Skip to content

A Comparative Study of the performance of CNN from scratch compared to a transfer learning approach (InceptionV3)

Notifications You must be signed in to change notification settings

taissirboukrouba/Comparing-CNN-InceptionV3-on-Intel-Image-Classification

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 

Repository files navigation

Comparing-CNN-InceptionV3-on-Intel-Image-Classification

image

This repo is a comparative study of performance between CNN from scratch and using transfer learning :

Data Table Summary

Properties Details
Name Intel Image Classificaiton
Source Kaggle
Author PUNEET BANSAL
Data Type Images
Problem Type Image Classificaiton
Files 3 Folders (14K Training , 3K Testing , 7K Prediction)
Classes 6 Classes (Buildings, Forest, Glacier, Mountain, Sea, Street)

Preprocessing Steps :

  • Image Resizing : Image has been resized to (150,150)
  • Image Normalisation : Normalising pixel values from 0-255 to between 0-1
  • Encoding Image Labels : Transforming labels from categorical format to numerical using Label Encoder

Modelling :

In this part we will go through the different CNN version that we created from scratch and using transfer learning with different versions compare the results including fine-tuning :

  • Simple CNN :
    • 1 convolutional layer (32 Units)
    • 1 MaxPooling layer
    • 2 Dense Layers ( 128 Units , 6 )
    • 10 Epochs
  • Deep CNN :
    • 4 convolutional layers (2 * 32 Units + 64 Units + 128 Units)
    • 4 Maxpooling Layers
    • 3 Dense Layers (128 Units + 64 Units , 6 )
    • 10 Epochs
  • Deep CNN + Fine-tuning :
    • Same previous Architecure
    • Change optimizer to Adamax
    • Added Callbacks (early stopping + LR Scheduler)
    • 20 Epochs
  • InceptionV3 I :
    • Excluding Original Classification layers (include_top = False)
    • Using ImageNet Weights (weights='imagenet')
    • Unfreezing the first 25 layers (inception1.layers[-25:])
    • Adding Costume classificaiton head :
      • Dense layer (1024 units)
      • Dropout layer Dropping 20% of units (0.2)
      • Classification layer (last dense layer) (6)
    • Optimizer Adamax with starting learning rate of 0.0001
    • Callbacks (early stopping + retreive best weights)
    • 20 Epochs
  • InceptionV3 II :
    • Same Previous Architecture
    • Unfreezing the last 25 layers instead (inception2.layers[25:])
    • 20 Epochs
  • InceptionV3 III :
    • Using the full architecture (48 Layers)
    • Different Costume classificaiton head :
      • Batch Normalization Layer
      • 2 Dense Layers (256 Units + 128 Units)
      • 1 Classification layer (6)
    • Same Callbacks
    • 20 Epochs

Results :

I - Simple CNN :

  • Training : The Simple CNN did a horrible job with evaluation of 68% Accuracy & 140% Loss

image

  • Prediciton : The bad training explains the bad predictions done by the simple CNN failing to classify a lot of image

image

II - Deep CNN :

  • Training : The Deep CNN did a better job than the normal CNN with evaluation of 81% Accuracy but still having errors with 82% Loss

image

  • Prediciton : The defective training explains the wrong predictions done by the Deep CNN failing to classify some images

image

III - Deep CNN + Fine-tuning :

  • Training : The Fine-tubned Deep CNN did a better than the previous models with evaluation of 82% Accuracy but still having errors with 49% Loss (still high loss)

image

  • Prediciton : The improvement in the training results can be seen in the predictions done by the fine-tuned CNN

image

IV - InceptionV3 (ver 1) :

  • Training : We can already see the supremacy of transfer learning here with evaluation of 90% Accuracy but still having errors with 27% Loss

image

  • Prediciton : The difference cannot be really seen here , but overall it's doing way better than previous models

image

IV - InceptionV3 (ver 2) :

  • Training : Freezing the first 25 layers insead showed difference with 91% Accuracy but still having errors with 25% Loss

image

  • Prediciton : The difference cannot be really seen here , but overall it's doing way better than previous models

image

IV - InceptionV3 (ver 3) :

  • Training : Using the whole model made the performance inferrior than previous versions of inception with evaluation of 89% Accuracy but having more errors with 30% Loss

image

  • Prediciton : The difference cannot be really seen here , but overall it's doing a bit worse than the other inception versions

image

Comparing Model Performance :

  • The Transfer learning models are doing exponentionally better than the best from-scratch CNN model ( the fine-tuned version)

image

Bonus :

  • This is the confusion matrices of each models giving better context at the overall performances

image

About

A Comparative Study of the performance of CNN from scratch compared to a transfer learning approach (InceptionV3)

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published