tf_op_mapper_nhwc.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.

from x2paddle.decoder.tf_decoder import TFGraph
from x2paddle.core.op_mapper import OpMapper
from x2paddle.core.util import *
from x2paddle import program
from x2paddle import gen_name
import traceback
import math
import inspect
import numpy
import sys


# compute padding size for SAME mode
def get_same_padding(in_size, kernel_size, stride):
    new_size = int(math.ceil(in_size * 1.0 / stride))
    pad_size = (new_size - 1) * stride + kernel_size - in_size
    if pad_size < 0:
        pad_size = 0
    pad0 = int(pad_size / 2)
    pad1 = pad_size - pad0
    return [pad0, pad1]


class TFOpMapperNHWC(OpMapper):
    directly_map_ops = {
        'Relu': ['relu'],
        'Relu6': ['relu6'],
        'Abs': ['abs'],
        'Sigmoid': ['sigmoid'],
        'Exp': ['exp'],
        'Rsqrt': ['rsqrt'],
        'Sqrt': ['sqrt'],
        'swish_f32': ['swish'],
        'Tanh': ['tanh'],
        'Softplus': ['softplus'],
        'LeakyRelu': ['leaky_relu', {
            'alpha': 'alpha'
        }],
        'Floor': ['floor'],
        'Erf': ['erf']
    }
    elementwise_ops = {
        'Add': 'elementwise_add',
        'AddV2': 'elementwise_add',
        'RealDiv': 'elementwise_div',
        'Sub': 'elementwise_sub',
        'Maximum': 'elementwise_max',
        'Minimum': 'elementwise_min',
        'LessEqual': 'less_equal',
        'GreaterEqual': 'greater_equal',
        'Mul': 'elementwise_mul',
        'FloorDiv': 'elementwise_floordiv'
    }

    def __init__(self, decoder):
        super(TFOpMapperNHWC, self).__init__()
        self.decoder = decoder
        self.graph = decoder.tf_graph
        self.weights = dict()
        self.omit_nodes = list()
        self.used_custom_layers = dict()
        program.clear()

        not_placeholder = list()
        for name in self.graph.input_nodes:
            if self.graph.get_node(
                    name).layer_type != "Placeholder" and self.graph.get_node(
                        name
                    ).layer_type != "OneShotIterator" and self.graph.get_node(
                        name).layer_type != "IteratorV2":
                not_placeholder.append(name)
        for name in not_placeholder:
            idx = self.graph.input_nodes.index(name)
            del self.graph.input_nodes[idx]

        program.inputs = self.graph.input_nodes
        program.outputs = self.graph.output_nodes

        unsupported_ops = set()
        sys.stderr.write("Total nodes: {}\n".format(len(self.graph.topo_sort)))
        for i, node_name in enumerate(self.graph.topo_sort):
            sys.stderr.write("\rConverting node {} ...     ".format(i + 1))
            node = self.graph.get_node(node_name)
            op = node.layer_type
            if op in self.directly_map_ops:
                if len(unsupported_ops) > 0:
                    continue
                self.directly_map(node)
            elif op in self.elementwise_ops:
                if len(unsupported_ops) > 0:
                    continue
                self.elementwise_map(node)
            elif hasattr(self, op):
                if len(unsupported_ops) > 0:
                    continue
                func = getattr(self, op)
                try:
                    func(node)
                except Exception as e:
                    unsupported_ops.add(op)
                    print("\n{}\n".format(traceback.format_exc()))
            else:
                unsupported_ops.add(op)
        if len(unsupported_ops) > 0:
            print("\n========= {} OPs are not supported yet ===========".format(
                len(unsupported_ops)))
            for op in unsupported_ops:
                print("========== {} ============".format(op))
            sys.exit(-1)
        sys.stderr.write("\nDone!\n")

    def directly_map(self, node):
        assert node.layer_type in self.directly_map_ops
        op_info = self.directly_map_ops[node.layer_type]
        input = self.graph.get_node(node.layer.input[0])
        attr = dict()
        for param in op_info[1:]:
            tf_param_name = list(param.keys())[0]
            pd_param_name = list(param.values())[0]
            tf_param = node.get_attr(tf_param_name)
            attr[pd_param_name] = tf_param

        program.add_layer(
            kernel="fluid.layers.{}".format(op_info[0]),
            inputs={"x": input.name},
            outputs=[node.name],
            **attr)

