This is the submission of the Fellowship.AI Challenge: Cohort 19.
Use a pretrained ResNet 50 and train on the STL10 dataset for 5 epochs and then visualize the layer activations using GRADCAM for a mislabeled image of your weakest class. What additional steps would you take to improve the model without changing the number of epochs?
ResNet 50
- main https://pytorch.org/docs/stable/torchvision/models.html#torchvision.models.resnet50
- tutorial https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html
STL10
- main https://cs.stanford.edu/~acoates/stl10/
- tutorial https://github.com/uoguelph-mlrg/Cutout
- sota https://paperswithcode.com/sota/image-classification-on-stl-10
Please see the details in notebook FellowshipAI_Challenge.ipynb