A PyTorch repo for Planet: Understanding the Amazon from Space Kaggle Competition.

This repository's main objective is to serve as a starting point to PyTorch. The parts that are covered are:

• A simple data loader that inherits torch.utils.data.Dataset. Adds default transformations from torchvision.transforms.
• Definition of a simple convolutional network and all necessary building blocks for training/validation.
• Loading pretrained models e.g. ResNet
• Visualizing training in tensorboard.
• Predicting on new images.

Requirements

• PyTorch. For installation refer to http://pytorch.org/.
• TensorFlow is required for visualization. CPU version is enough. For installation https://www.tensorflow.org/.
• Tensorboard logger. To install pip install tensorboard_logger. For usage click here.
• Only Python 3+ is supported.

Data

The data we are using are taken from Planet: Understanding the Amazon from Space Kaggle competition. More info can be found at the competition website. In a nutshell we want to label satellite image chips with atmospheric conditions and various classes of land cover/land use. It is a multi-labeling problem with 17 different classes. In the competition algorithms were scored using the mean F2 score.

Here we only use the jpg images. Note that the test-jpg/ and test-additional-jpg/ need to be merged for prediction. The data can still be downloaded here. The csv files don't need to be downloaded as we provide our own for train and validation set.

Training

python train.py

        --model         # specify model, (PlanetSimpleNet, PlanetResNet18)
        --patience      # early stopping
        --batch_size    # batch size
        --nepochs       # max epochs
        --nocuda        # no cuda
        -v              # verbose, extra info during training
        --nworkers      # number of workers
        --seed          # random seed

Models are saved in saved-models/ using the model name and the training run. Note that input images should be at least 224x224 as it is a requirement for pretrained models. Pretrained models are finetuned with a small learning rate.

Logging

Logs are saved in the logs folder. Each model in its own folder with a folder for each run. tensorboard --logdir yourlogdir will start tensorboard. For remote working add the following to .ssh/config for tunneling:

Host name
        HostName hostaddress
        User username
        IdentityFile path_to_id_rsa
        LocalForward 8888 127.0.0.1:6006

Tensorboard can be viewed at 127.0.0.1:8888 in your browser.

Predicting

python predict.py <saved_model>

        --batch_size    # batch size
        --nocuda        # no cuda
        --nworkers      # number of workers
        --output_file   # naming for output csv

Default saved model to load is PlanetSimpleNet. This should give you around 0.87 in the public leaderboard. PlanetResNet18 should give around 0.92.

Progress

• Dataloader with transformations.
• Simple convolutional network.
• Pretrained models.
• Tensorboard integration.
• Predict labels on test images.