kietzmannlab/blt-pytorch-ccn23

BLT-pytorch repository for our CCN 2023 paper

BLT 🍔 model for pytorch

Based on pytorch 1.12.1 and python 3.10

Training:

• Training files are available in "BLT_train/"
• You can call "train_net.py" with args that can be configured.
• Make sure to define 'dataset_path' in train_net.py
• You can write new models and place them in "models/". Make sure to update loading "get_network_model" function in helper_funcs, and other dependencies.
• The miniecoset dataset can be found at - https://osf.io/msna2/

Current performance:

• On one A100, we can train a BLT RNN with a 10 timestep unroll and a batch size of 1024. It takes ~3hrs for 60 epochs. The final timestep accuracy is 50% on the testplus set of miniecoset-100.

Evaluation:

• Trained and untrained model weights can be found at - https://doi.org/10.17605/OSF.IO/PW238
• Evaluation files are available in "BLT_analyse/"
• Run "extract_actvs.py" to extract the activations for the 'testplus' split of MiniEcoset and readout weights required for further analysis
• Make sure to define 'dataset_path' in extract_actvs.py
• The model files under "models/" are simply modified from the model files in "BLT_train/" to extract the relevant metrics
• Run "analysis.ipynb" to interact with the saved metrics and visualise the results

blt-pytorch-ccn23's Issues

Variable names such as l, o violate Flake8

Hello,

Just a minor issue to bring to your attention in using variable names, while running pre-commit checks Flake8 complaint about 'E741 ambiguous variable name 'l'' coming from two instances in the model definition :

BLT-pytorch-CCN23/BLT_train/models/BLT_net.py

Line 118 in 7e801e0

outputs = [None for l in range(self.timesteps)]

It violates PEP naming conventions.

Cheers,
Varun