Project I | Deep Learning: Image Classification with CNN

Image Classification with CNN

Build a Convolutional Neural Network (CNN) model to classify images from a given dataset into predefined categories/classes.

Task Descriptions and Project Instructions

Project Results

In this project, we processed the CIFAR10 dataset of images and explored different classifiers results:

• Building a sequential CNN model from scratch
• Fine tuned the model and got 84 % of accuracy and 0.47 of loss
• Transfer learning from VGG19 that got us 85 % of accuracy and 0.5 of loss with only 2 unfrozen layers

We deployed a simple gradio demo showing the classification results on unseen images.

Repository Folders and Files

Here is a short description of the folder and files available on the repository.

Documents

Group 4 - Image Classification with CNN - Presentation Slides
Presentation_P1_G4

Saved models

We didn't save all of the models because they were already inferior than the ones we had previously - The winner model is challenger8_fixed.keras or challenger8.pkl

• challenger1.pkl:saved model
• challenger2.pkl:saved model
• challenger3.pkl:saved model
• challenger6.pkl:saved model
• challenger7.pkl:saved model
• challenger8.pkl:saved model
• challenger8_fixed.keras:saved model
• model_02.keras:saved model
• model_03.keras:saved model
• model_04.keras:saved model
• model_05.keras:saved model
• model_06.keras:saved model
• model_07.keras:saved model
• model_08.keras:saved model

Notebooks

• Workflow_Project1_CIFAR10_CNN_G4: Our workflow where we worked to obtain eacg model

• Baseline_ID1: Training a baseline CNN from scratch

• challenger_ID2: Baseline + 1 CV 32 + Dropout

• challenger_ID4: challenger_ID2 + 2 CV 32 + GlobalPooling2D

• challenger_ID5: 2 CV 64 + Dense 256 + GlobalPooling2D

• challenger_ID6: 4 CV 64 + Dense 128 + GlobalPooling2D

• challenger_ID7: 3 CV 64 + Dense 128 + GlobalPooling2D

• challenger_ID8: Challenger_ID4 + Padding + 64 on the middle layers

• challenger_ID9: Conv64x2 + Conv128x2 + BatchNormalization + GlobalPooling2D

• challenger_ID10: challenger_ID8 + TopKAccuracy

• challenger_ID11: challenger_ID10 + SGD + LR=0,05

• challenger_ID12: Conv64x2 + Conv128x2 + BatchNormalization + GlobalPooling2D + SGD+LR=0,05

• challenger_ID13: challenger_ID11 + Adding 64x2 + 128x2 batch 512¶

• challenger_ID14: challenger_ID13 + Padding at the start + batch 512 + 80 epochs

• challenger_ID14_1: challenger_ID13+ Padding at the start + batch 128 + reduce ron plateau + 80 epochs

• challenger_ID15: challenger_ID14_1 + data augmentation horizontal flips = (0,1 , 0,1)

• challenger_ID16: challenger_ID14_1 + data augmentation horizontal flips = (0,025 , 0,025)

• challenger_ID17: challenger_ID14_1 + data augmentation horizontal flips = (0,1 , 0,1)

• challenger_ID18: challenger_ID14_1 + deeper CNN + BN + augmentation + SGD (lr=0.1) + batch 256 + 150 epochs

• challenger_IDVGG19: Transfer of vgg19, freeze and unfreeze the last 2 columns

• Deployment: Deploy a demo of our best model on Gradio

Python Modules

deployment.py: A .py to call deployment from terminal if we want.