[Relax][Frontend][TFLite] Implement DETECTION_POSTPROCESS tflite operator

[Aharrypotter]

Summary

• Implemented the TFLite DETECTION_POSTPROCESS operator conversion to Relax IR.
• Wires up the previously unimplemented operator to support object detection post-processing workflows in Relax.
• Relates to [Tracking Issue] Add missing vision/image operators #18928

Changes

• Operator Registration: Implemented convert_detection_postprocess in python/tvm/relax/frontend/tflite/tflite_frontend.py.
• Core Logic:
  • Integrated multibox_transform_loc for coordinate decoding and variance scaling.
  • Supported use_regular_nms attribute to switch between all-class NMS and class-agnostic NMS paths.
  • Leveraged all_class_non_max_suppression for efficient box filtering.
• Output Alignment: Used topk, gather_nd, and where operators to ensure the output tensors (boxes, classes, scores, num_detections) match the TFLite specification in terms of shape and layout.
• Attribute Validation: Added strict validation for required custom options such as num_classes, max_detections, and scaling factors.

Validation

Verified with linting and pre-commit hooks:

# Lint check
python -m ruff check python/tvm/relax/frontend/tflite/tflite_frontend.py

# Pre-commit checks
python -m pre_commit run --files python/tvm/relax/frontend/tflite/tflite_frontend.py

Result:

• Passed: All static checks and style guidelines are met.

[gemini-code-assist]

Code Review

This pull request implements the DETECTION_POSTPROCESS operator in the TFLite frontend for Relax, including attribute validation and NMS logic. Feedback highlights several critical issues: invalid Python unpacking of relax.Call objects (e.g., from topk), premature access to struct_info.dtype before normalization, a potential shape mismatch in mask generation, and robustness concerns regarding dynamic batch sizes in slicing operations.

[tlopex]

Several issues that need to be addressed:

1. relax.Call cannot be unpacked directly

Relax ops return relax.Call objects, not Python tuples, so direct unpacking will fail at graph construction time. This applies to at least these places:

• relax.op.vision.multibox_transform_loc(...)
• both relax.op.topk(...) calls

Please emit the call first, then extract the fields with relax.TupleGetItem, e.g.

ret = bb.emit(relax.op.vision.multibox_transform_loc(...))
transformed_boxes = relax.TupleGetItem(ret, 0)
transformed_scores = relax.TupleGetItem(ret, 1)

2. Control-flow / indentation issue

The all_class_non_max_suppression block and the assignments after if use_regular_nms: appear to be indented between the if and else branches. As written, this does not look like valid Python control flow. Please restructure this section so the if and else branches are each self-contained.

3. strided_slice with dynamic batch_size

Here:

relax.op.strided_slice(top_index_pairs, begin=[0, 0, 1], end=[batch_size, max_detections, 2])
if batch_size is symbolic, embedding it directly in a Python list will not work correctly. Please switch to a symbolic-friendly form, or avoid slicing on the batch dimension directly.

4. Unrelated formatting change

The single-line reformatting of batch_shape in convert_batch_matmul appears unrelated to this PR. Please revert it to keep the diff focused.

5. Missing test coverage

Please add an end-to-end correctness test against a TFLite model that uses DETECTION_POSTPROCESS (for example, an SSD MobileNet variant).

[Aharrypotter]

Currently thinking about how to write appropriate test cases.

[tlopex]

btw, please have a look at those gemini reviews, which I think maybe helpful

[tlopex]

@Aharrypotter Could you have a look and fix the issue?

[Aharrypotter]

Sure, will push the fix today.

[Aharrypotter]

I also fixed a bug in vision.py: the legalization path was not preserving the expected TensorFlow-style output format for DETECTION_POSTPROCESS. That created an implementation mismatch rather than just test noise, so I fixed the legalization logic and kept the updated tests aligned with the corrected behavior.

Please take a look @tlopex

[Aharrypotter]

I will debug the error later.