Weights & Biases x Reproducibility Challenge

Hey all!

Weights & Biases is supporting the Reproducibility Challenge 2020 and it would be awesome if you guys consider to participate. If you’re interested in ML research, reading and reproducing papers, and writing a report on that, this might be the perfect platform for you to demonstrate all those abilities in a single challenge.

All the details about the initiative can be found on this doc. You can join the Slack workspace using this link. Alternate link (Expires in 30 days).

I am handling the initiative primarily, so you can reach out to me in case of any questions or doubts.
We look forward to seeing your submission for the challenge.

P.S: Weights & Biases is providing $500 upon successful submission to the challenge by completing the checklist provided in the doc linked above to cover your computation costs. Additionally your report will be featured on our Gallery where we have reports on showcase by authors from OpenAI, Stanford, Berkeley and other top notch AI institutions.

We also are going to release a lot of exciting resources for the challenge next week.

Reference discussion

CC: @morgan


We had been working on an example submission for the ML Reproducibility Challenge to help anyone participating in the challenge to structure their submission appropriately. For this, we picked the CVPR 2020 paper - “ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks”. For this we created a Weights & Biases dashboard, two W&B reports and a GitHub repository. We have couple more resources coming soon. Here are the links:

  1. Weights & Biases Dashboard for ECANets
  2. W&B Reproducibility Challenge Report
  3. W&B Report on ECANets
  4. GitHub repository of ECANets reimplementation

P.S - You can export the reports on W&B to the LaTex template of the ML Reproducibility Challenge 2020 by simply clicking on the “Download Report” button followed by the “LaTex” tab and finally from the dropdown select “Reproducibility Challenge” which will download a .zip file containing all the LaTex files that you can submit directly to the workshop.

Instance Segmentation + Bounding Box results of per epoch from a Mask RCNN with a pretrained ECANet-50 backbone on MS-COCO 2017: