[MXNET-898] ONNX import/export: Sample_multinomial, ONNX export: Glob…

…alLpPool, LpPool (apache#13500)

* ONNX import/export: Sample_multinomial

* ONNX export: GlobalLpPool, LpPool

* Handle default p_value

* Add tests for multinomial, lppool, globallppool

* add a comment about shape test

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

@@ -586,6 +586,7 @@ def convert_pooling(node, **kwargs):
     pool_type = attrs["pool_type"]
     stride = eval(attrs["stride"]) if attrs.get("stride") else None
     global_pool = get_boolean_attribute_value(attrs, "global_pool")
+    p_value = attrs.get('p_value', 'None')
 
     pooling_convention = attrs.get('pooling_convention', 'valid')
 
@@ -598,26 +599,51 @@ def convert_pooling(node, **kwargs):
 
     pad_dims = list(parse_helper(attrs, "pad", [0, 0]))
     pad_dims = pad_dims + pad_dims
-    pool_types = {"max": "MaxPool", "avg": "AveragePool"}
-    global_pool_types = {"max": "GlobalMaxPool", "avg": "GlobalAveragePool"}
+    pool_types = {"max": "MaxPool", "avg": "AveragePool", "lp": "LpPool"}
+    global_pool_types = {"max": "GlobalMaxPool", "avg": "GlobalAveragePool",
+                         "lp": "GlobalLpPool"}
+
+    if pool_type == 'lp' and p_value == 'None':
+        raise AttributeError('ONNX requires a p value for LpPool and GlobalLpPool')
 
     if global_pool:
-        node = onnx.helper.make_node(
-            global_pool_types[pool_type],
-            input_nodes,  # input
-            [name],
-            name=name
-        )
+        if pool_type == 'lp':
+            node = onnx.helper.make_node(
+                global_pool_types[pool_type],
+                input_nodes,  # input
+                [name],
+                p=int(p_value),
+                name=name
+            )
+        else:
+            node = onnx.helper.make_node(
+                global_pool_types[pool_type],
+                input_nodes,  # input
+                [name],
+                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 pool_type == 'lp':
+            node = onnx.helper.make_node(
+                pool_types[pool_type],
+                input_nodes,  # input
+                [name],
+                p=int(p_value),
+                kernel_shape=kernel,
+                pads=pad_dims,
+                strides=stride,
+                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
+            )
 
     return [node]
 
@@ -1689,3 +1715,26 @@ def convert_logsoftmax(node, **kwargs):
         name=name
     )
     return [node]
+
+
+@mx_op.register("_sample_multinomial")
+def convert_multinomial(node, **kwargs):
+    """Map MXNet's multinomial operator attributes to onnx's
+    Multinomial operator and return the created node.
+    """
+    name, input_nodes, attrs = get_inputs(node, kwargs)
+    dtype = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[np.dtype(attrs.get("dtype", 'int32'))]
+    sample_size = convert_string_to_list(attrs.get("shape", '1'))
+    if len(sample_size) < 2:
+        sample_size = sample_size[-1]
+    else:
+        raise AttributeError("ONNX currently supports integer sample_size only")
+    node = onnx.helper.make_node(
+        "Multinomial",
+        input_nodes,
+        [name],
+        dtype=dtype,
+        sample_size=sample_size,
+        name=name,
+    )
+    return [node]

python/mxnet/contrib/onnx/onnx2mx/_import_helper.py

@@ -18,7 +18,7 @@
 # coding: utf-8_
 # pylint: disable=invalid-name
 """Operator attributes conversion"""
-from ._op_translations import identity, random_uniform, random_normal
+from ._op_translations import identity, random_uniform, random_normal, sample_multinomial
 from ._op_translations import add, subtract, multiply, divide, absolute, negative, add_n
 from ._op_translations import tanh, arccos, arcsin, arctan, _cos, _sin, _tan
 from ._op_translations import softplus, shape, gather, lp_pooling, size
@@ -48,6 +48,7 @@
     'RandomNormal'      : random_normal,
     'RandomUniformLike' : random_uniform,
     'RandomNormalLike'  : random_normal,
+    'Multinomial'       : sample_multinomial,
     # Arithmetic Operators
     'Add'               : add,
     'Sub'               : subtract,

python/mxnet/contrib/onnx/onnx2mx/_op_translations.py

@@ -38,6 +38,19 @@ def random_normal(attrs, inputs, proto_obj):
     new_attr = translation_utils._fix_attribute_names(new_attr, {'mean' : 'loc'})
     return 'random_uniform', new_attr, inputs
 
