Intel Scene Classification Challenge

Author: Yash Bhalgat | Rank 3rd Public Leaderboard | Rank 6th Private Leaderboard

Competition Link. You can get the data from here: Drive link

Requirements:

• Python3.6
• pytorch==1.0.x
• torchvision==0.2.2
• albumentations==0.1.12
• pretrainedmodels==0.7.4
• fastai==0.7.0
• numpy==1.15.4
• matplotlib==2.2.3
• PIL==5.1.0
• tqdm==4.25.0
• pickle==4.0

Data

When you download the data, you must organize the images into three folders: train, valid and test. The train and valid folders must have subfolders corresponding to the class names. The final directory structure for the imgs folder should look like:

imgs
├── test
├── train
│   ├── buildings
│   ├── forest
│   ├── glacier
│   ├── mountain
│   ├── sea
│   └── street
└── valid
    ├── buildings
    ├── forest
    ├── glacier
    ├── mountain
    ├── sea
    └── street

To extract the test images, you can simply parse the test_WyRytb0.csv file.

I chose a 80%-20% train-validation split. That means, you should transfer 20% of the files from each subfolder of the train directory to the valid directory. To do so, you can use the following command appropriately:

shuf -n <num_files> -e train/<class_name>/* | xargs -i mv {} valid/<class_name>/

Substitute num_files with the number of files (20%) you want to move and class_name with one of buildings forest glacier mountain sea street.

Training

There are two scripts which can be used to train the models: train_evaluate_scene_classification.py and fastai_full.py. The first file purely uses torch and torchvision. The second file exploits the abstractions provided by the fastai library to train the models. You can go through the code for the details of the implementation. The details can also be found in the submitted report.

To train any model using train_evaluate_scene_classification.py, you might have to edit the lines 208-238. These lines basically load the pretrained weights and replace the last fully-connected layer to accomodate these 6 classes.

For example, while using the xception network, the last layer is replaced as follows:

num_ftrs = model_ft.last_linear.in_features
model_ft.last_linear = nn.Linear(num_ftrs, NUM_CLASSES)

Hyper-parameters

You can edit the file fine_tuning_config_file.py to modify the hyperparameters, as per your usage.

Evaluation

For evaluation on the test-set, the scripts dump_output.py and create_submission.py are useful.

• Once a model is trained, dump_output.py runs the model on the test images and saves the output logits (output of the fully connected layer) to a dump_<model_name>.pkl file. These dump files are useful while performing ensembling of different trained models.
• During ensembling, we average the logit outputs of different models and then use them to get the predicted labels. In case of usine a single model, we just use the logit outputs (which are already dumped/saved) to compute the predicted labels. This is done using the create_submission.py file.
• test.py basically combines these two scripts' functionalities

Miscellaneous Scripts

1. create_features.py - There were some other experiments I performed, like using a SVM classifier on top of a trained ResNet (or any other network) used as a featurizer. You can use this script to explore this method further. :)
2. check_validation.py - You can use this script to inspect which validation images are being misclassified.
3. patchwise.py - This was my attempt to use a patchwise classifier (more like the recent BagNet paper). Didn't work so well.

Thank you. For any questions about this implementation, feel free to reach out to me at [email protected]