Caltech200 bird dataset classifier challenge

This is my project for the image classification challenge as part of the image recognition course at MVA.

For more information on the methods I used, please read the file MVA_RecVis_TD3_report.pdf

Codebase structure

• Cropping: cropper.py is used to crop an initial dataset into a cropped, cleaner dataset
• Training and testing:
  • data.py defines the data transforms and augmentations
  • models is the folder containing all model files.
  • model.py defines the model used as Net. In practice, model.py is a symlink to a file in the folder models.
  • main.py is a simple training script with an argparse CLI.
  • evaluate.py is a simple evaluation script featuring an argparse CLI to evaluate on the test dataset.
• Distributing computations:
  • computers_name is the list of the adresses of computers usable for training, one on each line. It is not included here.
  • distrib_train.py is a more advanced distributed training script using Pytorch torch.nn.parallel module and DistributedDataParallel function. It features an argparse CLI, but should be launched using Pytorch's launch.py script. See also launch.sh for launching.
  • launch.sh is a bash script to launch a worker on a specified number of nodes, taking node adresses from computers_name. It spawns a tmux on each node to allow for easily attaching to its shell and easy killing, and logs to the folders logs/stdout and logs/stderr
  • kill.sh does the inverse of launch.sh, it kills workers on a specified number of nodes.
• Visualising features
  • visualise_features.py is a script to visualise the outputs of a neural network on the dataset using PCA and t-SNE. It uses the model Net defined in featurizer.py. I used it to visualise the activations of the layer just before classification, in order to understand better the inner workings of my models.
  • featurizer.py defines the model used for this. Usually a symlink to a model in models.
  • TSNE_embeddings is a folder containing pictures of such embeddings.
• workshop.ipynb is a notebook that I used to quickly try different things.
• MVA_RecVis_TD3_report.pdf is my formal report where I detail the methods I used.