+def sample_multinomial(attrs, inputs, proto_obj):
+    """Draw random samples from a multinomial distribution."""
+    try:
+        from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
+    except ImportError:
+        raise ImportError("Onnx and protobuf need to be installed. "
+                          + "Instructions to install - /~https://github.com/onnx/onnx")
+    new_attrs = translation_utils._remove_attributes(attrs, ['seed'])
+    new_attrs = translation_utils._fix_attribute_names(new_attrs, {'sample_size': 'shape'})
+    new_attrs['dtype'] = TENSOR_TYPE_TO_NP_TYPE[int(attrs.get('dtype', 6))]
+    return 'sample_multinomial', new_attrs, inputs
+
+
 # Arithmetic Operations
 def add(attrs, inputs, proto_obj):
     """Adding two tensors"""
@@ -382,6 +395,7 @@ def global_lppooling(attrs, inputs, proto_obj):
                                                                 'kernel': (1, 1),
                                                                 'pool_type': 'lp',
                                                                 'p_value': p_value})
+    new_attrs = translation_utils._remove_attributes(new_attrs, ['p'])
     return 'Pooling', new_attrs, inputs
 
 def linalg_gemm(attrs, inputs, proto_obj):
@@ -671,11 +685,12 @@ def lp_pooling(attrs, inputs, proto_obj):
     new_attrs = translation_utils._fix_attribute_names(attrs,
                                                        {'kernel_shape': 'kernel',
                                                         'strides': 'stride',
-                                                        'pads': 'pad',
-                                                        'p_value': p_value
+                                                        'pads': 'pad'
                                                        })
+    new_attrs = translation_utils._remove_attributes(new_attrs, ['p'])
     new_attrs = translation_utils._add_extra_attributes(new_attrs,
-                                                        {'pooling_convention': 'valid'
+                                                        {'pooling_convention': 'valid',
+                                                         'p_value': p_value
                                                         })
     new_op = translation_utils._fix_pooling('lp', inputs, new_attrs)
     return new_op, new_attrs, inputs

python/mxnet/contrib/onnx/onnx2mx/_translation_utils.py

@@ -94,6 +94,7 @@ def _fix_pooling(pool_type, inputs, new_attr):
     stride = new_attr.get('stride')
     kernel = new_attr.get('kernel')
     padding = new_attr.get('pad')
+    p_value = new_attr.get('p_value')
 
     # Adding default stride.
     if stride is None:
@@ -138,7 +139,10 @@ def _fix_pooling(pool_type, inputs, new_attr):
             new_pad_op = symbol.pad(curr_sym, mode='constant', pad_width=pad_width)
 
     # Apply pooling without pads.
-    new_pooling_op = symbol.Pooling(new_pad_op, pool_type=pool_type, stride=stride, kernel=kernel)
+    if pool_type == 'lp':
+        new_pooling_op = symbol.Pooling(new_pad_op, pool_type=pool_type, stride=stride, kernel=kernel, p_value=p_value)
+    else:
+        new_pooling_op = symbol.Pooling(new_pad_op, pool_type=pool_type, stride=stride, kernel=kernel)
     return new_pooling_op
 
 def _fix_bias(op_name, attrs, num_inputs):

tests/python-pytest/onnx/test_cases.py

@@ -79,7 +79,6 @@
              'test_softplus'
              ],
     'import': ['test_gather',
-               'test_global_lppooling',
                'test_softsign',
                'test_reduce_',
                'test_mean',
@@ -89,7 +88,6 @@
                'test_averagepool_2d_precomputed_strides',
                'test_averagepool_2d_strides',
                'test_averagepool_3d',
-               'test_LpPool_',
                'test_split_equal',
                'test_hardmax'
                ],

tests/python-pytest/onnx/test_node.py

@@ -56,6 +56,24 @@ def get_rnd(shape, low=-1.0, high=1.0, dtype=np.float32):
         return np.random.choice(a=[False, True], size=shape).astype(np.float32)
 
