[PT FE] Support aten::aminmax for pytorch models

[LucaTamSapienza]

Details:

• Implemented aten::aminmax operation
• Implemented test for aminmax op
• registered inside op_table.cpp

Tickets:

• [Good First Issue]: Support aten::aminmax for pytorch models #23327

[LucaTamSapienza]

I have some issues when testing the "out" parameter in my code, I'm getting an error that I've never encountered before:

___________________________________ TestAminMax.test_aminmax[ ie_device:CPU - precision:FP16 - dim:None - keepdim:False - inputs:[0, 1, 2, 3, 4, -1] - mode:out - dtype:float32 ] ____________________________________

mod = <torch.ScriptMethod object at 0x136d11ae0>, inputs = [tensor([ 0.,  1.,  2.,  3.,  4., -1.]), tensor([-1.,  4.,  0.,  0.,  0.,  0.])], running_what = 'trace'

    def run_mod_and_filter_tensor_outputs(mod, inputs, running_what):
        try:
            if isinstance(inputs, dict) and example_inputs_is_kwarg:
                outs = wrap_retval(mod(**inputs))
            else:
>               outs = wrap_retval(mod(*_clone_inputs(inputs)))
E               RuntimeError: tensor does not have a device

../opvenv/lib/python3.9/site-packages/torch/jit/_trace.py:476: RuntimeError

The above exception was the direct cause of the following exception:

self = <openvino.frontend.pytorch.ts_decoder.TorchScriptPythonDecoder object at 0x161a52a90>, pt_module = aten_aminmax(), graph_element = None
example_input = [tensor([ 0.,  1.,  2.,  3.,  4., -1.]), tensor([-1.,  4.,  0.,  0.,  0.,  0.])], alias_db = None, shared_memory = True, skip_freeze = True, constant_cache = None, module_extensions = None

    def __init__(
            self,
            pt_module,
            graph_element=None,
            example_input=None,
            alias_db=None,
            shared_memory=True,
            skip_freeze=False,
            constant_cache=None,
            module_extensions=None):
        Decoder.__init__(self)
        # We store every decoder created by this decoder so that all them are not deleted until the first decoder is deleted
        self.m_decoders = []
        self._input_signature = None
        self._shared_memory = shared_memory
        self._input_is_list = False
        self.constant_cache = constant_cache if constant_cache is not None else dict()
        self.module_extensions = module_extensions
        if graph_element is None:
            try:
>               pt_module = self._get_scripted_model(
                    pt_module, example_input, skip_freeze)

../bin/arm64/Release/python/openvino/frontend/pytorch/ts_decoder.py:41: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
../bin/arm64/Release/python/openvino/frontend/pytorch/ts_decoder.py:133: in _get_scripted_model
    scripted = torch.jit.trace(
../opvenv/lib/python3.9/site-packages/torch/jit/_trace.py:806: in trace
    return trace_module(
../opvenv/lib/python3.9/site-packages/torch/jit/_trace.py:1102: in trace_module
    _check_trace(
../opvenv/lib/python3.9/site-packages/torch/utils/_contextlib.py:115: in decorate_context
    return func(*args, **kwargs)
../opvenv/lib/python3.9/site-packages/torch/jit/_trace.py:567: in _check_trace
    traced_outs = run_mod_and_filter_tensor_outputs(traced_func, inputs, "trace")
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

mod = <torch.ScriptMethod object at 0x136d11ae0>, inputs = [tensor([ 0.,  1.,  2.,  3.,  4., -1.]), tensor([-1.,  4.,  0.,  0.,  0.,  0.])], running_what = 'trace'

    def run_mod_and_filter_tensor_outputs(mod, inputs, running_what):
        try:
            if isinstance(inputs, dict) and example_inputs_is_kwarg:
                outs = wrap_retval(mod(**inputs))
            else:
                outs = wrap_retval(mod(*_clone_inputs(inputs)))
            outs = [out for out in outs if isinstance(out, torch.Tensor)]
            return outs
        except Exception as e:
            graph_diff_errors, tensor_compare_errors = graph_diagnostic_info()
            msg = f"encountered an exception while running the {running_what} with test inputs.\nException:\n{indent(str(e))}"
>           raise TracingCheckError(
                graph_diff_errors,
                tensor_compare_errors,
                extra_msg=msg,
            ) from e
E           torch.jit._trace.TracingCheckError: Tracing failed sanity checks!
E           encountered an exception while running the trace with test inputs.
E           Exception:
E           	tensor does not have a device

../opvenv/lib/python3.9/site-packages/torch/jit/_trace.py:482: TracingCheckError

This Tensor does not have a device error is driving me crazy, I don't understand what might be causing it, I'm testing only with CPU since GPU isn't related to conversion testing.
Do you have any suggestions on how to solve it?

[mvafin]

out is a tuple in torch, so this will not work. You would need to access each element of the tuple and update it using low level functions of context and decoder like add_tensor_to_context. That may be tricky to do, so I would suggest to implement this conversion only for case when out is not provided. Just validate that out is None

[LucaTamSapienza]

out is a tuple in torch, so this will not work. You would need to access each element of the tuple and update it using low level functions of context and decoder like add_tensor_to_context. That may be tricky to do, so I would suggest to implement this conversion only for case when out is not provided. Just validate that out is None

Yes, I knew I had to return a tuple since PyTorch was returning one, but since a tuple is immutable it was causing problems in my mind...
Anyway, I added a check and an assert in case out is used as a parameter.

[LucaTamSapienza]

Hi @mvafin,
I apologize for not informing you after making the changes regarding your request. This must have slowed down the process of closing this PR. However, if you could give me your review regarding these changes, I would appreciate it greatly. Thank you very much, and sorry again.

Best regards,
Luca

[mvafin]

Minor change, in general looks good

[mvafin]

build_jenkins