From 7720c11d865bfcdbf7ed3cc3291d83494c826401 Mon Sep 17 00:00:00 2001 From: Vandana Kannan Date: Thu, 18 Oct 2018 15:07:56 -0700 Subject: [PATCH] ONNX export: Fully connected operator w/o bias, ReduceSum, Square (#12646) * ONNX export: Fully connected operator with no bias * ONNX export: Helper function to convert bool string attributes to int * ONNX export: ReduceSum operator * ONNX import/export: Make pow backward compatible * ONNX export: Square operator --- .../contrib/onnx/mx2onnx/_op_translations.py | 157 ++++++++++++++++-- .../contrib/onnx/onnx2mx/_op_translations.py | 11 +- .../onnx/export/mxnet_export_test.py | 26 ++- .../onnx/export/onnx_backend_test.py | 2 + 4 files changed, 179 insertions(+), 17 deletions(-) diff --git a/python/mxnet/contrib/onnx/mx2onnx/_op_translations.py b/python/mxnet/contrib/onnx/mx2onnx/_op_translations.py index 20cfe072f9f5..a740b1a52d6c 100644 --- a/python/mxnet/contrib/onnx/mx2onnx/_op_translations.py +++ b/python/mxnet/contrib/onnx/mx2onnx/_op_translations.py @@ -127,6 +127,14 @@ def convert_string_to_list(string_val): return result_list +def get_boolean_attribute_value(attrs, attr_name): + """ Helper function to convert a string version + of Boolean attributes to integer for ONNX. + Takes attribute dictionary and attr_name as + parameters. + """ + return 1 if attrs.get(attr_name, 0) in ["True", "1"] else 0 + @mx_op.register("null") def convert_weights_and_inputs(node, **kwargs): """Helper function to convert weights and inputs. @@ -214,17 +222,42 @@ def convert_fully_connected(node, **kwargs): onnx = import_onnx_modules() name = node["name"] inputs = node["inputs"] + attrs = node["attrs"] + initializer = kwargs["initializer"] + + no_bias = get_boolean_attribute_value(attrs, "no_bias") + input_node_id = kwargs["index_lookup"][inputs[0][0]] weight_node_id = kwargs["index_lookup"][inputs[1][0]] - bias_node_id = kwargs["index_lookup"][inputs[2][0]] + proc_nodes = kwargs["proc_nodes"] - input_node = proc_nodes[input_node_id] - weights_node = proc_nodes[weight_node_id] - bias_node = proc_nodes[bias_node_id] + input_node = proc_nodes[input_node_id] input_name = input_node.name + + weights_node = proc_nodes[weight_node_id] weights_name = weights_node.name - bias_name = bias_node.name + + fcnode = [] + + if no_bias == 0: + bias_node_id = kwargs["index_lookup"][inputs[2][0]] + bias_node = proc_nodes[bias_node_id] + bias_name = bias_node.name + else: + data_type = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[np.dtype('int64')] + bias_name = "bias" + str(kwargs["idx"]) + tensor_node = onnx.helper.make_tensor_value_info(bias_name, data_type, (1,)) + initializer.append( + onnx.helper.make_tensor( + name=bias_name, + data_type=data_type, + dims=(1,), + vals=[0], + raw=False, + ) + ) + fcnode.append(tensor_node) node = onnx.helper.make_node( "Gemm", @@ -237,7 +270,9 @@ def convert_fully_connected(node, **kwargs): name=name ) - return [node] + fcnode.append(node) + + return fcnode @mx_op.register("BatchNorm") @@ -587,10 +622,8 @@ def convert_dot(node, **kwargs): trans_a_node = None trans_b_node = None - trans_a = 1 if ("transpose_a" in attrs) and \ - attrs.get("transpose_a") in ["True", "1"] else 0 - trans_b = 1 if ("transpose_b" in attrs) and \ - attrs.get("transpose_b") in ["True", "1"] else 0 + trans_a = get_boolean_attribute_value(attrs, "transpose_a") + trans_b = get_boolean_attribute_value(attrs, "transpose_b") op_name = "transpose" + str(kwargs["idx"]) create_helper_trans_node(op_name, input_node_a, 'a') @@ -732,8 +765,8 @@ def convert_pooling(node, **kwargs): kernel = eval(attrs["kernel"]) pool_type = attrs["pool_type"] stride = eval(attrs["stride"]) if attrs.get("stride") else None - global_pool = True if "global_pool" in attrs and\ - attrs.get("global_pool") == "True" else False + global_pool = get_boolean_attribute_value(attrs, "global_pool") + node_inputs = node["inputs"] input_node_idx = kwargs["index_lookup"][node_inputs[0][0]] input_node = proc_nodes[input_node_idx] @@ -2053,7 +2086,31 @@ def convert_power(node, **kwargs): "Pow", [input_node_a, input_node_b], [name], - name=None + name=name + ) + return [node] + +@mx_op.register("broadcast_power") +def convert_broadcast_power(node, **kwargs): + """Map MXNet's _power operator attributes to onnx's Pow operator + and return the created node. + """ + onnx = import_onnx_modules() + name = node["name"] + proc_nodes = kwargs["proc_nodes"] + inputs = node["inputs"] + + input_node_a_id = kwargs["index_lookup"][inputs[0][0]] + input_node_b_id = kwargs["index_lookup"][inputs[1][0]] + + input_node_a = proc_nodes[input_node_a_id].name + input_node_b = proc_nodes[input_node_b_id].name + + node = onnx.helper.make_node( + "Pow", + [input_node_a, input_node_b], + [name], + name=name ) return [node] @@ -2127,3 +2184,77 @@ def convert_spacetodepth(node, **kwargs): name=name, ) return [node] + +@mx_op.register("square") +def