Intel Scene Classification Competition

Requirements

• fastai==1.0.49 (instructions)
• numpy
• pandas
• pytorch

Code Structure

.
├── intel_submissions/
├── 1-intel_fastai_res50.ipynb
├── 2-intel_fastai_res50_places.ipynb
├── 3-intel_fastai_res50_knn.ipynb
├── 4-intel_fastai_res50_image_filtering_knn.ipynb
├── ensembling_submissions.ipynb
├── final_submission.csv
├── lowest_confidence.csv
└── README.md

Steps to run notebooks

It is assumed that cuda and cudnn are installed. All the experiments were performed on Google Colab which have all the dependencies prinstalled. The data should be placed in directory called data for running the notebooks.

The code experiments were run in the order of numeric prefixes in their filenames. These notebooks were used to generate intemediate models, some of which were used in the final ensemble.

The final submission can be generated by running ensembling_submissions.ipynb which creates final_submission.csv by ensembling submission files located in intel_submissions/ directory.

lowest_confidence.csv contains list of images which were mislabeled or confusing and were filtered at later steps.