DeepDetect Files

Features

• add external pytorch model import skill (3ec64fc)
• add tensor-backed pytorch worker dataset adapter (f8c4ec9)
• added vitpose eval losses (b27951b)
• additions to pytorch backends to accomodate new detection models (02e43ea)
• C++ data augmentation wired to input pull connnector (0d85ac4)
• cli: repository cleansing override (fa0a840)
• hardening of pytorch backend (622da53)
• initial keypoint pose estimator input connector (687ac8b)
• initial pytorch backend structure and messaging layers (c67555c)
• input c++ connector to pytorch backend pull mode (aa6485f)
• map equivalence between python worker and C++ implementations (7bc6cdb)
• model flops estimate (3526835)
• model flops estimate (d4f6148)
• model: initial vitpose with hungarian for slot assignment (b703e02)
• model: regular vitpose with per object keypoints (cec4791)
• model: RT-DETRv4 for pytorch backend (086bacc)
• per-class map in backend and python workers metrics (1b44cff)
• sam2 pytorch worker, inference only (9d340ce)
• sdk: initial python backend sdk (a4a17e6)
• sdk: initial torch/torchvision python backend sdk example (bfd862c)
• torch-only reference detector worker and updated documentation (a20f583)
• training observability mechanism for agents (993145e)
• vit patch_embed resampling from timm in vitpose (7c4eae3)
• vitpose photometric augmentations (f2c30f0)
• vitpose slot loss control and balancing (51eecb1)

Bug Fixes

• add frontmatter to DD skill (ec4b612)
• build with new libs (a29806a)
• graceful exit with pytorch workers (139a060)
• inference vitpose slot keypoint coordinates resizing (8594c2c)
• missing vitpose files (12a859b)
• pytorch backend map (0be18e1)
• remove env name from Visdom plot titles so we can compare runs (725a2ac)
• resume cli when init weights are enabled (31042b3)
• test metric iterations with pytorch workers (e80791e)
• vitpose layer wise decay only when backbone is pretrained (135bb4b)
• wheel building (b903d85)
• wheels tests (def8918)

Docker images

• CPU version: docker pull docker.jolibrain.com/deepdetect_cpu:v0.30.0
• GPU (CUDA only): docker pull docker.jolibrain.com/deepdetect_gpu:v0.30.0
• GPU (CUDA and TensorRT): docker pull docker.jolibrain.com/deepdetect_gpu_tensorrt:v0.30.0
• All images are available from https://docker.jolibrain.com/.

Python wheels

• CPU wheel: python -m pip install --extra-index-url https://www.deepdetect.com/download/wheels/simple deepdetect-cpu==0.30.0
• GPU wheel: python -m pip install --extra-index-url https://www.deepdetect.com/download/wheels/simple deepdetect-gpu==0.30.0
• Wheel indexes: https://www.deepdetect.com/download/wheels/simple/deepdetect-cpu/ and https://www.deepdetect.com/download/wheels/simple/deepdetect-gpu/