Add ONNX config for RT-DETR (and RT-DETRv2) #2201

qubvel · 2025-02-27T18:28:13Z

What does this PR do?

Add ONNX config for RT-DETR (and RT-DETRv2), continue of

[RT-Detr] - Add RT-Detr Onnx Config #2040

Also, FP16 ONNX tests are broken for RT-DETR but will be fixed in the coming version of transformers. What would be a better way to skip them?

PR in transformers:

Fix fp16 ONNX export for RT-DETR and RT-DETRv2 transformers#36460

Who can review?

@echarlaix

HuggingFaceDocBuilderDev · 2025-02-27T18:45:43Z

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.

xenova

Thanks!

I tested the exporter on the set of rtdetr_v2 models and these are the warnings. Most are not issues, but some may be, and could require some modifications to huggingface/transformers#36460 (review).

/usr/local/lib/python3.11/dist-packages/optimum/exporters/onnx/model_configs.py:2668: UserWarning: Exporting model with image `height=64` which is less than minimal 320, setting `height` to 320.
  warnings.warn(
/usr/local/lib/python3.11/dist-packages/optimum/exporters/onnx/model_configs.py:2673: UserWarning: Exporting model with image `width=64` which is less than minimal 320, setting `width` to 320.
  warnings.warn(
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr/modeling_rt_detr_resnet.py:107: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if num_channels != self.num_channels:
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:989: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  grid_w = torch.arange(int(width), device=device).to(dtype)
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:990: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  grid_h = torch.arange(int(height), device=device).to(dtype)
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:300: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if attn_weights.size() != (batch_size * self.num_heads, target_len, source_len):
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:336: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if attn_output.size() != (batch_size * self.num_heads, target_len, self.head_dim):
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:1747: TracerWarning: torch.as_tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  spatial_shapes = torch.as_tensor(spatial_shapes_list, dtype=torch.long, device=source_flatten.device)
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:1638: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  valid_wh = torch.tensor([width, height], device=device).to(dtype)
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:1647: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  anchors = torch.where(valid_mask, anchors, torch.tensor(torch.finfo(dtype).max, dtype=dtype, device=device))
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:191: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if not is_torchdynamo_compiling() and (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() != sequence_length:
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:212: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if reference_points.shape[-1] == 2:
/usr/local/lib/python3.11/dist-packages/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py:218: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  elif reference_points.shape[-1] == 4:

The if checks should be able to be ignored, since most are error checking. The only possibly concerning ones are:

grid_w = torch.arange(int(width), device=device).to(dtype) - tracing issues with int()
grid_h = torch.arange(int(height), device=device).to(dtype) - same issue as 1
spatial_shapes = torch.as_tensor(spatial_shapes_list, dtype=torch.long, device=source_flatten.device) - since the values use during export will be used in the final model (i.e., not dynamic).
valid_wh = torch.tensor([width, height], device=device).to(dtype) - same issue as 3

xenova · 2025-02-28T11:11:32Z

optimum/exporters/tasks.py

+        ),
+        "rt-detr-v2": supported_tasks_mapping(
+            "object-detection",
+            onnx="RTDetrOnnxConfig",


Can you add another RTDetrV2OnnxConfig class in model_configs.py and use that here? (similar to how other configs do it).

class RTDetrV2OnnxConfig(RTDetrOnnxConfig): pass

xenova · 2025-02-28T11:18:00Z

tests/exporters/exporters_utils.py

+    "rt-detr": "PekingU/rtdetr_r18vd",
+    "rt-detr-v2": "PekingU/rtdetr_v2_r18vd",


If we have any tiny-random models, we can add them here. Although, considering that these models are pretty small already, we can probably just use these ones.

echarlaix

LGTM! Thanks for the addition @qubvel

echarlaix · 2025-03-03T16:13:21Z

optimum/exporters/onnx/model_configs.py

+        if kwargs["height"] < 320:
+            warnings.warn(
+                f"Exporting model with image `height={kwargs['height']}` which is less than minimal 320, setting `height` to 320."
+            )
+            kwargs["height"] = 320
+        if kwargs["width"] < 320:
+            warnings.warn(
+                f"Exporting model with image `width={kwargs['width']}` which is less than minimal 320, setting `width` to 320."
+            )
+            kwargs["width"] = 320


why is it needed and where does the value 320 comes from ?

