PaddlePaddle/PaddleOCR · error · ValueError

Invalid input name {repr(name)} found in `dynamic_shapes`

Error message

Invalid input name {repr(name)} found in `dynamic_shapes`

What it means

During _convert_trt (building TensorRT engines from the Paddle model), the code cross-checks the keys of the dynamic_shapes dict against predictor.get_input_names(). This ValueError fires when dynamic_shapes contains an input name the model does not have — a stale or hand-edited shape spec left over from a different model. The check runs before any engine building, so no work is wasted.

Source

Thrown at tools/infer/utility.py:474

            ), f"The `{type(trt_config)}` don't have the attribute `{attr_name}`!"
            setattr(trt_config, attr_name, trt_cfg_setting[attr_name])

    def _get_predictor(model_file, params_file):
        # HACK
        config = inference.Config(str(model_file), str(params_file))
        config.enable_use_gpu(100, device_id)
        # NOTE: Disable oneDNN to circumvent a bug in Paddle Inference
        config.disable_mkldnn()
        config.disable_glog_info()
        return inference.create_predictor(config)

    dynamic_shape_input_data = dynamic_shape_input_data or {}

    predictor = _get_predictor(pp_model_file, pp_params_file)
    input_names = predictor.get_input_names()
    for name in dynamic_shapes:
        if name not in input_names:
            raise ValueError(
                f"Invalid input name {repr(name)} found in `dynamic_shapes`"
            )
    for name in input_names:
        if name not in dynamic_shapes:
            raise ValueError(f"Input name {repr(name)} not found in `dynamic_shapes`")
    for name in dynamic_shape_input_data:
        if name not in input_names:
            raise ValueError(
                f"Invalid input name {repr(name)} found in `dynamic_shape_input_data`"
            )

    trt_inputs = []
    for name, candidate_shapes in dynamic_shapes.items():
        # XXX: Currently we have no way to get the data type of the tensor
        # without creating an input handle.
        handle = predictor.get_input_handle(name)
        dtype = _pd_dtype_to_np_dtype(handle.type())
        min_shape, opt_shape, max_shape = candidate_shapes

View on GitHub (pinned to 2661c7c0ef)

Solutions

1. Inspect actual input names: print(predictor.get_input_names()) or use paddle_infer to list them
2. Fix the dynamic_shapes keys (usually in inference.yml Hpi.backend_configs.paddle_infer.trt_dynamic_shapes) to exactly match the model inputs
3. Re-download the matching model package so yml and model agree

Example fix

# inference.yml — before
trt_dynamic_shapes:
  x: [[1,3,48,320],[1,3,48,640],[1,3,48,1280]]  # model input is 'images'

# after
trt_dynamic_shapes:
  images: [[1,3,48,320],[1,3,48,640],[1,3,48,1280]]

predictor = build_raw_predictor(model)  # or read yml + model once
input_names = set(predictor.get_input_names())
assert set(dynamic_shapes) <= input_names, f'dynamic_shapes keys not in {input_names}'

try:
    _convert_trt(dynamic_shapes, model_file, params_file, ...)
except ValueError as e:
    if 'Invalid input name' in str(e):
        raise SystemExit('dynamic_shapes keys must equal predictor.get_input_names()')
    raise

When it happens

Trigger: Passing dynamic_shapes={'x': ...} to a model whose input is named 'images' (or vice versa); reusing an inference.yml from one architecture with another model.

Common situations: Copying TRT config between det and rec models whose input tensor names differ; older/newer exports renaming inputs; hand-written shape dictionaries using guessed names.

Related errors

• Input name {repr(name)} not found in `dynamic_shapes`
• Invalid input name {repr(name)} found in `dynamic_shape_inpu
• Configuration Error: 'trt_dynamic_shapes' must be defined in
• Unsupported data type: {pd_dtype}

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/afd6fb9ce975d4f5. Report an issue: GitHub.