turbo-goggles

Utility repo for my master thesis at the University of Amsterdam. Rest of my implementation is found in machine.

BabyAI treegen

Segmenter to analyze the Baby language from BabyAI.

Required packages:

• gym
• tqdm
• babyai

Inspect CNN

Scripts to do some introspection in trained CNN models in the BabyAI setting.

Required packages:

• PyTorch
• machine

Option Critic

Option-Critic implementation from here, adjusted for my needs.

Required packages:

• numpy
• matplotlib
• scipy

Plotting

GNUplot scripts for visualizing the performance of my models, and a big python plotter to mass create plots from tensorboard progress.

Required packages:

• matplotlib
• numpy

Train diagnostic

Files for training the diagnostic classifier over the hidden state of the ACModel LSTM. Additional functionality include gathering experience for datasets, examining the output of the trained models and online evaluation in the RL setting, as well as zero-shot and transfer learning experiments.

Also contains evaluation script for the diagnostic classifier for the creation of a confusion matrix.

Required packages:

• torch
• numpy
• machine

Utils

Some extra utilities:

• replace_infile.sh loops over (SLURM output) files and tries to replace some text inside each file.
• pull_jobs.sh connects to Lisa, and creates .csv files of tensorboard plots using extract_tb_data.py. Next, they are copied over to be used in plotting.
• average_all.sh averages data over runs. Uses average_data.py to average over .csv files.
• download_model.sh downloads models from Lisa using job ids.
• download_runs.sh downloads tensorboard run files from Lisa using job ids.
• download_logs.sh downloads log result files generated from jobs running on Lisa.