Add ONNX Support for Decision Transformer Model

[ra9hur]

Fixes #2032

@IlyasMoutawwakil ,

As per your suggestions, I have made necessary changes to support decision transformer model to ONNX. I ran the below command:

optimum-cli export onnx -m 'edbeeching/decision-transformer-gym-hopper-medium' --task 'reinforcement-learning' --framework 'pt' onnx/

Attaching the terminal output:
output_log.txt

The model was exported successfully with a warning:

The ONNX export succeeded with the warning: The maximum absolute difference between the output of the reference model and the ONNX exported model is not within the set tolerance 1e-05:
- state_preds: max diff = nan
- action_preds: max diff = nan
- return_preds: max diff = nan
- last_hidden_state: max diff = nan.

since max diff is nan, printed the results of the reference and ONNX models. Have attached the results in the below text file.
Note: Have extracted results from output_log.txt file to a separate text file for clarity.
output.txt

Excluding nan, if we closely observe, the difference seems to be well within set tolerance limits.

Please review and let me know your suggestions.

[ra9hur]

Thanks @IlyasMoutawwakil for the detail review comments, will check and re-submit again by next week.

[ra9hur]

@IlyasMoutawwakil ,
I have made necessary changes as per your suggestions.

optimum-cli command runs successfully.

optimum-cli export onnx -m 'edbeeching/decision-transformer-gym-hopper-medium' --task 'reinforcement-learning' --framework 'pt' onnx/

Here is the terminal output for your reference.
success_output_log.txt

Please let me know, if it is good for approval.

[IlyasMoutawwakil]

cool addition ! I committed a couple of suggestions, mainly about keeping the naming of dimensions closer to the config attributes.

[HuggingFaceDocBuilderDev]

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.

[ra9hur]

Thanks @IlyasMoutawwakil , your inputs were very valuable, it was a great learning experience while working on this PR !!

[ra9hur]

@IlyasMoutawwakil ,

Got to test this again and there were a few issues as in the attached log.

OrderedDict_Issue.txt

Further analyzing, changes had to be done to base.py. Output was sensitive to the order of dict elements.
Please review the changes.

[ra9hur]

@IlyasMoutawwakil ,

Can you please review the changes.

[IlyasMoutawwakil]

triggering tests and checking today !

[ra9hur]

Sure, thanks @IlyasMoutawwakil

The branch might be out of sync with the main repo. Please let me know if I should re-sync, copy the PR changes and then commit again.

[IlyasMoutawwakil]

@ra9hur yes please, a rebase or merge would be great ! also a make style for code quality.

[ra9hur]

@IlyasMoutawwakil , I have merged changes in main to this feature branch.

Pardon my ignorance, couldn't quite get what you meant by make style. In one of previous commits, you had made changes for variable names consistency. Is that what you meant ?

Last change that I made was only ordering of outputs in OrderDict in base.py, so nothing major.

Any references so that I can follow ?

[IlyasMoutawwakil]

The export seems to fail due to some outputs mismatch https://github.com/huggingface/optimum/actions/runs/11931260970/job/33256099991?pr=2038#step:5:7057
does it work for you locally ?

Pardon my ignorance, couldn't quite get what you meant by make style

I mean doing pip install -e .[quality] and then make style in terminal, so that ruff and black apply the code quality changes necessary.

[ra9hur]

Thanks for your guidance !!
make style made changes to 3 files, have pushed those changes. Hope, its all good now.

Regarding output mismatch, it is working fine locally.

Validating ONNX model /home/raghu/DL/topics/RL/Godot/onnx/model.onnx...
	-[✓] ONNX model output names match reference model (state_preds, last_hidden_state, return_preds, action_preds)
	-[✓] ONNX model output names match reference model (state_preds, last_hidden_state, return_preds, action_preds)
	- Validating ONNX Model output "state_preds":
	- Validating ONNX Model output "state_preds":
		-[✓] (2, 16, 11) matches (2, 16, 11)
		-[✓] (2, 16, 11) matches (2, 16, 11)
		-[✓] all values close (atol: 1e-05)
		-[✓] all values close (atol: 1e-05)
	- Validating ONNX Model output "action_preds":
	- Validating ONNX Model output "action_preds":
		-[✓] (2, 16, 3) matches (2, 16, 3)
		-[✓] (2, 16, 3) matches (2, 16, 3)
		-[✓] all values close (atol: 1e-05)
		-[✓] all values close (atol: 1e-05)
	- Validating ONNX Model output "return_preds":
	- Validating ONNX Model output "return_preds":
		-[✓] (2, 16, 1) matches (2, 16, 1)
		-[✓] (2, 16, 1) matches (2, 16, 1)
		-[✓] all values close (atol: 1e-05)
		-[✓] all values close (atol: 1e-05)
	- Validating ONNX Model output "last_hidden_state":
	- Validating ONNX Model output "last_hidden_state":
		-[✓] (2, 48, 128) matches (2, 48, 128)
		-[✓] (2, 48, 128) matches (2, 48, 128)
		-[✓] all values close (atol: 1e-05)
		-[✓] all values close (atol: 1e-05)
The ONNX export succeeded and the exported model was saved at: /home/raghu/DL/topics/RL/Godot/onnx

[IlyasMoutawwakil]

@ra9hur the export does work for reinforcement-learning task but not for feature-extraction, because the outputs defined for reinforcement learning are specific to this architecture, and it still expects the common single feature extraction output "feature-extraction": OrderedDict({"last_hidden_state": {0: "batch_size", 1: "sequence_length"}}),.
I tried changing things directly on your fork, moved the outputs to the model config directly and cleaned up some minor things.
I also ran tests locally:

tests/exporters/onnx/test_exporters_onnx_cli.py::OnnxCLIExportTestCase::test_exporters_cli_pytorch_cpu_136_decision_transformer_feature_extraction_default_edbeeching_decision_transformer_gym_hopper_medium PASSED
tests/exporters/onnx/test_exporters_onnx_cli.py::OnnxCLIExportTestCase::test_exporters_cli_pytorch_cpu_137_decision_transformer_no_task_edbeeching_decision_transformer_gym_hopper_medium PASSED
tests/exporters/onnx/test_exporters_onnx_cli.py::OnnxCLIExportTestCase::test_exporters_cli_pytorch_cpu_138_decision_transformer_reinforcement_learning_default_edbeeching_decision_transformer_gym_hopper_medium PASSED

we get 3/3 tests that are passing, for both supported tasks and no_task which infers it automatically.
tell me if this works for you.

[ra9hur]

@IlyasMoutawwakil , I tried your changes locally and executes without errors. Good to go from my side.
Thanks for helping me to troubleshoot and fix those issues !!

Checking new build errors, none of those seem to be specific to this PR.

[ra9hur]

@IlyasMoutawwakil , Tried running pytest on the tests folder locally while referring to this link: https://huggingface.co/docs/transformers/testing

python -m pytest -n auto --dist=loadfile -s -v ./tests/

But, I am not able to exactly reproduce errors as we see here.
Can you please share how to run run these tests locally ?

[IlyasMoutawwakil]

pip install -e .[tests,onnxruntime] then pytest tests/exporters/onnx -s -vvvv -k "decision_transformer" where -k "decision_transformer" is for targeting decision transformer tests

[IlyasMoutawwakil]

the failing test is only related to the missing hf token, PRs from a fork will usually fail that test for security reasons.

[ra9hur]

Thanks for the guidance !!