv2d_12[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_3 (MaxPooling2D)  (None, 15, 15, 240)  0        concatenate_2[0][0] 
__________________________________________________________________________________________________ 
activation_8 (Activation)       (None, 15, 15, 112)  0        batch_normalization_8[0][0] 
__________________________________________________________________________________________________ 
activation_9 (Activation)       (None, 15, 15, 48)   0        batch_normalization_9[0][0] 
__________________________________________________________________________________________________ 
activation_10 (Activation)      (None, 15, 15, 32)   0        batch_normalization_10[0][0] 
__________________________________________________________________________________________________ 
conv2d_13 (Conv2D)              (None, 15, 15, 48)   11568    max_pooling2d_3[0][0] 
__________________________________________________________________________________________________ 
concatenate_3 (Concatenate)     (None, 15, 15, 240)  0        activation_8[0][0] activation_9[0][0] activation_10[0][0] conv2d_13[0][0] 
__________________________________________________________________________________________________ 
conv2d_14 (Conv2D)              (None, 15, 15, 96)   23136    concatenate_3[0][0] 
__________________________________________________________________________________________________ 
conv2d_15 (Conv2D)              (None, 15, 15, 64)   138304   concatenate_3[0][0] 
__________________________________________________________________________________________________ 
conv2d_16 (Conv2D)              (None, 15, 15, 32)   192032   concatenate_3[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_11 (BatchNo (None, 15, 15, 96)   384      conv2d_14[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_12 (BatchNo (None, 15, 15, 64)   256      conv2d_15[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_13 (BatchNo (None, 15, 15, 32)   128      conv2d_16[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_4 (MaxPooling2D)  (None, 15, 15, 240)  0        concatenate_3[0][0] 
__________________________________________________________________________________________________ 
activation_11 (Activation)      (None, 15, 15, 96)   0        batch_normalization_11[0][0] 
__________________________________________________________________________________________________ 
activation_12 (Activation)      (None, 15, 15, 64)   0        batch_normalization_12[0][0] 
__________________________________________________________________________________________________ 
activation_13 (Activation)      (None, 15, 15, 32)   0        batch_normalization_13[0][0] 
__________________________________________________________________________________________________ 
conv2d_17 (Conv2D)              (None, 15, 15, 32)   7712     max_pooling2d_4[0][0] 
__________________________________________________________________________________________________ 
concatenate_4 (Concatenate)     (None, 15, 15, 224)  0        activation_11[0][0] activation_12[0][0] activation_13[0][0] conv2d_17[0][0] 
__________________________________________________________________________________________________ 
conv2d_18 (Conv2D)              (None, 15, 15, 80)   18000    concatenate_4[0][0] 
__________________________________________________________________________________________________ 
conv2d_19 (Conv2D)              (None, 15, 15, 80)   161360   concatenate_4[0][0] 
__________________________________________________________________________________________________ 
conv2d_20 (Conv2D)              (None, 15, 15, 32)   179232   concatenate_4[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_14 (BatchNo (None, 15, 15, 80)   320      conv2d_18[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_15 (BatchNo (None, 15, 15, 80)   320      conv2d_19[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_16 (BatchNo (None, 15, 15, 32)   128      conv2d_20[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_5 (MaxPooling2D)  (None, 15, 15, 224)  0        concatenate_4[0][0] 
__________________________________________________________________________________________________ 
activation_14 (Activation)      (None, 15, 15, 80)   0        batch_normalization_14[0][0] 
__________________________________________________________________________________________________ 
activation_15 (Activation)      (None, 15, 15, 80)   0        batch_normalization_15[0][0] 
__________________________________________________________________________________________________ 
activation_16 (Activation)      (None, 15, 15, 32)   0        batch_normalization_16[0][0] 
__________________________________________________________________________________________________ 
conv2d_21 (Conv2D)              (None, 15, 15, 32)   7200     max_pooling2d_5[0][0] 
__________________________________________________________________________________________________ 
concatenate_5 (Concatenate)     (None, 15, 15, 224)  0        activation_14[0][0] activation_15[0][0] activation_16[0][0] conv2d_21[0][0] 
__________________________________________________________________________________________________ 
conv2d_22 (Conv2D)              (None, 15, 15, 48)   10800    concatenate_5[0][0] 
__________________________________________________________________________________________________ 
conv2d_23 (Conv2D)              (None, 15, 15, 96)   193632   concatenate_5[0][0] 
__________________________________________________________________________________________________ 
conv2d_24 (Conv2D)              (None, 15, 15, 32)   179232   concatenate_5[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_17 (BatchNo (None, 15, 15, 48)   192      conv2d_22[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_18 (BatchNo (None, 15, 15, 96)   384      conv2d_23[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_19 (BatchNo (None, 15, 15, 32)   128      conv2d_24[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_6 (MaxPooling2D)  (None, 15, 15, 224)  0        concatenate_5[0][0] 
__________________________________________________________________________________________________ 
activation_17 (Activation)      (None, 15, 15, 48)   0        batch_normalization_17[0][0] 
__________________________________________________________________________________________________ 
activation_18 (Activation)      (None, 15, 15, 96)   0        batch_normalization_18[0][0] 
__________________________________________________________________________________________________ 
activation_19 (Activation)      (None, 15, 15, 32)   0        batch_normalization_19[0][0] 
__________________________________________________________________________________________________ 
conv2d_25 (Conv2D)              (None, 15, 15, 32)   7200     max_pooling2d_6[0][0] 
__________________________________________________________________________________________________ 
concatenate_6 (Concatenate)     (None, 15, 15, 208)  0        activation_17[0][0] activation_18[0][0] activation_19[0][0] conv2d_25[0][0] 
__________________________________________________________________________________________________ 
conv2d_26 (Conv2D)              (None, 15, 15, 112)  23408    concatenate_6[0][0] 
__________________________________________________________________________________________________ 
conv2d_27 (Conv2D)              (None, 15, 15, 48)   89904    concatenate_6[0][0] 