    def elementwise_map(self, node):
        assert node.layer_type in self.elementwise_ops
        op_type = self.elementwise_ops[node.layer_type]
        x = self.graph.get_node(node.layer.input[0])
        y = self.graph.get_node(node.layer.input[1])
        program.add_layer(
            kernel="fluid.layers.{}".format(op_type),
            inputs={"x": x.name,
                    "y": y.name},
            outputs=[node.name])

    def Placeholder(self, node):
        shape = node.out_shapes[0]
        assert len(shape) != 0, "Unknown shape of input nodes[{}].".format(
            node.layer_name)
        dtype = node.dtype
        program.add_layer(
            kernel="fluid.data",
            inputs={},
            outputs=[node.name],
            dtype=string(dtype),
            shape=shape,
            name=string(node.name))

    def Const(self, node):
        shape = node.out_shapes[0]
        dtype = node.dtype
        value = node.value
        initializer = "Constant(0.0)"
        if len(shape) == 0:
            assert value.size == 1, "Unexpected situation happend"
            shape = [1]
            initializer = "Constant({})".format(value)

        program.parameters[node.name] = node.value
        program.add_layer(
            kernel="fluid.layers.create_parameter",
            inputs={},
            outputs=[node.name],
            dtype=string(dtype),
            shape=shape,
            name=string(node.name),
            default_initializer=initializer)

    def Transpose(self, node):
        input = self.graph.get_node(node.layer.input[0])
        perm = self.graph.get_node(node.layer.input[1])
        assert perm.layer_type == "Const", "Perm of transpose OP should be Const"
        perm = perm.value.tolist()

        program.add_layer(
            kernel="fluid.layers.transpose",
            inputs={"x": input.name},
            outputs=[node.name],
            perm=perm)

    def Fill(self, node):
        dims = self.graph.get_node(node.layer.input[0])
        input_value = self.graph.get_node(node.layer.input[1])
        inputs = dict()
        attr = dict()
        assert input_value.layer_type == "Const", "Value of fill OP should be Const"
        if dims.layer_type == "Const":
            attr["shape"] = dims.value.tolist()
        else:
            inputs["shape"] = dims.name
        attr["dtype"] = string(input_value.dtype)
        attr["value"] = input_value.value

        program.add_layer(
            "fluid.layers.fill_constant",
            inputs=inputs,
            outputs=[node.name],
            **attr)

    def DepthToSpace(self, node):
        input = self.graph.get_node(node.layer.input[0])

        block_size = node.get_attr("block_size")
        data_format = node.get_attr("data_format").decode()
        n, h, w, c = input.out_shapes[0]

        input_name = input.name
        if data_format == "NHWC":
            transpose_name = gen_name("depth_to_space", "transpose")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            input_name = transpose_name

        shape = [0, block_size * block_size, -1, h, w]
        reshape_name = gen_name("depth_to_space", "reshape")
        program.add_layer(
            kernel="fluid.layers.reshape",
            inputs={"x": input_name},
            outputs=[reshape_name],
            shape=shape)

        transpose_name = gen_name("depth_to_space", "transpose")
        program.add_layer(
            kernel="fluid.layers.transpose",
            inputs={"x": reshape_name},
            outputs=[transpose_name],
            perm=[0, 2, 1, 3, 4])

        reshape_name = gen_name("depth_to_space", "reshape")
        program.add_layer(
            kernel="fluid.layers.reshape",
            inputs={"x": transpose_name},
            outputs=[reshape_name],
            shape=[0, c, h, w])

        program.add_layer(
            kernel="fluid.layers.pixel_shuffle",
            inputs={"x": reshape_name},
            outputs=[node.name],
            upscale_factor=block_size)

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def MaxPool(self, node):
        input = self.graph.get_node(node.layer.input[0])

        k_size = node.get_attr("ksize")
        strides = node.get_attr("strides")
        data_format = node.get_attr("data_format").decode()
        pad_mode = node.get_attr("padding").decode()

        input_name = input.name
        if data_format == "NHWC":
            transpose_name = gen_name("max_pool", "transpose")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            strides = [strides[i] for i in [0, 3, 1, 2]]
            k_size = [k_size[i] for i in [0, 3, 1, 2]]
            input_name = transpose_name

        program.add_layer(
            kernel="fluid.layers.pool2d",
            inputs={"input": input_name},
            outputs=[node.name],
            pool_size=k_size[2:4],
            pool_type=string("max"),
            pool_stride=strides[2:4],
            pool_padding=string(pad_mode))