The default height and width are set to 64, but export does not work for those values, so I override it with the minimum value divisible by 32, which is supported for export (should be greater than 300)

Thanks for the explanation! In this case shouldn't we check for num_queries in the config directly as this could vary between models ? (can be set to any value by default in cases this value cannot be extracted from the self._config)

Sounds good, I've updated the implementation

perfect thanks @qubvel !

echarlaix · 2025-03-03T17:42:01Z

Failing tests unrelated, merging, thanks again @qubvel !

YHallouard and others added 8 commits February 27, 2025 15:16

Add RTDetr Onnx Config

e92cabe

add test and comment

6c8106c

fix style

8c04a0b

Update tests large model

e35483c

fix expected configs model type

30d1c94

Fix inputs shape and remove pixel_mask

9a4a4e3

Add rt-detr-v2

07a1368

Add tests

3ece80f

qubvel mentioned this pull request Feb 27, 2025

Fix fp16 ONNX export for RT-DETR and RT-DETRv2 huggingface/transformers#36460

Open

xenova approved these changes Feb 28, 2025

View reviewed changes

Add separate V2 config

2e0290b

echarlaix approved these changes Mar 3, 2025

View reviewed changes

Update min size according to num_queries

04c3ec3

echarlaix merged commit b6c2b5c into huggingface:main Mar 3, 2025
33 of 37 checks passed

echarlaix mentioned this pull request Mar 3, 2025

[RT-Detr] - Add RT-Detr Onnx Config #2040

Closed

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add ONNX config for RT-DETR (and RT-DETRv2) #2201

Add ONNX config for RT-DETR (and RT-DETRv2) #2201

qubvel commented Feb 27, 2025 •

edited

Loading

HuggingFaceDocBuilderDev commented Feb 27, 2025

xenova left a comment •

edited

Loading

xenova Feb 28, 2025

xenova Feb 28, 2025

echarlaix left a comment •

edited

Loading

echarlaix Mar 3, 2025

qubvel Mar 3, 2025 •

edited

Loading

echarlaix Mar 3, 2025

qubvel Mar 3, 2025

echarlaix Mar 3, 2025

echarlaix commented Mar 3, 2025 •

edited

Loading

		"rt-detr": "PekingU/rtdetr_r18vd",
		"rt-detr-v2": "PekingU/rtdetr_v2_r18vd",

Add ONNX config for RT-DETR (and RT-DETRv2) #2201

Add ONNX config for RT-DETR (and RT-DETRv2) #2201

Conversation

qubvel commented Feb 27, 2025 • edited Loading

What does this PR do?

Who can review?

HuggingFaceDocBuilderDev commented Feb 27, 2025

xenova left a comment • edited Loading

Choose a reason for hiding this comment

xenova Feb 28, 2025

Choose a reason for hiding this comment

xenova Feb 28, 2025

Choose a reason for hiding this comment

echarlaix left a comment • edited Loading

Choose a reason for hiding this comment

echarlaix Mar 3, 2025

Choose a reason for hiding this comment

qubvel Mar 3, 2025 • edited Loading

Choose a reason for hiding this comment

echarlaix Mar 3, 2025

Choose a reason for hiding this comment

qubvel Mar 3, 2025

Choose a reason for hiding this comment

echarlaix Mar 3, 2025

Choose a reason for hiding this comment

echarlaix commented Mar 3, 2025 • edited Loading

qubvel commented Feb 27, 2025 •

edited

Loading

xenova left a comment •

edited

Loading

echarlaix left a comment •

edited

Loading

qubvel Mar 3, 2025 •

edited

Loading

echarlaix commented Mar 3, 2025 •

edited

Loading