__________________________________________________________________________________________________ 
conv2d_28 (Conv2D)              (None, 15, 15, 32)   166432   concatenate_6[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_20 (BatchNo (None, 15, 15, 112)  448      conv2d_26[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_21 (BatchNo (None, 15, 15, 48)   192      conv2d_27[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_22 (BatchNo (None, 15, 15, 32)   128      conv2d_28[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_7 (MaxPooling2D)  (None, 15, 15, 208)  0        concatenate_6[0][0] 
__________________________________________________________________________________________________ 
activation_20 (Activation)      (None, 15, 15, 112)  0        batch_normalization_20[0][0] 
__________________________________________________________________________________________________ 
activation_21 (Activation)      (None, 15, 15, 48)   0        batch_normalization_21[0][0] 
__________________________________________________________________________________________________ 
activation_22 (Activation)      (None, 15, 15, 32)   0        batch_normalization_22[0][0] 
__________________________________________________________________________________________________ 
conv2d_29 (Conv2D)              (None, 15, 15, 48)   10032    max_pooling2d_7[0][0] 
__________________________________________________________________________________________________ 
concatenate_7 (Concatenate)     (None, 15, 15, 240)  0        activation_20[0][0] activation_21[0][0] activation_22[0][0] conv2d_29[0][0] 
__________________________________________________________________________________________________ 
conv2d_30 (Conv2D)              (None, 7, 7, 96)     207456   concatenate_7[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_23 (BatchNo (None, 7, 7, 96)     384      conv2d_30[0][0] 
__________________________________________________________________________________________________ 
activation_23 (Activation)      (None, 7, 7, 96)     0        batch_normalization_23[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_8 (MaxPooling2D)  (None, 7, 7, 240)    0        concatenate_7[0][0] 
__________________________________________________________________________________________________ 
concatenate_8 (Concatenate)     (None, 7, 7, 336)    0        activation_23[0][0] max_pooling2d_8[0][0] 
__________________________________________________________________________________________________ 
conv2d_31 (Conv2D)              (None, 7, 7, 176)    59312    concatenate_8[0][0] 
__________________________________________________________________________________________________ 
conv2d_32 (Conv2D)              (None, 7, 7, 160)    484000   concatenate_8[0][0] 
__________________________________________________________________________________________________ 
conv2d_33 (Conv2D)              (None, 7, 7, 96)     806496   concatenate_8[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_24 (BatchNo (None, 7, 7, 176)    704      conv2d_31[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_25 (BatchNo (None, 7, 7, 160)    640      conv2d_32[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_26 (BatchNo (None, 7, 7, 96)     384      conv2d_33[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_9 (MaxPooling2D)  (None, 7, 7, 336)    0        concatenate_8[0][0] 
__________________________________________________________________________________________________ 
activation_24 (Activation)      (None, 7, 7, 176)    0        batch_normalization_24[0][0] 
__________________________________________________________________________________________________ 
activation_25 (Activation)      (None, 7, 7, 160)    0        batch_normalization_25[0][0] 
__________________________________________________________________________________________________ 
activation_26 (Activation)      (None, 7, 7, 96)     0        batch_normalization_26[0][0] 
__________________________________________________________________________________________________ 
conv2d_34 (Conv2D)              (None, 7, 7, 96)     32352    max_pooling2d_9[0][0] 
__________________________________________________________________________________________________ 
concatenate_9 (Concatenate)     (None, 7, 7, 528)    0        activation_24[0][0] activation_25[0][0] activation_26[0][0] conv2d_34[0][0] 
__________________________________________________________________________________________________ 
conv2d_35 (Conv2D)              (None, 7, 7, 176)    93104    concatenate_9[0][0] 
__________________________________________________________________________________________________ 
conv2d_36 (Conv2D)              (None, 7, 7, 160)    760480   concatenate_9[0][0] 
__________________________________________________________________________________________________ 
conv2d_37 (Conv2D)              (None, 7, 7, 96)     1267296  concatenate_9[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_27 (BatchNo (None, 7, 7, 176)    704      conv2d_35[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_28 (BatchNo (None, 7, 7, 160)    640      conv2d_36[0][0] 
__________________________________________________________________________________________________ 
batch_normalization_29 (BatchNo (None, 7, 7, 96)     384      conv2d_37[0][0] 
__________________________________________________________________________________________________ 
max_pooling2d_10 (MaxPooling2D) (None, 7, 7, 528)    0        concatenate_9[0][0] 
__________________________________________________________________________________________________ 
activation_27 (Activation)      (None, 7, 7, 176)    0        batch_normalization_27[0][0] 
__________________________________________________________________________________________________ 
activation_28 (Activation)      (None, 7, 7, 160)    0        batch_normalization_28[0][0] 
__________________________________________________________________________________________________ 
activation_29 (Activation)      (None, 7, 7, 96)     0        batch_normalization_29[0][0] 
__________________________________________________________________________________________________ 
conv2d_38 (Conv2D)              (None, 7, 7, 96)     50784    max_pooling2d_10[0][0] 
__________________________________________________________________________________________________ 
concatenate_10 (Concatenate)    (None, 7, 7, 528)    0        activation_27[0][0] activation_28[0][0] activation_29[0][0] conv2d_38[0][0] 
__________________________________________________________________________________________________ 
average_pooling2d (AveragePooli (None, 1, 1, 528)    0        concatenate_10[0][0] 
__________________________________________________________________________________________________ 
dropout (Dropout)               (None, 1, 1, 528)    0        average_pooling2d[0][0] 
__________________________________________________________________________________________________ 
flatten (Flatten)               (None, 528)          0        dropout[0][0] 
__________________________________________________________________________________________________ 
dense (Dense)                   (None, 10)           5290     flatten[0][0] 
__________________________________________________________________________________________________ 
activation_30 (Activation)      (None, 10)           0        dense[0][0] 
================================================================================================== 
Total params: 5,963,658 
Trainable params: 5,959,050 
Non-trainable params: 4,608 
__________________________________________________________________________________________________ 
None

