paddle_op_mapper.py

#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import math
import sys
import x2paddle
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
from onnx import helper, onnx_pb


class PaddleOpMapper(object):
    def __init__(self):
        self.paddle_onnx_dtype_map = {
            core.VarDesc.VarType.FP32: onnx_pb.TensorProto.FLOAT,
            core.VarDesc.VarType.FP64: onnx_pb.TensorProto.DOUBLE,
            core.VarDesc.VarType.INT32: onnx_pb.TensorProto.INT32,
            core.VarDesc.VarType.INT16: onnx_pb.TensorProto.INT16,
            core.VarDesc.VarType.INT16: onnx_pb.TensorProto.UINT16,
            core.VarDesc.VarType.INT64: onnx_pb.TensorProto.INT64,
            core.VarDesc.VarType.BOOL: onnx_pb.TensorProto.BOOL
        }

        self.name_counter = dict()

    def get_name(self, op_name, var_name):
        name = 'p2o.{}.{}'.format(op_name, var_name)
        if name not in self.name_counter:
            self.name_counter[name] = 0
        else:
            self.name_counter[name] += 1
        return name + '.{}'.format(self.name_counter[name])

    def make_constant_node(self, name, dtype, value=None):
        if isinstance(value, list):
            dims = (len(value), )
        elif value is None:
            dims = ()
            value = []
        else:
            dims = ()
            value = [value]
        tensor = helper.make_tensor(
            name=name, data_type=dtype, dims=dims, vals=value)
        node = helper.make_node(
            'Constant', inputs=[], outputs=[name], value=tensor)
        return node

    def conv2d(self, op, block):
        kernel_shape = block.var(op.input('Filter')[0]).shape
        node = helper.make_node(
            'Conv',
            inputs=op.input('Input') + op.input('Filter'),
            outputs=op.output('Output'),
            dilations=op.attr('dilations'),
            kernel_shape=kernel_shape[-2:],
            strides=op.attr('strides'),
            group=op.attr('groups'),
            pads=op.attr('paddings') + op.attr('paddings'))
        return node

    def relu(self, op, block):
        node = helper.make_node(
            'Relu', inputs=op.input('X'), outputs=op.output('Out'))
        return node

    def elementwise_add(self, op, block):
        axis = op.attr('axis')
        x_shape = block.var(op.input('X')[0]).shape
        y_shape = block.var(op.input('Y')[0]).shape
        if len(y_shape) == 1 and axis == 1:
            shape_name = self.get_name(op.type, 'shape')
            shape_value = [1] * len(x_shape)
            shape_value[axis] = y_shape[0]
            shape_node = self.make_constant_node(
                shape_name, onnx_pb.TensorProto.INT64, shape_value)
            temp_value = self.get_name(op.type, 'temp')
            y_node = helper.make_node(
                'Reshape',
                inputs=[op.input('Y')[0], shape_name],
                outputs=[temp_value])
            node = helper.make_node(
                'Add',
                inputs=[op.input('X')[0], temp_value],
                outputs=op.output('Out'))
            return [shape_node, y_node, node]
        elif len(x_shape) == len(y_shape):
            node = helper.make_node(
                'Add',
                inputs=[op.input('X')[0], op.input('Y')[0]],
                outputs=op.output('Out'))
            return node
        else:
            raise Excpetion("Unexpected situation happend in elementwise_add")

    def pool2d(self, op, block):
        pool_type = {
            'max': ('MaxPool', 'GlobalMaxPool'),
            'avg': ('AveragePool', 'GlobalAveragePool')
        }
        if op.attr('global_pooling'):
            node = helper.make_node(
                pool_type[op.attr('pooling_type')][1],
                inputs=op.input('X'),
                outputs=op.output('Out'),
            )
        else:
            node = helper.make_node(
                pool_type[op.attr('pooling_type')][0],
                inputs=op.input('X'),
                outputs=op.output('Out'),
                kernel_shape=op.attr('ksize'),
                strides=op.attr('strides'),
                pads=op.attr('paddings') + op.attr('paddings'))
        return node

    def softmax(self, op, block):
        node = helper.make_node(
            'Softmax',
            inputs=op.input('X'),
            outputs=op.output('Out'),
            axis=op.attr('axis'))
        return node

