Saoussen-CH / Image-classification-Facebook-PyTorch-Challenge

Image Classification of 102 flower species using ResNet200 pretained model in PyTorch

Image-classification-Facebook-PyTorch-Challenge

Framework and Libraries : PyTorch, Numpy

Visualisation Libraries : matplotlib, seaborn

Platform : Google Colaboratory GPU

Phases :

Data Exploration and preparation :

• 102 Categories of Flowers Dataset : Unbalanced Dataset
• Training Dataset : 6552 images
• Validation Dataset : 817 images
• Test Dataset : 818 images

Techniques :

• Data Augmentation
• Implementation of a Weighted balancing Sampler

Modeling and validation Phase :

Techniques :

• Transfer Learning : Used a Pretained ResNet 200 network architecture trained on ImageNet dataset.
• Replacing ResNet200 classifier with a re-implementation of the Adaptive-average-maxpooling classifier proposed by faster.ai
• Implementation of a training and validation methods
• Used Cross-validation technique -Fine-tuning the model by unfreezing some feature extraction layers weights and retraining for more epochs with new learning rate.

Testing Phase:

• Implementation of a prediction function
• Testing the trained model on the unseen test dataset resulted on a :

Top_1 accuracy : 99,87 %

Top_5 accuracy : 100 %

• Visualization of top_5 class predictions of a new flower image