Grounding DINO For zero-shot object detection go ahead and build from the following example:

[lfreee]

*************** EP Error ***************
EP Error /onnxruntime_src/onnxruntime/python/onnxruntime_pybind_state.cc:754 std::unique_ptronnxruntime::IExecutionProvider onnxruntime::python::CreateExecutionProviderInstance(const onnxruntime::SessionOptions&, const string&, const ProviderOptionsMap&) Invalid TensorRT EP option: trt_engine_hw_compatible
when using [('TensorrtExecutionProvider', {'trt_engine_cache_enable': True, 'trt_max_workspace_size': 4294967296, 'trt_engine_cache_path': '/root/sevnce/liufurui/vision-at-a-clip-main/gdino/data', 'trt_engine_hw_compatible': True, 'trt_sparsity_enable': True, 'trt_build_heuristics_enable': True, 'trt_builder_optimization_level': 0, 'trt_fp16_enable': False}), 'CUDAExecutionProvider', 'CPUExecutionProvider']
Falling back to ['CUDAExecutionProvider', 'CPUExecutionProvider'] and retrying.

[rhysdg]

Hey @lfreee! Apologies for missing this issue here, I've had a pretty busy schedule recently. This was identified recently as a deprecated execution provider argument and I'll send through and update shortly.

Note that I have a big update coming through shortly that outdoes the full onnxruntime approach by pulling out the BERT stage and compillng with a TensorRT backend. End to end I'm getting about 200-300 ms on my RTX 3080.

The remainder of the network remian a bit of a problem with full compilation - The swin Transformer backbone has some unsupported operations etc so I'm still looking into that