    def scale(self, op, block):
        scale = op.attr('scale')
        bias = op.attr('bias')
        if math.fabs(scale - 1.0) < 1e-06 and math.fabs(bias - 0.0) < 1e-06:
            node = helper.make_node(
                'Identity', inputs=op.input('X'), outputs=op.output('Out'))
            return node
        else:
            scale_name = self.get_name(op.type, 'scale')
            bias_name = self.get_name(op.type, 'bias')
            scale_node = self.make_constant_node(
                scale_name, onnx_pb.TensorProto.FLOAT, scale)
            bias_node = self.make_constant_node(bias_name,
                                                onnx_pb.TensorProto.FLOAT, bias)
            temp_tensor_name = self.get_name(op.type, 'temporary')
            if op.attr('bias_after_scale'):
                node1 = helper.make_node(
                    'Mul',
                    inputs=[scale_name, op.input('X')[0]],
                    outputs=[temp_tensor_name])
                node2 = helper.make_node(
                    'Add',
                    inputs=[bias_name, temp_tensor_name],
                    outputs=op.output('Out'))
            else:
                node1 = helper.make_node(
                    'Add',
                    inputs=[bias_name, op.input('X')[0]],
                    outputs=temp_tensor_name)
                node2 = helper.make_node(
                    'Mul',
                    inputs=[scale_name, temp_tensor_name],
                    outputs=[op.output('Out')])
            return [scale_node, bias_node, node1, node2]

    def mul(self, op, block):
        x_shape = block.var(op.input('X')[0]).shape
        y_shape = block.var(op.input('Y')[0]).shape
        out_shape = list(block.var(op.output('Out')[0]).shape)
        x_num_col_dims = op.attr('x_num_col_dims')
        y_num_col_dims = op.attr('y_num_col_dims')
        flatten_x_name = 'flatten_{}'.format(op.input('X')[0])
        flatten_y_name = 'flatten_{}'.format(op.input('Y')[0])
        shape_name = 'temp_shape_{}'.format(op.output('Out')[0])
        temp_out_name = 'temp_{}'.format(op.output('Out')[0])
        flatten_x = helper.make_node(
            'Flatten',
            inputs=op.input('X'),
            outputs=[flatten_x_name],
            axis=x_num_col_dims)
        flatten_y = helper.make_node(
            'Flatten',
            inputs=op.input('Y'),
            outputs=[flatten_y_name],
            axis=y_num_col_dims)
        shape_node = self.make_constant_node(
            shape_name, onnx_pb.TensorProto.INT64, out_shape)
        node = helper.make_node(
            'MatMul',
            inputs=[flatten_x_name, flatten_y_name],
            outputs=[temp_out_name])
        reshape_out = helper.make_node(
            'Reshape',
            inputs=[temp_out_name, shape_name],
            outputs=op.output('Out'))
        return [flatten_x, flatten_y, shape_node, node, reshape_out]

    def batch_norm(self, op, block):
        kwargs = {
            'epsilon': op.attr('epsilon'),
            'momentum': op.attr('momentum')
        }
        inputs = op.input('X') + op.input('Scale') + op.input(
            'Bias') + op.input('Mean') + op.input('Variance')
        node = helper.make_node(
            'BatchNormalization',
            inputs=inputs,
            outputs=op.output('Y'),
            **kwargs)
        return node

    def concat(self, op, block):
        node = helper.make_node(
            'Concat',
            inputs=op.input('X'),
            outputs=op.output('Out'),
            axis=op.attr('axis'))
        return node

    def depthwise_conv2d(self, op, block):
        return self.conv2d(op, block)

    def relu6(self, op, block):
        min_name = self.get_name(op.type, 'min')
        max_name = self.get_name(op.type, 'max')
        min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
                                           0)
        max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
                                           op.attr('threshold'))
        node = helper.make_node(
            'Clip',
            inputs=[op.input('X')[0], min_name, max_name],
            outputs=op.output('Out'),
        )
        return [min_node, max_node, node]

    def shape(self, op, block):
        node = helper.make_node(
            'Shape', inputs=op.input('Input'), outputs=op.output('Out'))
        return node

    def split(self, op, block):
        sections = op.attr('sections')
        if len(sections) > 0:
            node = helper.make_node(
                'Split',
                inputs=op.input('X'),
                outputs=op.output('Out'),
                axis=op.attr('axis'),
                split=sections)
        else:
            node = helper.make_node(
                'Split',
                inputs=op.input('X'),
                outputs=op.output('Out'),
                axis=op.attr('axis'))
        return node

    def slice(self, op, block):
        axes = op.attr('axes')
        starts = op.attr('starts')
        ends = op.attr('ends')
        axes_name = self.get_name(op.type, 'axes')
        starts_name = self.get_name(op.type, 'starts')
        ends_name = self.get_name(op.type, 'ends')

        axes_node = self.make_constant_node(axes_name,
                                            onnx_pb.TensorProto.INT64, axes)
        starts_node = self.make_constant_node(starts_name,
                                              onnx_pb.TensorProto.INT64, starts)
        ends_node = self.make_constant_node(ends_name,
                                            onnx_pb.TensorProto.INT64, ends)
        node = helper.make_node(
            "Slice",
            inputs=[op.input('Input')[0], starts_name, ends_name, axes_name],
            outputs=op.output('Out'),
        )
        return [starts_node, ends_node, axes_node, node]

