IEEE's Signal Processing Society - Camera Model Identification

Code for Kaggle competition: https://www.kaggle.com/c/sp-society-camera-model-identification

This repo contains code for training and predicting single Resnet50 model which achieves 98% accuracy on private LB.

Run training

Download competition's dataset https://www.kaggle.com/c/sp-society-camera-model-identification/data

You can use more data from flickr to reduce overfitting.

Next, download pretrained weight from https://download.pytorch.org/models/resnet50-19c8e357.pth

Next, call python train.py --train_files train_files --val_files val_files --pretrained_weights_path resnet50-19c8e357.pth --batch_size 128 --model_save_path model.pth

You will need some time to train a model. It takes ~4 hours on a single Tesla M40.

Run prediction

Just call

python predict.py --test_files test_files --batch_size 128 --model_path model.pth --submit_path submit.csv

In the finish, there will be a file submit.csv which you will be able to submit on Kaggle.