[v1.x] Update onnx support to work with onnx 1.7.0 with most CV models (

#19017)

* fix pooling_convention warning when convert model to onnx (#18529)

* fix  pooling_convention warning

* fix pooling_convention warning

* fix lint

Co-authored-by: JackieWu <[email protected]>

* Prevent uninitialized variable error.

* Initial work to get Dropout to work with onnx 1.7

* Remove trailing whitespace for pylint.

* Fix tensor initialization for Dropout operator input.

* Update Clip operator to support latest ONNX opset versions by moving min/max attributes to inputs.

* Fix whitespace.

* Add support for importing Dropout operator in ONNX opset version >= 12.

* Add support for import ONNX opsets >= 11 to clip operator.

* Add optional opset_version parameter that defaults to latest opset version supported by onnx. Pass this parameter to each graph layer when exporting.

* Add optional parameter to create_model() that allows user to specify which onnx opset version they want to use when exporting, defaults to latest version supported by onnx.

* Use opset_version argument to determine operator format.

* Add a opset_version parameter to from_onnx() so at operator conversion time, we know what opset version to use.

* For Clip and Dropout operators, use opset version from passed proto_obj, which reflects what opset version the onnx model uses.

* Use same tolerances that are in master.

* Change Pad operator to use inputs instead of attributes for newer opset versions. Check opset version instead of ONNX version for Pooling operator.

* Add documentation opset_version parameter.

* Add opset_version parameters to unit tests.

* Add test script for testing inference with onnxruntime on CV models from gluon model zoo.

* Add license and clean up imports.

* Install onnxruntime in docker container for unit tests.

* Add onnxruntime to test dependencies.

* Install onnxruntime into CentOS docker image.

* Disable testing squeezenet models for now.

* Update onnx version.

* Fix typo.

* Use mx.image.imread instead of PIL module.

* ONNX import: use Conv pad attribute for symmetrical padding (#18675)

Signed-off-by: Serge Panev <[email protected]>

* Install onnx in CentOS containers when installing python.

* Update import and export of some ONNX ops to support newer opset versions - this gets all ONNX unit tests to pass with onnx 1.7.

* Re-enable squeezenet model testings in onnxruntime.

* Run the onnxruntime inference tests in the ONNX pipeline instead of normal unittests pipelines.

* Add missed return value.

* Refactor code based on review comment.

* Since the onnx tests are only run on ubuntu_cpu images, we don't need to install onnx and onnxruntime in the CentOS containers.

Co-authored-by: Liu, Hao <[email protected]>
Co-authored-by: JackieWu <[email protected]>
Co-authored-by: Joe Evans <[email protected]>
Co-authored-by: Serge Panev <[email protected]>

ci/docker/install/ubuntu_onnx.sh

@@ -30,5 +30,5 @@ echo "Installing libprotobuf-dev and protobuf-compiler ..."
 apt-get update || true
 apt-get install -y libprotobuf-dev protobuf-compiler
 
-echo "Installing pytest, pytest-cov, protobuf, Pillow, ONNX and tabulate ..."
-pip3 install pytest==3.6.3 pytest-cov==2.5.1 protobuf==3.5.2 onnx==1.3.0 Pillow==5.0.0 tabulate==0.7.5
+echo "Installing pytest, pytest-cov, protobuf, Pillow, ONNX, tabulate and onnxruntime..."
+pip3 install pytest==3.6.3 pytest-cov==2.5.1 protobuf==3.5.2 onnx==1.7.0 Pillow==5.0.0 tabulate==0.7.5 onnxruntime==1.4.0

ci/docker/runtime_functions.sh

@@ -1228,11 +1228,13 @@ unittest_centos7_gpu() {
 integrationtest_ubuntu_cpu_onnx() {
 	set -ex
 	export PYTHONPATH=./python/
-    export DMLC_LOG_STACK_TRACE_DEPTH=10
+	export MXNET_SUBGRAPH_VERBOSE=0
+	export DMLC_LOG_STACK_TRACE_DEPTH=10
 	tests/python-pytest/onnx/backend_test.py
 	pytest tests/python-pytest/onnx/mxnet_export_test.py
 	pytest tests/python-pytest/onnx/test_models.py
 	pytest tests/python-pytest/onnx/test_node.py
+	pytest tests/python-pytest/onnx/test_onnxruntime.py
 }
 
 integrationtest_ubuntu_gpu_python() {

python/mxnet/contrib/onnx/mx2onnx/_op_translations.py

@@ -175,7 +175,7 @@ def convert_weights_and_inputs(node, **kwargs):
                 data_type=data_type,
                 dims=dims,
                 vals=np_arr.flatten().tolist(),
-                raw=False,
+                raw=False
             )
         )
 