    def fill_constant(self, op, block):
        value = op.attr('value')
        dtype = op.attr('dtype')
        shape = op.attr('shape')
        value = np.ones(shape) * value
        if dtype == 2:
            value = value.astype('int32')
        node = helper.make_node(
            'Constant',
            inputs=[],
            outputs=op.output('Out'),
            value=helper.make_tensor(
                name=op.output('Out')[0],
                data_type=self.paddle_onnx_dtype_map[dtype],
                dims=shape,
                vals=value.tolist()))
        return node

    def transpose2(self, op, block):
        node = helper.make_node(
            'Transpose',
            inputs=op.input('X'),
            outputs=op.output('Out'),
            perm=op.attr('axis'))
        return node

    def reshape2(self, op, block):
        input_names = op.input_names
        if len(op.input('ShapeTensor')) > 1:
            cast_shape_nodes = list()
            cast_shape_names = list()
            for i in range(len(op.input('ShapeTensor'))):
                dim = op.input('ShapeTensor')[i]
                temp_name = self.get_name(op.type, 'shape.cast')
                node = helper.make_node(
                    'Cast',
                    inputs=[dim],
                    outputs=[temp_name],
                    to=onnx_pb.TensorProto.INT64)
                cast_shape_nodes.append(node)
                cast_shape_names.append(temp_name)

            temp_name = self.get_name(op.type, 'shape.concat')
            shape_node = helper.make_node(
                'Concat', inputs=cast_shape_names, outputs=[temp_name], axis=-1)
            node = helper.make_node(
                'Reshape',
                inputs=[op.input('X')[0], temp_name],
                outputs=op.output('Out'))
            return cast_shape_nodes + [shape_node, node]
        else:
            temp_name = self.get_name(op.type, 'shape.cast')
            cast_shape_node = helper.make_node(
                'Cast',
                inputs=op.input('ShapeTensor'),
                outputs=[temp_name],
                to=onnx_pb.TensorProto.INT64)
            node = helper.make_node(
                'Reshape',
                inputs=[op.input('X')[0], temp_name],
                outputs=op.output('Out'))
            return [cast_shape_node, node]

    def dropout(self, op, block):
        dropout_mode = op.attr('dropout_implementation')
        dropout_prob = op.attr('dropout_prob')
        if dropout_mode == 'upscale_in_train':
            node = helper.make_node(
                'Identity', inputs=op.input('X'), outputs=op.output('Out'))
            return node
        elif dropout_mode == 'downgrade_in_infer':
            scale_name = self.get_name(op.type, 'scale')
            scale_node = self.make_constant_node(
                scale_name, onnx_pb.TensorProto.FLOAT, 1 - dropout_prob)
            node = helper.make_node(
                "Mul",
                inputs=[op.input('X')[0], scale_name],
                outputs=op.output('Out'))
            return [scale_node, node]
        else:
            raise Exception("Unexpected situation happend")

    def reduce_mean(self, op, block):
        node = helper.make_node(
            'ReduceMean',
            inputs=op.input('X'),
            outputs=op.output('Out'),
            axes=op.attr('dim'),
            keepdims=op.attr('keep_dim'))
        return node

    def nearest_interp(self, op, block):
        input_names = op.input_names
        coordinate_transformation_mode = 'half_pixel'
        if op.attr('align_corners'):
            coordinate_transformation_mode = 'align_corners'
        if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
            node = helper.make_node(
                'Resize',
                inputs=[op.input('X')[0], '',
                        op.input('OutSize')[0]],
                outputs=op.output('Out'),
                mode='nearest',
                coordinate_transformation_mode=coordinate_transformation_mode)
        elif 'Scale' in input_names and len(op.input('Scale')) > 0:
            node = helper.make_node(
                'Resize',
                inputs=[op.input('X')[0],
                        op.input('Scale')[0]],
                outputs=op.output('Out'),
                mode='nearest',
                coordinate_transformation_mode=coordinate_transformation_mode)
        else:
            out_shape = [op.attr('out_h'), op.attr('out_w')]
            scale = op.attr('scale')
            if out_shape.count(-1) > 0:
                scale_name = self.get_name(op.type, 'scale')
                scale_node = self.make_constant_node(
                    scale_name, onnx_pb.TensorProto.FLOAT, [1, 1, scale, scale])
                roi_name = self.get_name(op.type, 'roi')
                roi_node = self.make_constant_node(roi_name,
                                                   onnx_pb.TensorProto.FLOAT,
                                                   [1, 1, 1, 1, 1, 1, 1, 1])
                node = helper.make_node(
                    'Resize',
                    inputs=[op.input('X')[0], roi_name, scale_name],
                    outputs=op.output('Out'),
                    mode='nearest',
                    coordinate_transformation_mode=coordinate_transformation_mode
                )
                return [scale_node, roi_node, node]
            else:
                raise Exception("Unexpected situation happend")
        return node