convert_square(node, **kwargs): + """Map MXNet's square operator attributes to onnx's Pow operator + and return the created node. + """ + onnx = import_onnx_modules() + name = node["name"] + proc_nodes = kwargs["proc_nodes"] + inputs = node["inputs"] + + input_node_a_id = kwargs["index_lookup"][inputs[0][0]] + input_node_a = proc_nodes[input_node_a_id].name + + initializer = kwargs["initializer"] + data_type = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[np.dtype('int64')] + + power2_name = "square_tensor" + str(kwargs["idx"]) + tensor_node = onnx.helper.make_tensor_value_info(power2_name, data_type, (1,)) + initializer.append( + onnx.helper.make_tensor( + name=power2_name, + data_type=data_type, + dims=(1,), + vals=[2], + raw=False, + ) + ) + + node = onnx.helper.make_node( + "Pow", + [input_node_a, power2_name], + [name], + name=name + ) + return [tensor_node, node] + +@mx_op.register("sum") +def convert_sum(node, **kwargs): + """Map MXNet's sum operator attributes to onnx's ReduceSum operator + and return the created node. + """ + onnx = import_onnx_modules() + name = node["name"] + proc_nodes = kwargs["proc_nodes"] + inputs = node["inputs"] + attrs = node["attrs"] + + mx_axis = attrs.get("axis", None) + axes = convert_string_to_list(str(mx_axis)) if mx_axis is not None else None + + keepdims = get_boolean_attribute_value(attrs, "keepdims") + + input_node_id = kwargs["index_lookup"][inputs[0][0]] + input_node = proc_nodes[input_node_id].name + + if axes: + node = onnx.helper.make_node( + 'ReduceSum', + inputs=[input_node], + outputs=[name], + axes=axes, + keepdims=keepdims, + name=name + ) + else: + node = onnx.helper.make_node( + 'ReduceSum', + inputs=[input_node], + outputs=[name], + keepdims=keepdims, + name=name + ) + return [node] diff --git a/python/mxnet/contrib/onnx/onnx2mx/_op_translations.py b/python/mxnet/contrib/onnx/onnx2mx/_op_translations.py index 7040103c005a..fedd7134c3d4 100644 --- a/python/mxnet/contrib/onnx/onnx2mx/_op_translations.py +++ b/python/mxnet/contrib/onnx/onnx2mx/_op_translations.py @@ -534,10 +534,15 @@ def squareroot(attrs, inputs, proto_obj): def power(attrs, inputs, proto_obj): """Returns element-wise result of base element raised to powers from exp element.""" new_attrs = translation_utils._fix_attribute_names(attrs, {'exponent':'exp'}) - if 'broadcast' in attrs and attrs['broadcast'] == 1: + if 'broadcast' in attrs: new_attrs = translation_utils._remove_attributes(new_attrs, ['broadcast']) - return 'broadcast_power', new_attrs, inputs - return 'pow', new_attrs, inputs + if attrs['broadcast'] == 1: + return 'broadcast_power', new_attrs, inputs + else: + mxnet_op = symbol.pow(inputs[0], inputs[1]) + return mxnet_op, new_attrs, inputs + mxnet_op = symbol.broadcast_power(inputs[0], inputs[1]) + return mxnet_op, new_attrs, inputs def exponent(attrs, inputs, proto_obj): """Elementwise exponent of input array.""" diff --git a/tests/python-pytest/onnx/export/mxnet_export_test.py b/tests/python-pytest/onnx/export/mxnet_export_test.py index 7cbc98003827..9f91369d667e 100644 --- a/tests/python-pytest/onnx/export/mxnet_export_test.py +++ b/tests/python-pytest/onnx/export/mxnet_export_test.py @@ -214,6 +214,30 @@ def test_spacetodepth(): npt.assert_almost_equal(output[0], numpy_op) +@with_seed() +def test_square(): + input1 = np.random.randint(1, 10, (2, 3)).astype("float32") + + ipsym = mx.sym.Variable("input1") + square = mx.sym.square(data=ipsym) + model = mx.mod.Module(symbol=square, data_names=['input1'], label_names=None) + model.bind(for_training=False, data_shapes=[('input1', np.shape(input1))], label_shapes=None) + model.init_params() + + args, auxs = model.get_params() + params = {} + params.update(args) + params.update(auxs) + + converted_model = onnx_mxnet.export_model(square, params, [np.shape(input1)], np.float32, "square.onnx") + + sym, arg_params, aux_params = onnx_mxnet.import_model(converted_model) + result = forward_pass(sym, arg_params, aux_params, ['input1'], input1) + + numpy_op = np.square(input1) + + npt.assert_almost_equal(result, numpy_op) + if __name__ == '__main__': test_models("bvlc_googlenet", (1, 3, 224, 224), (1, 1000)) test_models("bvlc_reference_caffenet", (1, 3, 224, 224), (1, 1000)) @@ -224,4 +248,4 @@ def test_spacetodepth(): test_model_accuracy("inception_v1", (1, 3, 224, 224)) test_model_accuracy("inception_v2", (1, 3, 224, 224)) - unittest.main() \ No newline at end of file + unittest.main() diff --git a/tests/python-pytest/onnx/export/onnx_backend_test.py b/tests/python-pytest/onnx/export/onnx_backend_test.py index add5e8307753..7c3d42175b77 100644 --- a/tests/python-pytest/onnx/export/onnx_backend_test.py +++ b/tests/python-pytest/onnx/export/onnx_backend_test.py @@ -64,6 +64,8 @@ 'test_reduce_max', 'test_reduce_mean', 'test_reduce_prod', + 'test_reduce_sum_d', + 'test_reduce_sum_keepdims_random', 'test_squeeze', 'test_softmax_example', 'test_softmax_large_number',