ONNX Export Support for CrossEntropyLoss

[KsenijaS]

Add ONNX export support for torch.nn.CrossEntropyLoss.

[dr-ci]

💊 CircleCI build failures summary and remediations

As of commit 2a3fdda (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no CircleCI failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions on the GitHub issue tracker.

This comment has been revised 51 times.

[facebook-github-bot]

@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[lara-hdr]

OnnxBackendSimpleModelTest.test_gradient_of_add_and_mul_cpu is failing on the CI, could you run it locally and see if you can repro the error?

[lara-hdr]

LGTM.Thanks!

[KsenijaS]

@houseroad can you please review?

[houseroad]

LG

[facebook-github-bot]

@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[neginraoof]

LGTM.

[KsenijaS]

@houseroad can you please merge this PR if there are no more comments

[facebook-github-bot]

@houseroad merged this pull request in 944ea4c.

[ezyang]

It looks like celu cuda needed to be blacklisted too

21:01:33 =================================== FAILURES ===================================
21:01:33 ___________________ OnnxBackendNodeModelTest.test_celu_cuda ____________________
21:01:33 
21:01:33 test_self = <caffe2.python.onnx.tests.onnx_backend_test.OnnxBackendNodeModelTest testMethod=test_celu_cuda>
21:01:33 device = 'CUDA'
21:01:33 
21:01:33     def run(test_self, device):  # type: (Any, Text) -> None
21:01:33         if model_test.model_dir is None:
21:01:33             model_dir = self.prepare_model_data(model_test)
21:01:33         else:
21:01:33             model_dir = model_test.model_dir
21:01:33         model_pb_path = os.path.join(model_dir, 'model.onnx')
21:01:33         model = onnx.load(model_pb_path)
21:01:33         model_marker[0] = model
21:01:33         if hasattr(self.backend, 'is_compatible') \
21:01:33            and callable(self.backend.is_compatible) \
21:01:33            and not self.backend.is_compatible(model):
21:01:33             raise unittest.SkipTest('Not compatible with backend')
21:01:33 >       prepared_model = self.backend.prepare(model, device)
21:01:33 
21:01:33 ../.local/lib/python3.6/site-packages/onnx/backend/test/runner/__init__.py:278: 
21:01:33 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
21:01:33 ../.local/lib/python3.6/site-packages/caffe2/python/onnx/backend.py:713: in prepare
21:01:33     init_net, predict_net = cls._onnx_model_to_caffe2_net(model, device, opset_version, False)
21:01:33 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
21:01:33 
21:01:33 cls = <class 'caffe2.python.onnx.backend.Caffe2Backend'>
21:01:33 onnx_model = ir_version: 7
21:01:33 producer_name: "backend-test"
21:01:33 graph {
21:01:33   node {
21:01:33     input: "X"
21:01:33     output: "Y"
21:01:33     op_type: "Celu"
21:01:33     at...  dim {
21:01:33             dim_value: 1
21:01:33           }
21:01:33         }
21:01:33       }
21:01:33     }
21:01:33   }
21:01:33 }
21:01:33 opset_import {
21:01:33   domain: ""
21:01:33   version: 12
21:01:33 }
21:01:33 
21:01:33 device = 'CUDA', opset_version = 12, include_initializers = False
21:01:33 
21:01:33     @classmethod
21:01:33     def _onnx_model_to_caffe2_net(cls, onnx_model, device, opset_version, include_initializers):
21:01:33         device_option = get_device_option(Device(device))
21:01:33     
21:01:33         onnx_model = onnx.utils.polish_model(onnx_model)
21:01:33         init_model = cls.optimize_onnx(onnx_model, init=True)
21:01:33         pred_model = cls.optimize_onnx(onnx_model, predict=True)
21:01:33     
21:01:33         init_net = caffe2_pb2.NetDef()
21:01:33         pred_net = caffe2_pb2.NetDef()
21:01:33     
21:01:33         init_net.name = onnx_model.graph.name + '_init'
21:01:33         pred_net.name = onnx_model.graph.name + '_predict'
21:01:33     
21:01:33         if include_initializers:
21:01:33             init_net.op.extend(cls._create_tensor_filling_op(tp) for tp in onnx_model.graph.initializer)
21:01:33     
21:01:33         cls._dummy_name.reset(cls._all_names_in_graph(init_model.graph) | cls._all_names_in_graph(pred_model.graph))
21:01:33     
21:01:33         errors = []
21:01:33         for net, model in ( (init_net, init_model), (pred_net, pred_model) ):
21:01:33             net.device_option.CopyFrom(device_option)
21:01:33             for node in model.graph.node:
21:01:33                 try:
21:01:33                     c2ops = cls._onnx_node_to_caffe2_op(
21:01:33                         init_model, pred_model, node, opset_version)
21:01:33                 except Exception as e:
21:01:33                     msg = 'Error while processing node: {}. Exception: {}'.format(node, e)
21:01:33                     errors.append(msg)
21:01:33                     print('ONNX FATAL:', msg, file=sys.stderr)
21:01:33                     continue
21:01:33                 init_net.op.extend(c2ops.init_ops)
21:01:33                 net.op.extend(c2ops.ops)
21:01:33                 net.external_input.extend(c2ops.interface_blobs)
21:01:33             net.external_output.extend(
21:01:33                 value_info.name for value_info in model.graph.output)
21:01:33             net.external_input.extend(
21:01:33                 value_info.name for value_info in model.graph.input)
21:01:33     
21:01:33         if len(errors) > 0:
21:01:33             raise RuntimeError(
21:01:33                 "ONNX conversion failed, encountered {} errors:\n\n{}".format(
21:01:33 >                   len(errors), "\n\n".join(errors)))
21:01:33 E           RuntimeError: ONNX conversion failed, encountered 1 errors:
21:01:33 E           
21:01:33 E           Error while processing node: input: "X"
21:01:33 E           output: "Y"
21:01:33 E           op_type: "Celu"
21:01:33 E           attribute {
21:01:33 E             name: "alpha"
21:01:33 E             f: 2.0
21:01:33 E             type: FLOAT
21:01:33 E           }
21:01:33 E           . Exception: Don't know how to translate op Celu
21:01:33 

[houseroad]

@ezyang which ci runs this cuda test?

[ezyang]

https://ci.pytorch.org/jenkins/job/caffe2-builds/job/py3.6-devtoolset7-rocmrpm-centos7-test/12913//console