Feat - gdino extended ops, extended tensorrt execution provider settings

rhysdg / vision-at-a-clip

Low-latency ONNX and TensorRT based zero-shot classification and detection with contrastive language-image pre-training based prompts

16 stars 1 forks source link

Closed rhysdg closed 1 month ago

rhysdg commented 1 month ago

Adding a number of extended tesnorrt execution provider setting
Adding the ability to warmup a model
Adding some extended ops
Despite these changes we're only liking at about a 2x speedup with a FP16 TensorRT engine due to a number of incompatible nodes - working on spitting up BERT as a pre-initialised text encoder shortly