@@ -462,36 +462,73 @@ def convert_pad(node, **kwargs):
     """Map MXNet's pad operator attributes to onnx's Pad operator
     and return the created node.
     """
+    opset_version = kwargs["opset_version"]
     name, input_nodes, attrs = get_inputs(node, kwargs)
 
     mxnet_pad_width = convert_string_to_list(attrs.get("pad_width"))
     onnx_pad_width = transform_padding(mxnet_pad_width)
 
     pad_mode = attrs.get("mode")
+    pad_value = np.float32(attrs.get("constant_value", 0.0))
 
-    if pad_mode == "constant":
-        pad_value = float(attrs.get("constant_value")) \
-            if "constant_value" in attrs else 0.0
-        node = onnx.helper.make_node(
-            'Pad',
-            inputs=input_nodes,
-            outputs=[name],
-            mode='constant',
-            value=pad_value,
-            pads=onnx_pad_width,
-            name=name
-        )
+    if opset_version >= 11:
+        # starting with opset 11, pads and constant_value are inputs instead of attributes
+        from onnx.helper import make_tensor, make_tensor_value_info
+        initializer = kwargs["initializer"]
+        pads_input_name = name + "_pads"
+        pads_input_type = onnx.TensorProto.INT64
+        pads_input_shape = np.shape(np.array(onnx_pad_width))
+        pads_value_node = make_tensor_value_info(pads_input_name, pads_input_type, pads_input_shape)
+        pads_tensor_node = make_tensor(pads_input_name, pads_input_type, pads_input_shape, onnx_pad_width)
+        initializer.append(pads_tensor_node)
+        input_nodes.append(pads_input_name)
+
+        if pad_mode == "constant":
+            const_input_name = name + "_constant"
+            const_input_type = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[pad_value.dtype]
+            const_value_node = make_tensor_value_info(const_input_name, const_input_type, ())
+            const_tensor_node = make_tensor(const_input_name, const_input_type, (), [pad_value])
+            initializer.append(const_tensor_node)
+            input_nodes.append(const_input_name)
+            pad_node = onnx.helper.make_node(
+                "Pad",
+                input_nodes,
+                [name],
+                mode=pad_mode,
+                name=name
+            )
+            return [pads_value_node, const_value_node, pad_node]
+        else:
+            pad_node = onnx.helper.make_node(
+                "Pad",
+                input_nodes,
+                [name],
+                mode=pad_mode,
+                name=name
+            )
+            return [pads_value_node, pad_node]
     else:
-        node = onnx.helper.make_node(
-            'Pad',
-            inputs=input_nodes,
-            outputs=[name],
-            mode=pad_mode,
-            pads=onnx_pad_width,
-            name=name
-        )
-
-    return [node]
+        if pad_mode == "constant":
+            node = onnx.helper.make_node(
+                'Pad',
+                inputs=input_nodes,
+                outputs=[name],
+                mode='constant',
+                value=pad_value,
+                pads=onnx_pad_width,
+                name=name
+            )
+            return [node]
+        else:
+            node = onnx.helper.make_node(
+                'Pad',
+                inputs=input_nodes,
+                outputs=[name],
+                mode=pad_mode,
+                pads=onnx_pad_width,
+                name=name
+            )
+            return [node]
 
 
 def create_helper_trans_node(op_name, input_node, node_name):
@@ -639,6 +676,7 @@ def convert_pooling(node, **kwargs):
     MaxPool/AveragePool/GlobalMaxPool/GlobalAveragePool operators
     based on the input node's attributes and return the created node.
     """
+    opset_version = kwargs["opset_version"]
     name, input_nodes, attrs = get_inputs(node, kwargs)
 
     kernel = eval(attrs["kernel"])
@@ -648,13 +686,14 @@ def convert_pooling(node, **kwargs):
     p_value = attrs.get('p_value', 'None')
 
