Python shape calculations and static adaptation for ONNX export
Resolve an ONNX export blocked by Python shape calculations, preserve pooling semantics and validate a fixed-shape model against PyTorch and ONNX Runtime.
About 6 min readSpecific problems encountered in model training, export and edge deployment, with versions, test conditions and unresolved limits preserved.
Every article is available in full in Chinese and English.
8 articles
Resolve an ONNX export blocked by Python shape calculations, preserve pooling semantics and validate a fixed-shape model against PyTorch and ONNX Runtime.
About 6 min readSeparate AMP skipped updates, non-finite losses and logging failures, and capture the batch, model, optimizer and random state needed to investigate a training NaN.
About 8 min readSet up ONNX Runtime with a standalone QNN Plugin EP on Linux ARM64 and check ELF compatibility, ARM/DSP loading, device discovery and EPContext deployment.
About 10 min readDiagnose low DEIMv2 AP by checking weight loading, backbone preprocessing, EMA selection and a reproducible ONNX export baseline.
About 7 min readExamine DEIMv2 GridSample semantics, CPU and GPU fallback costs, discrete sampling changes and the limits of archived RK3576 and RK3588 measurements.
About 7 min readTrace differences between one-shot and iterative video localization through decoded frames, source timestamps, crop offsets and evaluation on the original timeline.
About 8 min readBuild a Jetson AGX Orin deployment baseline from BSP and CUDA libraries to TensorRT bindings, GPU containers, engine builds and numerical validation.
About 13 min readCheck trainable modules, optimizer state, multimodal inputs and independent evaluation when adapting Qwen3-VL with staged supervised fine-tuning.
About 7 min readNo articles match. Try another term or reset the filters.