Creating CNN image classifier on Stanford Cars dataset using fastai library.

Summary of best model based on experiments in Solution_notebook.ipynb:

Stage 1:

• initialize the model(resnet152 with pretrained weights on ImageNet).
• train final layer
• best learning rate: 0.001
• number of epochs: 16
• train loss: 0.364173
• validation loss: 0.630648
• validation accuracy: 0.810811

Stage 2:

• unfreeze all layers and train.
• best learning rate: 0.001
• number of epochs: 15
• using differential learning rate: [lr/9, lr/6, lr]. The earlier layers was trained on very small learning rate, 0.001/9, the middle layers trained on a bit higher learning rate, 0.001/6, and final layers trained using learning rate, 0.001
• train loss: 0.250652
• validation loss: 0.475406
• validation accuracy: 0.867322

Stage 3:

• resize image to 299, freeze the layers and train the final layer.
• best learning rate: 0.001
• number of epochs: 8
• train loss: 0.213713
• validation loss: 0.444467
• validation accuracy: 0.878378

Final Testing:

• Use Test-Time-Augmentation(TTA). Taking series of different versions of the original image(cropping different areas, changing the zoom) and passing them through the model to get predictions. The average output predictions is then calculated based on different versions of the image and this is given as final output precition for the image
• Accuracy on test set: 0.8827

Testing Steps:

• dependency: PyTorch and fastai library. Install:

conda install -c pytorch -c fastai fastai

or

pip install fastai

• data should be put inside data folder using cars_train and cars_test as folder names.
• run testing.py

Model file can be found here. It's stored in git LFS as the size is about 200 mb.