 
+def _fix_attributes(attrs, attribute_mapping):
+    new_attrs = attrs
+    attr_modify = attribute_mapping.get('modify', {})
+    for k, v in attr_modify.items():
+        new_attrs[v] = new_attrs.pop(k, None)
+
+    attr_add = attribute_mapping.get('add', {})
+    for k, v in attr_add.items():
+        new_attrs[k] = v
+
+    attr_remove = attribute_mapping.get('remove', [])
+    for k in attr_remove:
+        if k in new_attrs:
+            del new_attrs[k]
+
+    return new_attrs
+
+
 def forward_pass(sym, arg, aux, data_names, input_data):
     """ Perform forward pass on given data
     :param sym: Symbol
@@ -118,7 +136,7 @@ def get_onnx_graph(testname, input_names, inputs, output_name, output_shape, att
             return model
 
         for test in test_cases:
-            test_name, mxnet_op, onnx_name, inputs, attrs, mxnet_specific = test
+            test_name, mxnet_op, onnx_name, inputs, attrs, mxnet_specific, fix_attrs, check_value, check_shape = test
             with self.subTest(test_name):
                 names, input_tensors, inputsym = get_input_tensors(inputs)
                 test_op = mxnet_op(*inputsym, **attrs)
@@ -131,33 +149,66 @@ def get_onnx_graph(testname, input_names, inputs, output_name, output_shape, att
                                                             onnx_name + ".onnx")
                     onnxmodel = load_model(onnxmodelfile)
                 else:
-                    onnxmodel = get_onnx_graph(test_name, names, input_tensors, onnx_name, outputshape, attrs)
+                    onnx_attrs = _fix_attributes(attrs, fix_attrs)
+                    onnxmodel = get_onnx_graph(test_name, names, input_tensors, onnx_name, outputshape, onnx_attrs)
 
                 bkd_rep = backend.prepare(onnxmodel, operation='export')
                 output = bkd_rep.run(inputs)
 
-                npt.assert_almost_equal(output[0], mxnet_output)
+                if check_value:
+                    npt.assert_almost_equal(output[0], mxnet_output)
+
+                if check_shape:
+                    npt.assert_equal(output[0].shape, outputshape)
 