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def Conv2D(self, node):
        input = self.graph.get_node(node.layer.input[0])
        kernel = self.graph.get_node(node.layer.input[1])

        k_size = kernel.out_shapes[0]
        strides = node.get_attr("strides")
        dilations = node.get_attr("dilations")
        data_format = node.get_attr("data_format").decode()
        pad_mode = node.get_attr("padding").decode()
        if data_format == "NHWC":
            n, h, w, c = input.out_shapes[0]
        else:
            n, c, h, w = input.out_shapes[0]

        if kernel.layer_type == 'Const':
            kernel_value = kernel.value
            kernel_weight_name = kernel.layer_name.replace('/', '_')
        else:
            kernel_value = self.decoder.infer_tensor(kernel)
            if kernel.layer_type == 'Split':
                kernel_weight_name = "{}_{}_kernel".format(node.layer_name,
                                                           kernel.layer_name)
            else:
                kernel_weight_name = kernel.layer_name.replace('/', '_')
        program.parameters[kernel_weight_name] = numpy.transpose(kernel_value,
                                                                 (3, 2, 0, 1))

        input_name = input.name
        if data_format == "NHWC":
            strides = [strides[i] for i in [0, 3, 1, 2]]
            dilations = [dilations[i] for i in [0, 3, 1, 2]]
            transpose_name = gen_name("conv2d", "transpose")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            input_name = transpose_name

        if c == -1:
            attr = {"shape": [0, k_size[2], 0, 0]}
            node.fluid_code.add_layer(
                "reshape", inputs=input, output=input, param_attr=attr)
            program.add_layer(
                kernel="fluid.layers.reshape",
                inputs={"x": input_name},
                outputs=[input_name],
                shape=[0, k_size[2], 0, 0])

        program.add_layer(
            kernel="fluid.layers.conv2d",
            inputs={"input": input_name},
            outputs=[node.name],
            bias_attr=False,
            param_attr=string(kernel_weight_name),
            num_filters=k_size[3],
            filter_size=k_size[0:2],
            stride=strides[2:4],
            dilation=dilations[2:4],
            padding=string(pad_mode))

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def BiasAdd(self, node):
        input = self.graph.get_node(node.layer.input[0])
        bias = self.graph.get_node(node.layer.input[1])
        program.add_layer(
            kernel="fluid.layers.elementwise_add",
            inputs={"x": input.name,
                    "y": bias.name},
            outputs=[node.name])

    def FusedBatchNorm(self, node):
        input = self.graph.get_node(node.layer.input[0])
        gamma = self.graph.get_node(node.layer.input[1])
        beta = self.graph.get_node(node.layer.input[2])
        moving_mean = self.graph.get_node(node.layer.input[3])
        moving_var = self.graph.get_node(node.layer.input[4])
        data_format = node.get_attr("data_format").decode()

        assert gamma.layer_type == "Const"
        assert beta.layer_type == "Const"
        assert moving_mean.layer_type == "Const"
        assert moving_var.layer_type == "Const"

        input_name = input.name
        if data_format == "NHWC":
            transpose_name = gen_name("batch_norm", "transpose")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            input_name = transpose_name

        program.add_layer(
            kernel="fluid.layers.batch_norm",
            inputs={"input": input_name},
            outputs=[node.name],
            epsilon=node.get_attr("epsilon"),
            param_attr=string(gamma.name),
            bias_attr=string(beta.name),
            moving_mean_name=string(moving_mean.name),
            moving_variance_name=string(moving_var.name),
            is_test=True)

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def Mean(self, node):
        input = self.graph.get_node(node.layer.input[0])
        reduce_idx = self.graph.get_node(node.layer.input[1])
        assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
        dims = reduce_idx.value.tolist()
        keep_dims = node.get_attr("keep_dims")

        program.add_layer(
            kernel="fluid.layers.reduce_mean",
            inputs={"input": input.name},
            outputs=[node.name],
            dim=dims,
            keep_dim=keep_dims)

    def Reshape(self, node):
        input = self.graph.get_node(node.layer.input[0])
        param = self.graph.get_node(node.layer.input[1])
        if param.layer_type == "Const":
            shape = param.value.tolist()
            program.add_layer(
                kernel="fluid.layers.reshape",
                inputs={"x": input.name},
                outputs=[node.name],
                shape=shape)
        else:
            program.add_layer(
                kernel="fluid.layers.reshape",
                inputs={"x": input.name,
                        "shape": param.name},
                outputs=[node.name])
        if param.layer_type != "Const":
            out_shape = numpy.array(node.out_shapes[0])
            if (out_shape > 0).any():
                out_shape[out_shape < 0] = 0
                program.add_layer(
                    kernel="fluid.layers.reshape",
                    inputs={"x": node.name},
                    outputs=[node.name],
                    shape=out_shape.tolist())

