leohaipengli/resnet50-ft

I implemented the second problem: "Run validation of the model every few training epochs on validation or test set of the dataset and save the model with the best validation error."

Basically I refactored the orignal imagenet_finetune.py.

List of files:

• config.py: contains the config of fine tuning
• dataloader.py: a wrapper class that allows easy cifar-10 data loading
• imagenet_finetune.py: contains the logic of fine tuning
• inference.py: contains convenient functions to do validation on train/test data set for given model
• resnet50cifar.py: contains the definition of the NN model

Instructions:

To do fine tuning with validation on train/test set and saving models, just run

python imagenet_finetune.py

It will output logs about errors and save the models including the best one. Finally it will output which epoch gives you the best model.

Result:

I plot a graph showing loss & accuracy of different epochs.