 
-# test_case = ("test_case_name", mxnet op, "ONNX_op_name", [input_list], attribute map, MXNet_specific=True/False)
+# test_case = ("test_case_name", mxnet op, "ONNX_op_name", [input_list], attribute map, MXNet_specific=True/False,
+# fix_attributes = {'modify': {mxnet_attr_name: onnx_attr_name},
+#                   'remove': [attr_name],
+#                   'add': {attr_name: value},
+# check_value=True/False, check_shape=True/False)
 test_cases = [
-    ("test_equal", mx.sym.broadcast_equal, "Equal", [get_rnd((1, 3, 4, 5)), get_rnd((1, 5))], {}, False),
-    ("test_greater", mx.sym.broadcast_greater, "Greater", [get_rnd((1, 3, 4, 5)), get_rnd((1, 5))], {}, False),
-    ("test_less", mx.sym.broadcast_lesser, "Less", [get_rnd((1, 3, 4, 5)), get_rnd((1, 5))], {}, False),
+    ("test_equal", mx.sym.broadcast_equal, "Equal", [get_rnd((1, 3, 4, 5)), get_rnd((1, 5))], {}, False, {}, True,
+     False),
+    ("test_greater", mx.sym.broadcast_greater, "Greater", [get_rnd((1, 3, 4, 5)), get_rnd((1, 5))], {}, False, {}, True,
+     False),
+    ("test_less", mx.sym.broadcast_lesser, "Less", [get_rnd((1, 3, 4, 5)), get_rnd((1, 5))], {}, False, {}, True,
+     False),
     ("test_and", mx.sym.broadcast_logical_and, "And",
-     [get_rnd((3, 4, 5), dtype=np.bool_), get_rnd((3, 4, 5), dtype=np.bool_)], {}, False),
+     [get_rnd((3, 4, 5), dtype=np.bool_), get_rnd((3, 4, 5), dtype=np.bool_)], {}, False, {}, True, False),
     ("test_xor", mx.sym.broadcast_logical_xor, "Xor",
-     [get_rnd((3, 4, 5), dtype=np.bool_), get_rnd((3, 4, 5), dtype=np.bool_)], {}, False),
+     [get_rnd((3, 4, 5), dtype=np.bool_), get_rnd((3, 4, 5), dtype=np.bool_)], {}, False, {}, True, False),
     ("test_or", mx.sym.broadcast_logical_or, "Or",
-     [get_rnd((3, 4, 5), dtype=np.bool_), get_rnd((3, 4, 5), dtype=np.bool_)], {}, False),
-    ("test_not", mx.sym.logical_not, "Not", [get_rnd((3, 4, 5), dtype=np.bool_)], {}, False),
-    ("test_square", mx.sym.square, "Pow", [get_rnd((2, 3), dtype=np.int32)], {}, True),
+     [get_rnd((3, 4, 5), dtype=np.bool_), get_rnd((3, 4, 5), dtype=np.bool_)], {}, False, {}, True, False),
+    ("test_not", mx.sym.logical_not, "Not", [get_rnd((3, 4, 5), dtype=np.bool_)], {}, False, {}, True, False),
+    ("test_square", mx.sym.square, "Pow", [get_rnd((2, 3), dtype=np.int32)], {}, True, {}, True, False),
     ("test_spacetodepth", mx.sym.space_to_depth, "SpaceToDepth", [get_rnd((1, 1, 4, 6))],
-     {'block_size': 2}, False),
+     {'block_size': 2}, False, {}, True, False),
     ("test_softmax", mx.sym.SoftmaxOutput, "Softmax", [get_rnd((1000, 1000)), get_rnd(1000)],
-     {'ignore_label': 0, 'use_ignore': False}, True),
-    ("test_fullyconnected", mx.sym.FullyConnected, "Gemm", [get_rnd((4,3)), get_rnd((4, 3)), get_rnd(4)],
-     {'num_hidden': 4, 'name': 'FC'}, True)
+     {'ignore_label': 0, 'use_ignore': False}, True, {}, True, False),
+    ("test_fullyconnected", mx.sym.FullyConnected, "Gemm", [get_rnd((4, 3)), get_rnd((4, 3)), get_rnd(4)],
+     {'num_hidden': 4, 'name': 'FC'}, True, {}, True, False),
+    ("test_lppool1", mx.sym.Pooling, "LpPool", [get_rnd((2, 3, 20, 20))],
+     {'kernel': (4, 5), 'pad': (0, 0), 'stride': (1, 1), 'p_value': 1, 'pool_type': 'lp'}, False,
+     {'modify': {'kernel': 'kernel_shape', 'pad': 'pads', 'stride': 'strides', 'p_value': 'p'},
+      'remove': ['pool_type']}, True, False),
+    ("test_lppool2", mx.sym.Pooling, "LpPool", [get_rnd((2, 3, 20, 20))],
+     {'kernel': (4, 5), 'pad': (0, 0), 'stride': (1, 1), 'p_value': 2, 'pool_type': 'lp'}, False,
+     {'modify': {'kernel': 'kernel_shape', 'pad': 'pads', 'stride': 'strides', 'p_value': 'p'},
+      'remove': ['pool_type']}, True, False),
+    ("test_globallppool1", mx.sym.Pooling, "GlobalLpPool", [get_rnd((2, 3, 20, 20))],
+     {'kernel': (4, 5), 'pad': (0, 0), 'stride': (1, 1), 'p_value': 1, 'pool_type': 'lp', 'global_pool': True}, False,
+     {'modify': {'p_value': 'p'},
+      'remove': ['pool_type', 'kernel', 'pad', 'stride', 'global_pool']}, True, False),
+    ("test_globallppool2", mx.sym.Pooling, "GlobalLpPool", [get_rnd((2, 3, 20, 20))],
+     {'kernel': (4, 5), 'pad': (0, 0), 'stride': (1, 1), 'p_value': 2, 'pool_type': 'lp', 'global_pool': True}, False,
+     {'modify': {'p_value': 'p'},
+      'remove': ['pool_type', 'kernel', 'pad', 'stride', 'global_pool']}, True, False),
+
+    # since results would be random, checking for shape alone
+    ("test_multinomial", mx.sym.sample_multinomial, "Multinomial",
+     [np.array([0, 0.1, 0.2, 0.3, 0.4]).astype("float32")],
+     {'shape': (10,)}, False, {'modify': {'shape': 'sample_size'}}, False, True)
 ]
 
 if __name__ == '__main__':