coremltools: ValueError: Torch var stride_width.1 not found in context

When converting a traced torchvision model, an expected input to a mul operation is not found: ValueError: Torch var stride_width.1 not found in context

Stack Trace

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/tmp/ipykernel_16536/1090456604.py in <module>
      5     traced_model = torch.jit.trace(model_to_trace, example_image).eval()
      6 
----> 7 detector_mlmodel = ct.convert(traced_model, inputs=[ct.ImageType(shape=(1, 3, 224, 224))])
      8 detector_mlmodel.save("segmenter.mlmodel")

/opt/conda/lib/python3.7/site-packages/coremltools/converters/_converters_entry.py in convert(model, source, inputs, outputs, classifier_config, minimum_deployment_target, convert_to, compute_precision, skip_model_load, compute_units, package_dir, debug)
    454         package_dir=package_dir,
    455         debug=debug,
--> 456         specification_version=specification_version,
    457     )
    458 

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/converter.py in mil_convert(model, convert_from, convert_to, compute_units, **kwargs)
    185         See `coremltools.converters.convert`
    186     """
--> 187     return _mil_convert(model, convert_from, convert_to, ConverterRegistry, MLModel, compute_units, **kwargs)
    188 
    189 

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/converter.py in _mil_convert(model, convert_from, convert_to, registry, modelClass, compute_units, **kwargs)
    214                             convert_to,
    215                             registry,
--> 216                             **kwargs
    217                          )
    218 

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/converter.py in mil_convert_to_proto(model, convert_from, convert_to, converter_registry, **kwargs)
    279     frontend_converter = frontend_converter_type()
    280 
--> 281     prog = frontend_converter(model, **kwargs)
    282 
    283     if convert_to.lower() != "neuralnetwork":

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/converter.py in __call__(self, *args, **kwargs)
    107         from .frontend.torch import load
    108 
--> 109         return load(*args, **kwargs)
    110 
    111 

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/load.py in load(model_spec, inputs, specification_version, debug, outputs, cut_at_symbols, **kwargs)
     55     inputs = _convert_to_torch_inputtype(inputs)
     56     converter = TorchConverter(torchscript, inputs, outputs, cut_at_symbols, specification_version)
---> 57     return _perform_torch_convert(converter, debug)
     58 
     59 

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/load.py in _perform_torch_convert(converter, debug)
     94 def _perform_torch_convert(converter, debug):
     95     try:
---> 96         prog = converter.convert()
     97     except RuntimeError as e:
     98         if debug and "convert function" in str(e):

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/converter.py in convert(self)
    279 
    280             # Add the rest of the operations
--> 281             convert_nodes(self.context, self.graph)
    282 
    283             graph_outputs = [self.context[name] for name in self.graph.outputs]

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/ops.py in convert_nodes(context, graph)
     87 
     88         context.prepare_for_conversion(node)
---> 89         add_op(context, node)
     90 
     91         # We've generated all the outputs the graph needs, terminate conversion.

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/ops.py in mul(context, node)
   1388 @register_torch_op
   1389 def mul(context, node):
-> 1390     inputs = _get_inputs(context, node, expected=2)
   1391     x, y = promote_input_dtypes(inputs)
   1392     res = mb.mul(x=x, y=y, name=node.name)

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/ops.py in _get_inputs(context, node, expected, min_expected)
    187     value of @expected.
    188     """
--> 189     inputs = [context[name] for name in node.inputs]
    190     if expected is not None:
    191         expected = [expected] if not isinstance(expected, (list, tuple)) else expected

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/ops.py in <listcomp>(.0)
    187     value of @expected.
    188     """
--> 189     inputs = [context[name] for name in node.inputs]
    190     if expected is not None:
    191         expected = [expected] if not isinstance(expected, (list, tuple)) else expected

