Add ONNX export support for RT-DETR models #1930
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| from transformers.onnx.utils import get_preprocessor | ||
|
|
||
| preprocessor = get_preprocessor(normalized_config._name_or_path) | ||
| if preprocessor is not None and hasattr(preprocessor, "size"): |
There was a problem hiding this comment.
I'm not an Optimum expert, just wanted to ask in what case preprocessor would be None and we'd proceed?
|
@xenova can you add a test? |
|
Done! @fxmarty :) I also set CHW axes to be static (must be what the model was trained on) |
|
static shapes might not be necessary actually: huggingface/transformers#31640 |
|
Hi, may I know that if it's possible to run with onnxruntime instead of using optimum and obtaining model processor using HF pipeline? Also, I tried to export with Cuda but it seems not working. Is there anything I can tweak? Thanks again for the great work! |
What does this PR do?
cc @merveenoyan
Fixes # (issue)
Before submitting
Who can review?