Follow
Prime Electronics and Satellitics Inc. 13
Sep 4, 2024 05:25 Increase logging level to extra verbose Sep 3, 2024 05:58 Increase logging level to verbose on onnx to engine script 04:00 trim sample video to keep only important part luigi
02:26 Make 2 distinct ONNX2TRT conversion scripts, one for JetPack 4.6, the other for JetPack 5.1 May 30, 2024 08:37 adjust visualisation style, allow draw_bbox() to show person id May 28, 2024 06:08 Add method to free buffers May 23, 2024 05:24 Implement pycuda backend for inference with TensorRT engine in additon to original polygraphy backend.
the default is polygraphy. You can set TRT_BACKEND to 'PYCUDA' to choose pycuda backend. May 21, 2024 10:36 Add 'get_model_format_and_input_shape' helper function Luigi
06:07 Return and show bounding box confidence May 19, 2024 10:09 Deal with multiple cameras Luigi
May 17, 2024 15:31 Show bounding box on screen too Luigi
15:02 Internalize input frame buffer Luigi
14:15 Unify batched and non-batched versions Luigi
07:26 Show FPS on demo video 07:13 Support Inference over batch with TensorRT Engine Model 06:07 Use dynamic batch size by default 06:00 Add script used to convert ONNX to fp32/fp16/int8/mixed engine 05:50 Use ONNXruntime instead of ONNX.checker.check_model to detect ONNX model as the latter is too strict, it sometimes rejects ONNX models May 16, 2024 10:12 Resize in keeping aspect ratio in visualization 09:51 fix postprocess_batch method in RTMO_GPU_Batch class May 15, 2024 03:35 Add model format (ONNX/Engine) & input size detection (based on file header, not on filename) for RTMO Apr 19, 2024 17:11 Add support TensorRT engine support for RTMO Mar 27, 2024 04:01 Re-quantize models in FP16 in keep positional encoding in FP32 to avoid bad accuracy Mar 26, 2024 05:59 Improve Debug Message Layout in Validate Funciton 05:32 Add appropriate pre- to input image and postprocessing to validate function Mar 25, 2024 10:52 Validate only on pose result, add visual check 09:15 Add comparison between joint coordinated from original and converted models 04:04 Add 'rtol', 'atol' arguments to control tolerance in model validation 02:29 Fix Input Type Error raised in infer() 01:21 Add utility to convert ONNX model in FP32/16 mixed precision Mar 19, 2024 08:37 Load ./libmmdeploy_tensorrt_ops.so for TensorRT EP if available 07:42 Remove fix_batch_dimension_all.sh 07:39 Remove models with fixed batch size as dynamic batch is generally supported Mar 18, 2024 09:41 Apply constant folding 02:52 Reperform shape inference on all models Mar 16, 2024 01:03 Bugfix: correct batch dimension fixation then regenerate ONNX models Luigi
Batch dimension has been fixed properly following the guide below
https://onnxruntime.ai/docs/tutorials/mobile/helpers/make-dynamic-shape-fixed.html Mar 9, 2024 02:13 Update README.md 02:10 Update README.md Mar 8, 2024 08:54 Add option --yolo_nas_pose, used to read YOLO NAS Pose model instead of RTMO 04:12 Fix typo in model names mentioned in demo.sh demo_batch.sh 04:09 Repair DEMO to run with TRT EP 02:37 Fix FP16 models having poor accuracy Keep only dynamic-batch ONNX models, because
it's compatible with TRT EP as well as we've just verified. Mar 7, 2024 09:03 Bugfix for batch demo: add padding if remaning number of frames is not equal to batch size 08:23 Make ONNX models compatible with ONNXruntime's TensorrtExecutionProvider - Use FP16 ONNX Model with Batch Size 4 for DEMO with TRT EP
- Perform shape inference on all ONNX models
- Removed int8 models which is not recommended due to performance drop
- Add ONNX models with fixed input batch size 1,2,4
- Enable TensorRT Execution Provider (TRT EP) Mar 5, 2024 05:54 Use RTMO_GPU_Batch to demo 05:51 Add new class 'RTMO_GPU_Batch' that can perform inference on batch of images 05:13 Optimization for CUDAExecutionProvider - Use IO Binding & Enable Use MAX Workspace to Optimize for RTMO
- Eliminate the dependency on rtmlib Mar 4, 2024 09:13 INT8 models obtained via dynamic quantization Mar 1, 2024 10:21 Update README.md 10:15 Add a sample video