An Interpretable Bengali Fish Recognizer

• Tech Stack:
• Language: Python
• Libraries: Matplotlib, FastAI, Grad-CAM, HuggingFace
• IDE: VS Code
• Notebook: Google Colab
• Website Integration: Link
• Github URL: Project Link
• Video Demonstration: Link

I collected images for 20 different fishes that mostly get sold in bengali fish markets using fastai’s search and download methods. Then I cleaned the data using fastai vision’s cleaner widgets. I prepared the dataloader with a batch size of 32. DenseNet-121, VGG-19 and ResNet-50 were chosen as the learner models based on my previous research experience. Then I trained them using fastai’s fine_tune() and fit_one_cycle() both methods. ResNet-50 showed the best performance. To find out the important regions for the predictions, I applied Grad-CAM, an XAI approach on all the models. Lastly I deployed the best model to huggingface and created a website integration using github pages.