     pooling_convention = attrs.get('pooling_convention', 'valid')
-
+    ceil_mode = False
     if pooling_convention == 'full':
-        pooling_warning = "Pooling: ONNX currently doesn't support pooling_convention. " \
-                          "This might lead to shape or accuracy issues. " \
-                          "https://github.com/onnx/onnx/issues/549"
-
-        logging.warning(pooling_warning)
+        if opset_version < 10:
+            pooling_warning = "Pooling: ONNX lower than 1.5.0 doesn't support pooling_convention. " \
+                              "This might lead to shape or accuracy issues. " \
+                              "https://github.com/onnx/onnx/issues/549"
+            logging.warning(pooling_warning)
+        ceil_mode = True
 
     pad_dims = list(parse_helper(attrs, "pad", [0, 0]))
     pad_dims = pad_dims + pad_dims
@@ -694,15 +733,27 @@ def convert_pooling(node, **kwargs):
                 name=name
             )
         else:
-            node = onnx.helper.make_node(
-                pool_types[pool_type],
-                input_nodes,  # input
-                [name],
-                kernel_shape=kernel,
-                pads=pad_dims,
-                strides=stride,
-                name=name
-            )
+            if opset_version >= 10:
+                node = onnx.helper.make_node(
+                    pool_types[pool_type],
+                    input_nodes,  # input
+                    [name],
+                    kernel_shape=kernel,
+                    pads=pad_dims,
+                    strides=stride,
+                    name=name,
+                    ceil_mode=ceil_mode
+                )
+            else:
+                node = onnx.helper.make_node(
+                    pool_types[pool_type],
+                    input_nodes,  # input
+                    [name],
+                    kernel_shape=kernel,
+                    pads=pad_dims,
+                    strides=stride,
+                    name=name
+                )
 
     return [node]
 
@@ -945,17 +996,35 @@ def convert_dropout(node, **kwargs):
     and return the created node.
     """
     name, input_nodes, attrs = get_inputs(node, kwargs)
+    opset_version = kwargs["opset_version"]
 
     probability = float(attrs.get("p", 0.5))
 
-    dropout_node = onnx.helper.make_node(
-        "Dropout",
-        input_nodes,
-        [name],
-        ratio=probability,
-        name=name
-    )
-    return [dropout_node]
+    if opset_version >= 12:
+        # opset >= 12 requires the ratio to be an input
+        initializer = kwargs["initializer"]
+        ratio_input_name = name + "_ratio"
+        value_node = onnx.helper.make_tensor_value_info(ratio_input_name,
+                                                        onnx.TensorProto.FLOAT, ())
+        tensor_node = onnx.helper.make_tensor(ratio_input_name, onnx.TensorProto.FLOAT,
+                                              (), [probability])
+        initializer.append(tensor_node)
+        dropout_node = onnx.helper.make_node(
+            "Dropout",
+            [input_nodes[0], ratio_input_name],
+            [name],
+            name=name
+        )
+        return [value_node, dropout_node]
+    else:
+        dropout_node = onnx.helper.make_node(
+            "Dropout",
+            input_nodes,
+            [name],
+            ratio=probability,
+            name=name
+        )
+        return [dropout_node]
 
 
 @mx_op.register("Flatten")
@@ -971,19 +1040,46 @@ def convert_clip(node, **kwargs):
     and return the created node.
     """
     name, input_nodes, attrs = get_inputs(node, kwargs)
+    opset_version = kwargs["opset_version"]
 
-    a_min = np.float(attrs.get('a_min', -np.inf))
-    a_max = np.float(attrs.get('a_max', np.inf))
+    a_min = float(attrs.get('a_min', -np.inf))
+    a_max = float(attrs.get('a_max', np.inf))
 