vii) Decay Parameter

Initialize the number of epochs and the initial learning rate. Next, create a polynomial decay function that will compute a dynamic learning rate. It will change/decay with every epoch in a polynomial manner.

NUM_EPOCHS = 50
INIT_LR = 5e-3
def poly_decay(epoch):
  maxEpochs = NUM_EPOCHS
  baseLR = INIT_LR
  power = 1.0
  alpha = baseLR * (1 - (epoch / float(maxEpochs))) ** power
  return alpha

viii) Preparing the Data

Load the CIFAR dataset as arrays in the float data type.

1. Compute the mean of each train and test dataset and subtract it. So that the pixel values are normalized.
2. Use the label binarizer to convert the labels from integers to vectors.
3. Use the image data generator to create variation in data.

((trainX, trainY), (testX, testY)) = cifar10.load_data()
trainX = trainX.astype("float")
testX = testX.astype("float")
                                # Step 1
mean = np.mean(trainX, axis=0)
trainX -= mean
testX -= mean
                                # Step 2
lb = LabelBinarizer()
trainY = lb.fit_transform(trainY)
testY = lb.transform(testY)
                                # Step 3
aug = ImageDataGenerator(width_shift_range=0.1,height_shift_range=0.1, horizontal_flip=True,fill_mode="nearest")

ix) Compiling and Fitting the Model

Define the callbacks and optimizers for the model. Create a model instance with arguments of width, height, depth, and classes(32,32,3,10).

1. Compile the model with categorical cross-entropy loss and set the metric to accuracy.
2. Fit the model with the data generator, validation data, steps per epoch, no of epochs, and callbacks.

callbacks=[LearningRateScheduler(poly_decay)]
opt = SGD(lr=INIT_LR, momentum=0.9)
model = GoogleNet(width=32, height=32, depth=3, classes=10)
                                    # Step 1
model.compile(loss="categorical_crossentropy", optimizer=opt,metrics=["accuracy"])
                                    # Step 2
model.fit(aug.flow(trainX, trainY, batch_size=64),validation_data=(testX, testY), steps_per_epoch=len(trainX) // 64,epochs=NUM_EPOCHS, callbacks=callbacks, verbose=1)

Initial epoch output:

Final epoch output:

# Final epoch
Epoch 50/50 781/781 [==============================] - 103s 132ms/step - loss: 0.0211 - accuracy: 0.9939 - val_loss: 0.3649 - val_accuracy: 0.9085

x) Evaluating the Model

Use the below code to checks the test accuracy. It comes out to be approximately 90 percent. (Top 20 for the CIFAR-10 dataset on kaggle)

score=model.evaluate(testX,testY)
print('Test Score=',score[0])
print('Test Accuracy=',score[1])

Output:

Test Score: 0.3751
Test Accuracy: 0.9077

Inference

Below are some of the inference results by using our trained GoogleNet model.

Conclusion

Coming to the end of this tutorial, where we did GoogleNet implementation in Keras. We did the training using the CIFAR-10 dataset, but you may use your own datsets fpr this purpose.

References:

• For more details on GoogleNet architecture, you may check the official paper by Christian Szegedy et al Going Deeper with convolutions (2015)
• Special regards to Adrian Rosebrock for providing an extended look at the GoogleNet model.