Meta Learning

Project on Meta-Learning for the "Advanced Topics on Computational Semantics" course (UvA, 2020)

How To

• Single-task training can be started by running the following line in your Command Line Interface: python3 train.py. Possible command-line options include:
  • batch_size to set the number of elements in a batch
  • random_seed to set a seed for reproducibility
  • epochs to set the number of epochs to run for
  • dataset, e.g. NLI or IBM to set the dataset you want to use (for single task only)
  • More parameters can be found by executing train.py --help or in the file train.py itself.

Similarly Multi-task training, training of Prototypical networks and Proto-MAML can be started with the python scripts multitask.py, prototypes.py and protomaml.py.

Dataset Locations

• .data/multinli for the NLI dataset
  • Downloading is taken care of by PyTorch
• .data/ibm for the IBM dataset
  • Download dataset 3.1 from here.
• .data/MRPC for the MRPC dataset
  • Download dataset from here.
• .data/pdb for the Penn Discourse Bank
  • Contact us for an appropriatly pre-processed version; or
  • Download the raw dataset from here.
• .data/SICK for SICK
  • Download dataset from here.