PyTorch 1.6.0 released today - breaks coremltools converter v4.0b2 in many cases with AttributeError: module 'torch._C' has no attribute '_jit_pass_canonicalize_ops'

[leovinus2001]

I just noticed that PyTorch 1.6.0 was released. Now, many test cases from the last 2 weeks break with the log below and I believe a similar issue was reported earlier in issue #751 for pytorch nightly

Reproducible::--------------------------

yes, with torch =1.6.0 and torchvision = 0.7.0

Example test cases that fail now are #823 #825 or
#736 (comment)

Log:--------------------------

Traceback (most recent call last):
File "testAssert.py", line 55, in
inputs= [ ct.TensorType(name="input1", shape=dummy_input.shape) ]
File "/Library/Python/3.7/lib/python/site-packages/coremltools/converters/_converters_entry.py", line 299, in convert
**kwargs
File "/Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/converter.py", line 120, in _convert
prog = frontend_converter(model, **kwargs)
File "/Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/converter.py", line 62, in call
return load(*args, **kwargs)
File "/Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/load.py", line 73, in load
converter = TorchConverter(torchscript, inputs, outputs, cut_at_symbols)
File "/Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 141, in init
raw_graph, params_dict = self._expand_and_optimize_ir(self.torchscript)
File "/Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 273, in _expand_and_optimize_ir
_torch._C._jit_pass_canonicalize_ops(graph)
AttributeError: module 'torch._C' has no attribute '_jit_pass_canonicalize_ops'

[leovinus2001]

While I do not know how to fix it properly, this is what I can see in case that is helpful for someone else.

1. coremltools/converters/mil/frontend/torch/converter.py
   The converter bails on the line 273
   _torch._C._jit_pass_canonicalize_ops(graph)

There are two alternatives without the _ops:
_torch._C._jit_pass_canonicalize(graph)
or
_torch._C._jit_pass_canonicalize_graph_fuser_ops(graph)

Neither one has lasting success as we then end up with (sometimes, not always)

2. in
   inputs=[ ct.TensorType(name="input1", shape=dummy_input.shape) ],
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/_converters_entry.py", line 299, in convert
   **kwargs
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/converter.py", line 120, in _convert
   prog = frontend_converter(model, **kwargs)
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/converter.py", line 62, in call
   return load(*args, **kwargs)
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/load.py", line 73, in load
   converter = TorchConverter(torchscript, inputs, outputs, cut_at_symbols)
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 143, in init
   raw_graph, params_dict, self.inputs, cut_at_symbols
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/internal_graph.py", line 241, in init
   new_node = InternalTorchIRNode(raw_node, parent=self)
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/internal_graph.py", line 142, in init
   for name in node.attributeNames()
   File "Library/Python/3.7/lib/python/site-packages/coremltools/converters/mil/frontend/torch/internal_graph.py", line 142, in
   for name in node.attributeNames()
   AttributeError: 'torch._C.Node' object has no attribute 'ival'

And that error is working on a constant which has 'value' but not 'ival '.

[hovhanns]

The same issue occurred in my case.
Used segmentation_models_pytorch

Steps to reproduce

import segmentation_models_pytorch as smp
def Smp_Unet(out_channels=1):
    model = smp.Unet("efficientnet-b0", classes=out_channels, encoder_weights="imagenet")
    model.segmentation_head[2] = nn.ReLU(inplace=True)
    return model
model = Smp_Unet()
model.load_state_dict(model_state_dict)
model.encoder.set_swish(memory_efficient=False)
model.eval()
dummy_input = torch.rand(1, 3, 512, 512)
tranced_model = torch.jit.trace(model, dummy_input)
core_ml_model = ct.convert(tranced_model, inputs=[ct.ImageType(name="image_in")])

Getting the AttributeError

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-13-719676927649> in <module>
      1 tranced_model = torch.jit.trace(model, dummy_input)
----> 2 core_ml_model = ct.convert(tranced_model, inputs=[ct.ImageType(name="image_in")])

~/.local/lib/python3.8/site-packages/coremltools/converters/_converters_entry.py in convert(model, source, inputs, outputs, classifier_config, minimum_deployment_target, **kwargs)
    290             raise ValueError("outputs must not be specified for PyTorch")
    291 
--> 292         proto_spec = _convert(
    293             model,
    294             convert_from="torch",

~/.local/lib/python3.8/site-packages/coremltools/converters/mil/converter.py in _convert(model, convert_from, convert_to, converter_registry, **kwargs)
    118     backend_converter = backend_converter_type()
    119 
--> 120     prog = frontend_converter(model, **kwargs)
    121     common_pass(prog)
    122     out = backend_converter(prog, **kwargs)

~/.local/lib/python3.8/site-packages/coremltools/converters/mil/converter.py in __call__(self, *args, **kwargs)
     60         from .frontend.torch import load
     61 
---> 62         return load(*args, **kwargs)
     63 
     64 

~/.local/lib/python3.8/site-packages/coremltools/converters/mil/frontend/torch/load.py in load(model_spec, debug, **kwargs)
     71     outputs = kwargs.get("outputs", None)
     72     cut_at_symbols = kwargs.get("cut_at_symbols", None)
---> 73     converter = TorchConverter(torchscript, inputs, outputs, cut_at_symbols)
     74 
     75     try:

~/.local/lib/python3.8/site-packages/coremltools/converters/mil/frontend/torch/converter.py in __init__(self, torchscript, inputs, outputs, cut_at_symbols)
    138         self.output_names = outputs
    139         self.context = TranscriptionContext()
--> 140         raw_graph, params_dict = self._expand_and_optimize_ir(self.torchscript)
    141         self.graph = InternalTorchIRGraph(
    142             raw_graph, params_dict, self.inputs, cut_at_symbols

~/.local/lib/python3.8/site-packages/coremltools/converters/mil/frontend/torch/converter.py in _expand_and_optimize_ir(torchscript)
    352         _torch._C._jit_pass_lint(graph)
    353         # Replaces a couple specific ops patterns (add, sub, mul, div, chunk).
--> 354         _torch._C._jit_pass_canonicalize_ops(graph)
    355         _torch._C._jit_pass_lint(graph)
    356         # From PyTorch code: This pass catches all of the small, easy to catch

AttributeError: module 'torch._C' has no attribute '_jit_pass_canonicalize_ops'

[leovinus2001]

As we have seen no patch from the coremltools team, may I politely enquire whether coremltools plans to be compatible with PyTorch 1.6 going forward?

[aseemw]

Yes, we are currently working on resolving the issues with pytorch 1.6.0 and the next coremltools beta release will support it.

[akash1301]

Is there any way to avoid this error by downgrading the PyTorch version?

[leovinus2001]

Maybe I misunderstand the question but PyTorch 1.5.1 works fine with coremltoolsv4.0b2 or TOT latest. I use "pip" to jump between forward and backward between sets of pytorch/torchvision etc of compatible versions. Going forward, we just need to know for sure that 1.6.x will work as well, and @aseemw confirmed earlier it will. It would be nice to know when that fix will drop of course, hint ;) And, please verify torchvision and audio as well with Pytorch 1.6, thanks.

[leovinus2001]

The 4.0b3 release of coremltools fixes this issue.