    def hard_sigmoid(self, op, block):
        slope = op.attr('slope')
        offset = op.attr('offset')
        node = helper.make_node(
            'HardSigmoid',
            inputs=op.input('X'),
            outputs=op.output('Out'),
            alpha=slope,
            beta=offset)
        return node

    def elementwise_mul(self, op, block):
        axis = op.attr('axis')
        x_shape = block.var(op.input('X')[0]).shape
        y_shape = block.var(op.input('Y')[0]).shape
        if len(y_shape) == 1 and axis == 1:
            shape_name = self.get_name(op.type, 'shape')
            shape_value = [1] * len(x_shape)
            shape_value[axis] = y_shape[0]
            shape_node = self.make_constant_node(
                shape_name, onnx_pb.TensorProto.INT64, shape_value)
            temp_value = self.get_name(op.type, 'temp')
            y_node = helper.make_node(
                'Reshape',
                inputs=[op.input('Y')[0], shape_name],
                outputs=[temp_value])
            node = helper.make_node(
                'Mul',
                inputs=[op.input('X')[0], temp_value],
                outputs=op.output('Out'))
            return [shape_node, y_node, node]
        elif len(x_shape) == len(y_shape):
            node = helper.make_node(
                'Mul',
                inputs=[op.input('X')[0], op.input('Y')[0]],
                outputs=op.output('Out'))
            return node
        else:
            raise Excpetion("Unexpected situation happend in elementwise_add")
        return node

    def feed(self, op, block):
        name = op.output('Out')[0]
        var = block.var(name)
        tensor_info = helper.make_tensor_value_info(
            name=name,
            shape=var.shape,
            elem_type=self.paddle_onnx_dtype_map[var.dtype])
        return tensor_info

    def fetch(self, op, block):
        name = op.input('X')[0]
        var = block.var(name)
        tensor_info = helper.make_tensor_value_info(
            name=name,
            shape=var.shape,
            elem_type=self.paddle_onnx_dtype_map[var.dtype])
        return tensor_info

    def convert_weights(self, program):
        var_names = program.global_block().vars
        nodes = list()
        for name in var_names:
            var = program.global_block().var(name)
            if name.endswith('feed') or name.endswith('fetch'):
                continue
            if not var.persistable:
                continue
            weight = np.array(fluid.global_scope().find_var(name).get_tensor())
            tensor = helper.make_tensor(
                name=name,
                dims=var.shape,
                data_type=self.paddle_onnx_dtype_map[var.dtype],
                vals=weight.flatten().tolist())
            node = helper.make_node(
                'Constant', inputs=[], outputs=[name], value=tensor)
            nodes.append(node)
        return nodes

    def convert(self, program, save_dir):
        weight_nodes = self.convert_weights(program)
        op_nodes = list()
        input_nodes = list()
        output_nodes = list()

        unsupported_ops = set()

        print("Translating PaddlePaddle to ONNX...\n")
        for block in program.blocks:
            for i, op in enumerate(block.ops):
                sys.stdout.write(
                    "\rTotal:{}, Current:{} : {}                   ".format(
                        len(block.ops), i + 1, op.type))
                if not hasattr(self, op.type):
                    unsupported_ops.add(op.type)
                    continue
                if len(unsupported_ops) > 0:
                    continue
                node = getattr(self, op.type)(op, block)
                if op.type == 'feed':
                    input_nodes.append(node)
                elif op.type == 'fetch':
                    output_nodes.append(node)
                else:
                    if isinstance(node, list):
                        op_nodes = op_nodes + node
                    else:
                        op_nodes.append(node)

        if len(unsupported_ops) > 0:
            print("\nThere's {} ops are not supported yet".format(
                len(unsupported_ops)))
            for op in unsupported_ops:
                print("=========== {} ===========".format(op))
            return

        graph = helper.make_graph(
            nodes=weight_nodes + op_nodes,
            name='onnx_model_from_paddle',
            initializer=[],
            inputs=input_nodes,
            outputs=output_nodes)
        model = helper.make_model(graph, producer_name='X2Paddle')
        onnx.checker.check_model(model)

        if not os.path.isdir(save_dir):
            os.makedirs(save_dir)
        with open(os.path.join(save_dir, 'x2paddle_model.onnx'), 'wb') as f:
            f.write(model.SerializeToString())
        print("\nTranslated model saved in {}".format(
            os.path.join(save_dir, 'x2paddle_model.onnx')))