-    clip_node = onnx.helper.make_node(
-        "Clip",
-        input_nodes,
-        [name],
-        name=name,
-        min=a_min,
-        max=a_max
-    )
-    return [clip_node]
+    if opset_version >= 11:
+        # opset >= 11 requires min/max to be inputs
+        initializer = kwargs["initializer"]
+        min_input_name = name + "_min"
+        max_input_name = name + "_max"
+        min_value_node = onnx.helper.make_tensor_value_info(min_input_name,
+                                                            onnx.TensorProto.FLOAT, ())
+        max_value_node = onnx.helper.make_tensor_value_info(max_input_name,
+                                                            onnx.TensorProto.FLOAT, ())
+        min_tensor_node = onnx.helper.make_tensor(min_input_name, onnx.TensorProto.FLOAT,
+                                                  (), [a_min])
+        max_tensor_node = onnx.helper.make_tensor(max_input_name, onnx.TensorProto.FLOAT,
+                                                  (), [a_max])
+        initializer.append(min_tensor_node)
+        initializer.append(max_tensor_node)
+        input_nodes.append(min_input_name)
+        input_nodes.append(max_input_name)
+        clip_node = onnx.helper.make_node(
+            "Clip",
+            input_nodes,
+            [name],
+            name=name
+        )
+        return [min_value_node, max_value_node, clip_node]
+
+    else:
+        clip_node = onnx.helper.make_node(
+            "Clip",
+            input_nodes,
+            [name],
+            name=name,
+            min=a_min,
+            max=a_max
+        )
+        return [clip_node]
 
 
 def scalar_op_helper(node, op_name, **kwargs):
@@ -2070,14 +2166,34 @@ def convert_topk(node, **kwargs):
     else:
         raise NotImplementedError("ONNX expects both value and indices as output")
 
-    topk_node = onnx.helper.make_node(
-        "TopK",
-        input_nodes,
-        outputs,
-        axis=axis,
-        k=k,
-        name=name
-    )
+    opset_version = kwargs['opset_version']
+    if opset_version >= 10:
+        from onnx.helper import make_tensor, make_tensor_value_info
+        initializer = kwargs["initializer"]
+        k_input_name = name + "_k"
+        k_input_type = onnx.TensorProto.INT64
+        k_value_node = make_tensor_value_info(k_input_name, k_input_type, ())
+        k_tensor_node = make_tensor(k_input_name, k_input_type, (), k)
+        initializer.append(k_tensor_node)
+        input_nodes.append(k_input_name)
+
+        topk_node = onnx.helper.make_node(
+            "TopK",
+            input_nodes,
+            outputs,
+            axis=axis,
+            name=name
+        )
+        return [k_value_node, topk_node]
+    else:
+        topk_node = onnx.helper.make_node(
+            "TopK",
+            input_nodes,
+            outputs,
+            axis=axis,
+            k=k,
+            name=name
+        )
 
     return [topk_node]
 

python/mxnet/contrib/onnx/mx2onnx/export_model.py

@@ -29,7 +29,7 @@
 
 
 def export_model(sym, params, input_shape, input_type=np.float32,
-                 onnx_file_path='model.onnx', verbose=False):
+                 onnx_file_path='model.onnx', verbose=False, opset_version=None):
     """Exports the MXNet model file, passed as a parameter, into ONNX model.
     Accepts both symbol,parameter objects as well as json and params filepaths as input.
     Operator support and coverage -
@@ -63,11 +63,15 @@ def export_model(sym, params, input_shape, input_type=np.float32,
 
     try:
         from onnx import helper, mapping
+        from onnx.defs import onnx_opset_version
     except ImportError:
         raise ImportError("Onnx and protobuf need to be installed. "
                           + "Instructions to install - https://github.com/onnx/onnx")
 
     converter = MXNetGraph()
+    if opset_version is None:
+        # default is to use latest opset version the onnx package supports
+        opset_version = onnx_opset_version()
 
     data_format = np.dtype(input_type)
     # if input parameters are strings(file paths), load files and create symbol parameter objects
@@ -76,11 +80,11 @@ def export_model(sym, params, input_shape, input_type=np.float32,
         sym_obj, params_obj = load_module(sym, params)
         onnx_graph = converter.create_onnx_graph_proto(sym_obj, params_obj, input_shape,
                                                        mapping.NP_TYPE_TO_TENSOR_TYPE[data_format],
-                                                       verbose=verbose)
+                                                       verbose=verbose, opset_version=opset_version)
     elif isinstance(sym, symbol.Symbol) and isinstance(params, dict):
         onnx_graph = converter.create_onnx_graph_proto(sym, params, input_shape,
                                                        mapping.NP_TYPE_TO_TENSOR_TYPE[data_format],
-                                                       verbose=verbose)
+                                                       verbose=verbose, opset_version=opset_version)
     else:
         raise ValueError("Input sym and params should either be files or objects")