Reading Codetr source code - part 1 out of 2 скачать в хорошем качестве

Reading Codetr source code - part 1 out of 2

This video dives into the source code of CoDETR - current SOTA object detection model. Part 1: setup, debugging setup, and transformer part of the model - encoder, decoder, deformable attention, one-to-one matching (aka Hungarian) and loss. Part 2 will be about aux heads: region proposal network, ROI heads, ATSS, their use to boost encoder training and aux queries setup to boost decoder training. Important links: Original paper https://arxiv.org/pdf/2211.12860 Source code: https://github.com/Sense-X/Co-DETR Jupyter notebook shown in the video: https://github.com/adensur/blog/blob/... 00:00 - Intro 04:47 - Setup, Dataloaders 11:06 - Transformer part of the model 15:00 - Backbone 20:44 - Positional Encoding 24:32 - Valid ratios and Reference points 30:26 - Encoder 33:35 - Mutli Scale Deformable Attention 43:30 - (Mixed) Query Selection 51:15 - Reference Points for Decoder 52:34 - Decoder code 58:53 - Converting Decoder output to Detections 01:02:26 - Loss, Hungarian Matching 01:10:52 - Outro