BErt-like Neurophysiological Data Representation

Last update: Dec 23, 2022

Related tags

Overview

BENDR

BErt-like Neurophysiological Data Representation

This repository contains the source code for reproducing, or extending the BERT-like self-supervision pre-training for EEG data from the article:

BENDR: using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data

To run these scripts, you will need to use the DN3 project. We will try to keep this updated so that it works with the latest DN3 release. If you are just looking for the BENDR model, and don't need to reproduce the article results per se, BENDR will be (or maybe already is if I forgot to update it here) integrated into DN3, in which case I would start there.

Currently, we recommend version 0.2. Feel free to open an issue if you are having any trouble.

More extensive instructions are upcoming, but in essence you will need to either:

a)  Download the TUEG dataset and pre-train new encoder and contextualizer weights, _or_
b)  Use the [pre-trained model weights](https://github.com/SPOClab-ca/BENDR/releases/tag/v0.1-alpha)

Once you have a pre-trained model:

1) Add the paths of the pre-trained weights to configs/downstream.yml
2) Edit paths to local copies of your datasets in configs/downstream_datasets.yml
3) Run downstream.sh

Comments

about the loss function

Very appreciate for your contribution.i am really interested in the self training in EEG. The only question is about calculating loss function. In your paper, The calculation of the denominator uses cosine similarity between the output of the transformer and the 20 distractors and the input of the transformer. However, in the code, the calculation of the denominator uses cosine similarity between the input of the transformer and the 20 distractors, and the output of the transformer. In other word, the output and the input switch positions. Are both the calculation approaches the same? Or why did you change the calculation approache in the code? Thanks!

opened by stickOverCarrot 2
About deploy downstream.yml and downstream_datasets.yml

Tranks for supplying your code. But when I follow your markdown, I meet some problems

This is my project files

This is my downstream.yml

This is my downstream_datasets.yml

opened by YoloEliwa 1
Pre-trained weights?

Not an issue per se, but you state the pre-trained weights for your paper are available in this repo, yet I have had a good look around and I haven't found them, nor a means of downloading them. Please can you let me know where I could find them? I'm really keen to try out this exciting architecture you've put together!

opened by SgtWhiskeyjack 1
result_tracking module

There's a reference that's in the module import: downstream.py from result_tracking import ThinkerwiseResultTracker that looks like some type of tracking code for experiments?

opened by bencten 1
dropout should change

Iteration: 4%|▍ | 13/330 [00:36<16:00, 3.03s/batches, bac=0.5, Accuracy=0.51, loss=0.695, lr=1.47e-6]D:\Anaconda\envs\LGG\lib\site-packages\torch\nn\functional.py:1338: UserWarning: dropout2d: Received a 3D input to dropout2d and assuming that channel-wise 1D dropout behavior is desired - input is interpreted as shape (N, C, L), where C is the channel dim. This behavior will change in a future release to interpret the input as one without a batch dimension, i.e. shape (C, H, W). To maintain the 1D channel-wise dropout behavior, please switch to using dropout1d instead. warnings.warn("dropout2d: Received a 3D input to dropout2d and assuming that channel-wise "

opened by zy2021314 0
A more detailed explanation

We need to use your code for research, may I ask when you can provide detailed explanation, because we have some difficulties in understanding the code without detailed explanation.

opened by EchizenMike 0
preload in downstream.yml

In the "downstream.yml" file, what is the function of the "preload"? What's mean if I specify "preload: True" or "preload: False"?

Thank you in advance

opened by frannfuri 0

Releases(v0.1-alpha)

v0.1-alpha(Feb 25, 2021)

This release mainly serves so that the weights used for fine-tuning in the article are available for download (and one can avoid pre-training if not needed).
Source code(tar.gz)
Source code(zip)
contextualizer.pt(584.30 MB)
encoder.pt(15.15 MB)

Owner

GitHub Repository

BErt-like Neurophysiological Data Representation

Related tags

Overview

BENDR

Comments

about the loss function

About deploy downstream.yml and downstream_datasets.yml

Pre-trained weights?

result_tracking module

dropout should change

A more detailed explanation

preload in downstream.yml

Releases(v0.1-alpha)

v0.1-alpha(Feb 25, 2021)

Owner

Snakemake workflow for converting FASTQ files to self-contained CRAM files with maximum lossless compression.

Analytical view of olist e-commerce in Brazil

Python Practicum - prepare for your Data Science interview or get a refresher.

Toolchest provides APIs for scientific and bioinformatic data analysis.

BErt-like Neurophysiological Data Representation

Spectacular AI SDK fuses data from cameras and IMU sensors and outputs an accurate 6-degree-of-freedom pose of a device.

pyhsmm MITpyhsmm - Bayesian inference in HSMMs and HMMs. MIT

A Python module for clustering creators of social media content into networks

The micro-framework to create dataframes from functions.

Py-price-monitoring - A Python price monitor

Tokyo 2020 Paralympics, Analytics

Automatic earthquake catalog building workflow: EQTransformer + Siamese EQTransformer + PickNet + REAL + HypoInverse

An Integrated Experimental Platform for time series data anomaly detection.

Python library for creating data pipelines with chain functional programming

MEAD: A Large-scale Audio-visual Dataset for Emotional Talking-face Generation [ECCV2020]

Udacity-api-reporting-pipeline - Udacity api reporting pipeline

X-news - Pipeline data use scrapy, kafka, spark streaming, spark ML and elasticsearch, Kibana

A program that uses an API and a AI model to get info of sotcks

peptides.py is a pure-Python package to compute common descriptors for protein sequences

An Indexer that works out-of-the-box when you have less than 100K stored Documents