miniplaces-2017

Code for miniplaces challenge 2017.

Original implementation uses ResNet-34 and achieves op-1 accuracy of 50.2% and top-5 accuracy of 78.8%.

For a detailed explanation of the code within, please refer to the project report.

Brief tour of the code base:

• The iPython notebooks Transforms.ipynb and Visualization.ipynb are used to generate visuals for the report
• The folder Resnet Code contains the code used to train and evaluate the model
  • accuracy.py simply calculates the accuracy attained given a target and output vector
  • fine_tuning_config_file.py contains constants (like batch size and learning rate)
  • metrics.py contains code to generate the output file for testing
  • train.py is the meat of the code base, containing code to define and train the model
  • runningAvg.py contains a class to keep track of running averages (for evaluation)
  • test_set.py contains the test set data loader
  • train_set.py contains the train/validation data loader
  • tester.py contains code that given a saved model, produces the output.

Datasets

More information, along with the dataset, can be found at the MiniPlaces Challenge repo.

To run the code

To train the code and generate a checkpoint each time an epoch is finished, type:

python train.py tr 

To produce an output.txt file containing the properly formatted output file for submission, type:

python train.py test '<path_to_checkpoint>'

This code assumes that the data folder is located outside of this directory at /data. Checkpoint files will be saved to ../../checkpoint.pth.tar.

References

Parts of our code were taken from the following tutorials. Proper citation was given in the write-up.

1. Madan, Spandan. "Pytorch Tutorial for Fine Tuning/Transfer Learning a Resnet for Image Classification." https://github.com/Spandan-Madan/Pytorch_fine_tuning_Tutorial
2. PyTorch. "ImageNet Training in Pytorch". https://github.com/pytorch/examples/blob/master/imagenet/main.py
3. ncullen93. "High-Level Training, Data Augmentation, and Utilities for Pytorch." https://github.com/ncullen93/torchsample