TimesFM #2135
TimesFM #2135
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. |
| def outputs(self) -> Dict[str, Dict[int, str]]: | ||
| return super().outputs |
There was a problem hiding this comment.
I think the error you are facing is because you haven't specified the names + shapes for the expected outputs:
i.e., mean_predictions and full_predictions (you can skip loss)
| @@ -2629,6 +2629,51 @@ class EncoderDecoderOnnxConfig(EncoderDecoderBaseOnnxConfig): | |||
| DEFAULT_ONNX_OPSET = 14 # uses SDPA in Transformers, hence opset>=14. | |||
|
|
|||
|
|
|||
| class TimesFMDummyInputGenerator(DummyInputGenerator): | |||
There was a problem hiding this comment.
could make sense to re-use DummyPatchTSTInputGenerator instead (with some extra modifications), let's wait for huggingface/transformers#34082 to be merged
There was a problem hiding this comment.
@echarlaix one question i had about the huggingface/transformers#34082 in the timesfm_config i do not think my TimesFmOnnxConfig is correct... do you agree?
There was a problem hiding this comment.
Yes it shouldn't inherit from OnnxSeq2SeqConfigWithPast (which should be used for the export of encoder-decoder models). I think you can directly remove TimesFmOnnxConfig in huggingface/transformers#34082 as it shouldn't be used anywhere from what I know (also I'm expecting anything onnx related to be deprecated in v5)
There was a problem hiding this comment.
in any case let me know if I can help @kashif
There was a problem hiding this comment.
ok let me remove it... it would be good to then get it working via transformers.js... let me remove it first and then get back to you!
There was a problem hiding this comment.
sure thanks, also cc @michaelbenayoun @JingyaHuang @IlyasMoutawwakil who can take a look while I'm ooo
|
thanks! |
|
|
||
| @property | ||
| def inputs(self) -> Dict[str, Dict[int, str]]: | ||
| return {"past_values": {0: "batch_size", 1: "sequence_length"}} |
There was a problem hiding this comment.
an issue is remaining before we can merge : it seems that the resulting onnx graph doesn't considered the batch_size as dynamic (won't accept one different than the one given during export) likely coming from :
@kashif any chance we can add the modification directly in the transformers modeling ? otherwise we need to patch the model before export
There was a problem hiding this comment.
Indeed, that does look to be the source of the loss of dynamic shapes. torch.onnx.export doesn't play well with python list operations, so we'd need to refactor it to use tensor operations, if possible.
What does this PR do?
Add support for TimesFM model