Image Classification Model in PyTorch for Indoor Scene Recognition.
This is a Google Colab Notebook. The notebook can also be found here : https://colab.research.google.com/drive/1ALGzKj3cafL0MPUNtbUNXzRc8S_C_lqM
The MIT-67 Indoor Scene Recognition Dataset is used here. The dataset has 15620 images in total distributed amongst 67 classes consiting of airport,trainstation,kitchen,library,etc. Link: http://web.mit.edu/torralba/www/indoor.html
The model used is Resnext101_32x16d. Link: https://github.com/facebookresearch/ResNeXt .
Transfer Learning was used in this task. Pre-trained weights of the model on the ImageNet dataset was used.
Along with transfer learning, data-augmentation, learning rate annealing, early stopping,etc. were also used in the training process.
This model was originally created for the kaggle challenge ( https://www.kaggle.com/c/qstp-deep-learning-2019 ) . The model acheived 3rd rank in the contest amongst 30 participants(Top 1 %).
The notebook contains the link to the weights of the final model, so that it can directly be used without training all over.