/opt/conda/lib/python3.7/site-packages/coremltools/converters/mil/frontend/torch/converter.py in __getitem__(self, torch_name)
     87                 return self._current_graph[idx][torch_name]
     88         raise ValueError(
---> 89             "Torch var {} not found in context {}".format(torch_name, self.name)
     90         )
     91 

ValueError: Torch var stride_width.1 not found in context 

Steps To Reproduce

import coremltools as ct
import torch, torchvision

detector_model = torchvision.models.detection.fasterrcnn_mobilenet_v3_large_fpn(weights="DEFAULT")
detector_model = detector_model.eval()

class FasterRCNN_MobileNetV3_AdapterModel(torch.nn.Module):
    """This adapter is only here to unbox the first output."""
    def __init__(self, model, w=2):
        super().__init__()
        self.model = model

    def forward(self, x):
        result = self.model(x)
        return result[0]['boxes'], result[0]['labels'], result[0]['scores']

adapted_detector_model = FasterRCNN_MobileNetV3_AdapterModel(detector_model)

model_to_trace = adapted_detector_model
with torch.inference_mode():
    example_image = torch.rand(1,3,224,224)
    out = model_to_trace(example_image)
    traced_model = torch.jit.trace(model_to_trace, example_image).eval()
    
detector_mlmodel = ct.convert(traced_model, inputs=[ct.ImageType(shape=(1, 3, 224, 224))])
detector_mlmodel.save("segmenter.mlmodel")
  • If the model conversion succeeds, but there is a numerical mismatch in predictions, please include the code used for comparisons.

System environment:

  • coremltools version: 6.2
  • OS: Linux (Linux foohostname 4.19.0-23-cloud-amd64 #1 SMP Debian 4.19.269-1 (2022-12-20) x86_64 GNU/Linux)
  • Any other relevant version information (e.g. PyTorch or TensorFlow version):
    • Python: 3.7
    • PyTorch: 1.12.1+cu102
    • Other libraries installed as dependencies of coremltools:
Requirement already satisfied: coremltools==6.2 in /opt/conda/lib/python3.7/site-packages (6.2)
Requirement already satisfied: tqdm in /opt/conda/lib/python3.7/site-packages (from coremltools==6.2) (4.64.1)
Requirement already satisfied: protobuf<=4.0.0,>=3.1.0 in /home/jupyter/.local/lib/python3.7/site-packages (from coremltools==6.2) (3.20.1)
Requirement already satisfied: packaging in /opt/conda/lib/python3.7/site-packages (from coremltools==6.2) (21.3)
Requirement already satisfied: numpy>=1.14.5 in /opt/conda/lib/python3.7/site-packages (from coremltools==6.2) (1.21.6)
Requirement already satisfied: sympy in /opt/conda/lib/python3.7/site-packages (from coremltools==6.2) (1.10.1)
Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in /opt/conda/lib/python3.7/site-packages (from packaging->coremltools==6.2) (3.0.9)
Requirement already satisfied: mpmath>=0.19 in /opt/conda/lib/python3.7/site-packages (from sympy->coremltools==6.2) (1.2.1)

Please advise. Thank you!

About this issue

  • Original URL
  • State: closed
  • Created 4 years ago
  • Comments: 15 (1 by maintainers)

Most upvoted comments

@junpeiz, Sounds good. Thank you for clarifying it. I’ll open new issue for torchvision::roi_align. Many of latest library like Clip requires to upgrade torch & torchvision latest version. Hopefully stride_width.1 will be fixed soon.

@ivyas21 Good, now we can confirm that stride_width.1 issue is introduced by newer version of PyTorch. It’s helpful for us to fix it in the future.

For the torchvision::roi_align not implemented issue, it’s expected, as coremltools currently doesn’t support that op. Feel free to write your own composite op by following https://coremltools.readme.io/docs/composite-operators, and also it seems that other users have proposed a solution to add that PR in https://github.com/apple/coremltools/pull/1509.

Let’s keep this thread focus on the stride_width.1 issue. For torchvision::roi_align feel free to open another issue as a feature request.