RuntimeError: tensor op type aten::view has more than one matched: [[('size', 'List[int]', 'None')], [('dtype', 'int', 'None')]]

[sayakpaul]

Describe the issue:

RuntimeError: tensor op type aten::view has more than one matched: [[('size', 'List[int]', 'None')], [('dtype', 'int', 'None')]]

Environment: Google Colab

Configuration:

I am referring to this example.

Log message:

/usr/local/lib/python3.7/dist-packages/nni/retiarii/experiment/pytorch.py in run(self, config, port, debug)
    255             assert config is not None, 'You are using classic search mode, config cannot be None!'
    256             self.config = config
--> 257             self.start(port, debug)
    258 
    259     def _check_exp_status(self) -> bool:

/usr/local/lib/python3.7/dist-packages/nni/retiarii/experiment/pytorch.py in start(self, port, debug)
    237         exp_status_checker = Thread(target=self._check_exp_status)
    238         exp_status_checker.start()
--> 239         self._start_strategy()
    240         # TODO: the experiment should be completed, when strategy exits and there is no running job
    241         _logger.info('Waiting for experiment to become DONE (you can ctrl+c if there is no running trial jobs)...')

/usr/local/lib/python3.7/dist-packages/nni/retiarii/experiment/pytorch.py in _start_strategy(self)
    175 
    176         _logger.info('Start strategy...')
--> 177         self.strategy.run(base_model_ir, self.applied_mutators)
    178         _logger.info('Strategy exit')
    179         # TODO: find out a proper way to show no more trial message on WebUI

/usr/local/lib/python3.7/dist-packages/nni/retiarii/strategy/tpe_strategy.py in run(self, base_model, applied_mutators)
     77                     model = mutator.apply(model)
     78                 # run models
---> 79                 submit_models(model)
     80                 self.running_models[self.model_id] = model
     81                 self.model_id += 1

/usr/local/lib/python3.7/dist-packages/nni/retiarii/execution/api.py in submit_models(*models)
     42     engine = get_execution_engine()
     43     get_and_register_default_listener(engine)
---> 44     engine.submit_models(*models)
     45 
     46 

/usr/local/lib/python3.7/dist-packages/nni/retiarii/execution/base.py in submit_models(self, *models)
     60     def submit_models(self, *models: Model) -> None:
     61         for model in models:
---> 62             data = self.pack_model_data(model)
     63             self._running_models[send_trial(data.dump())] = model
     64             self._history.append(model)

/usr/local/lib/python3.7/dist-packages/nni/retiarii/execution/base.py in pack_model_data(cls, model)
    111     @classmethod
    112     def pack_model_data(cls, model: Model) -> Any:
--> 113         return BaseGraphData(codegen.model_to_pytorch_script(model), model.evaluator)
    114 
    115     @classmethod

/usr/local/lib/python3.7/dist-packages/nni/retiarii/codegen/pytorch.py in model_to_pytorch_script(model, placement)
     14     total_pkgs = set()
     15     for name, cell in model.graphs.items():
---> 16         import_pkgs, graph_code = graph_to_pytorch_model(name, cell, placement=placement)
     17         graphs.append(graph_code)
     18         total_pkgs.update(import_pkgs)

/usr/local/lib/python3.7/dist-packages/nni/retiarii/codegen/pytorch.py in graph_to_pytorch_model(graph_name, graph, placement)
    137             if node.operation.type == 'shared':
    138                 submodule_name = _remove_prefix(node.operation.parameters['reference'], graph_name)
--> 139             edge_codes.append(node.operation.to_forward_code(submodule_name, node_name, inputs, inputs_value))
    140 
    141     output_names, _ = _format_inputs(graph.output_node)

/usr/local/lib/python3.7/dist-packages/nni/retiarii/operation_def/torch_op_def.py in to_forward_code(self, field, output, inputs, inputs_value)
    371         if self.type in TensorOps.comparison_ops:
    372             return f'{output} = ({inputs[0]} {TensorOps.comparison_ops[self.type]} {inputs[1]})'
--> 373         matched_args = TensorOps._get_matched_args(self.type, inputs)
    374         if matched_args is None:
    375             return TensorOpExceptions[self.type](output, inputs)

/usr/local/lib/python3.7/dist-packages/nni/retiarii/operation_def/torch_op_def.py in _get_matched_args(_type, inputs)
    361                 return ManuallyChooseDef[_type]
    362             else:
--> 363                 raise RuntimeError(f'tensor op type {_type} has more than one matched: {matched}')
    364         else:
    365             if _type in TensorOpExceptions:

RuntimeError: tensor op type aten::view has more than one matched: [[('size', 'List[int]', 'None')], [('dtype', 'int', 'None')]]

How to reproduce it?: Here's the Colab Notebook.

[QuanluZhang]

@sayakpaul , in v2.3 retiarii only supports torch 1.6 and 1.7.

torch 1.8 and 1.9 will be supported in NNI v2.4 which will be released soon. you can checkout this pr #3937

[sayakpaul]

Let me run the notebook by checking out the branch first.

[sayakpaul]

@QuanluZhang could you tell me how I should install nni from your branch so that it's compatible with Torch 1.9?

[QuanluZhang]

@sayakpaul you can install from source, please refer to this doc https://nni.readthedocs.io/en/stable/Tutorial/InstallationLinux.html#install-nni-through-source-code

[sayakpaul]

@QuanluZhang was able to finally nni (built from source). Here's the Colab Notebook in case anyone else faces a similar problem can get a workaround.

[sayakpaul]

@QuanluZhang on a single Tesla P100, how long do you think the example should take to complete? Having some timing information would be really helpful.