Deep Neural Networks, Homework 1

Recognizing and counting shapes.

Running experiments

python src/run.py configuration.yaml

Configuration

We can specify various training scenarios via YAML files. This allows us to experiment with different network architectures and hyperparameters.

The configuration should have the following fields:

• type: one of classification, counting-small, counting-large, depending on which problem we're attempting to train for. This controls:
  • the size of the final output layer
  • final transformation applied to output layer
  • transformations applied to targets, if necessary
  • the loss function used
  • the accuracy metric used
• batch_norm (defaults to false)
• dropout (defaults to false): false or p parameter for dropout layers
• lr
• epochs
• batch_size
• augmentations: list of augmentation specifications to expand the dataset with:
  • rotation (defaults to 0): integer 0-3 specifying number of 90° rotations to apply
  • vflip (defaults to false)
  • hflip (defaults to false)
• convolutions: a list of convolution layer specifications:
  • in_channels
  • out_channels
  • size (defaults to 3)
  • stride (defaults to 1)
  • padding (defaults to 1)
  • max_pool (defaults to false): whether to apply a 2x2 max pool layer
• linears: a list of desnse layer specifications:
  • in_features
  • out_features

Results

For each run, a timestamped directory is created under outputs/ with:

• A copy of the used configuration file
• A CSV file with train set losses, test set losses, and test set accuracies, for each epoch
• losses.png: a plot of the losses
• accuracies.png: a plot of the accuracies
• net.pt: weights of the trained model