Skip to content

DeepDetect v0.24.0
Compare
Choose a tag to compare
@Bycob Bycob released this 31 Mar 09:52
· 56 commits to master since this release
v0.24.0
41bd5c0

Features

  • add custom api path to swagger (4fe0df7)
  • add percent error measure display (1cc15d6)
  • api: add a model_stats field containing the number of parameters of the model (b562fee)
  • api: add labels in service info (66cbff5)
  • api: increase accepted header size (07f6ff3)
  • log model parameters and size at service startup (041b649)
  • regression: add l1 metric for regression (c82f08d)
  • torch: add radam optimizer (5bba045)
  • torch: add translation and bbox duplication to data augmentation (8752e1f)
  • torch: allow data aug to be only noise or distort (5a02234)
  • torch: allow data augmentation w/o db (f5b16b3)
  • torch: data augmentation w/o db for bbox (a99ca7b)
  • torch: set data augmentation factors as requested (e26a775)
  • torch: update torch to 1.13 (9c5da36)
  • trt: add int8 inference (a212a8e)
  • trt: recompile engine if wrong version is detected (0f0bb62)
  • upgrade to TensorRT 8.4.3 (1132760)

Bug Fixes

  • api: re-add parameters in info call (df318cb)
  • raise exception when a bbox file contains invalid classes (3a82a9d)
  • readme: correct docker tags for ci-master (49dde89)
  • regression: fix eucl metric in case of thresholded metric (a006615)
  • take into account false negatives when computing average precision (11905eb)
  • tensorrt: clarify conditions to rebuild engine (9d08b0a)
  • torch: add measures to output event when training not done (5714767)
  • torch: avoid race condition when building alphabet (b1accb7)
  • torch: correctly normalize l1 and l2 metrics in case of multi dim regression (cc9a636)
  • torch: data augmentation handle dummy bboxes correctly (53d0c39)
  • torch: dataset size is half the database size (9541de1)
  • torch: make multi dim regression for images work (00985bf)
  • torch: small glitches in data augmentation (678944a)
  • torch: when reading bbox dataset, also check that the class is not >= nclasses (7b2de88)
  • trace_yolox: bbox shifted by 1 when training yolox (487bad7)
  • trace_yolox: input shape for nonsquare images (6db03be)

Docker images:

  • CPU version: docker pull docker.jolibrain.com/deepdetect_cpu:v0.24.0
  • GPU (CUDA only): docker pull docker.jolibrain.com/deepdetect_gpu:v0.24.0
  • GPU (CUDA and Tensorrt) :docker pull docker.jolibrain.com/deepdetect_cpu_tensorrt:v0.24.0
  • GPU with torch backend: docker pull docker.jolibrain.com/deepdetect_gpu_torch:v0.24.0
  • All images available from https://docker.jolibrain.com/, list images with curl -X GET https://docker.jolibrain.com/v2/_catalog
Assets 2
Loading