    def Pad(self, node):
        input = self.graph.get_node(node.layer.input[0])
        paddings = self.graph.get_node(node.layer.input[1])
        assert paddings.layer_type == "Const", "Padding should be Const"
        paddings = paddings.value.flatten().tolist()

        if len(input.out_shapes[0]) == 4:
            if paddings[0] + paddings[1] + paddings[6] + paddings[7] == 0:
                new_padding = paddings[2:6]
                transpose_name = gen_name("pad", "transpose")
                program.add_layer(
                    kernel="fluid.layers.transpose",
                    inputs={"x": input.name},
                    outputs=[transpose_name],
                    perm=[0, 3, 1, 2])
                program.add_layer(
                    kernel="fluid.layers.pad2d",
                    inputs={"input": transpose_name},
                    outputs=[node.name],
                    paddings=new_padding)
                program.add_layer(
                    kernel="fluid.layers.transpose",
                    inputs={"x": node.name},
                    outputs=[node.name],
                    perm=[0, 2, 3, 1])
                return

        program.add_layer(
            kernel="fluid.layers.pad",
            inputs={"input": input.name},
            outputs=[node.name],
            paddings=paddings)

    def Squeeze(self, node):
        input = self.graph.get_node(node.layer.input[0])
        squeeze_dims = node.get_attr('squeeze_dims')
        program.add_layer(
            kernel="fluid.layers.squeeze",
            inputs={"input": input.name},
            outputs=[node.name],
            axes=squeeze_dims)

    def Softmax(self, node):
        input = self.graph.get_node(node.layer.input[0])
        axis = node.get_attr("axis")
        program.add_layer(
            kernel="fluid.layers.softmax",
            inputs={"input": input.name},
            outputs=[node.name],
            axis=axis)

    def Shape(self, node):
        input = self.graph.get_node(node.layer.input[0])
        program.add_layer(
            kernel="fluid.layers.shape",
            inputs={"input": input.name},
            outputs=[node.name])

    def ArgMax(self, node):
        input = self.graph.get_node(node.layer.input[0])
        axis = self.graph.get_node(node.layer.input[1])
        assert axis.layer_type == "Const", "ArgMax only support Const parameter"
        axis = axis.value
        program.add_layer(
            kernel="fluid.layers.argmax",
            inputs={"x": input.name},
            outputs=[node.name],
            axis=axis)

    def MatMul(self, node):
        x = self.graph.get_node(node.layer.input[0])
        y = self.graph.get_node(node.layer.input[1])
        transpose_a = node.get_attr('transpose_a')
        transpose_b = node.get_attr('transpose_b')
        if transpose_a is None:
            transpose_a = node.get_attr('adj_x')
        if transpose_b is None:
            transpose_b = node.get_attr('adj_y')
        program.add_layer(
            kernel="fluid.layers.matmul",
            inputs={"x": x.name,
                    "y": y.name},
            outputs=[node.name],
            transpose_x=transpose_a,
            transpose_y=transpose_b)

    def DepthwiseConv2dNative(self, node):
        input = self.graph.get_node(node.layer.input[0])
        kernel = self.graph.get_node(node.layer.input[1])
        assert kernel.layer_type == "Const", "Kernel of DepthwiseConv2DNative should be Const"

        in_shape = input.out_shapes[0]
        k_size = kernel.out_shapes[0]
        strides = node.get_attr("strides")
        dilations = node.get_attr("dilations")
        data_format = node.get_attr("data_format").decode()
        pad_mode = node.get_attr("padding").decode()

        program.parameters[kernel.layer_name.replace(
            '/', '_')] = numpy.transpose(kernel.value, (2, 3, 0, 1))

        input_name = input.name
        if data_format == "NHWC":
            in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
            strides = [strides[i] for i in [0, 3, 1, 2]]
            dilations = [dilations[i] for i in [0, 3, 1, 2]]
            transpose_name = gen_name('depthwise_conv2d', 'transpose')
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            input_name = transpose_name

        program.add_layer(
            kernel="fluid.layers.conv2d",
            inputs={"input": input_name},
            outputs=[node.name],
            num_filters=in_shape[1],
            filter_size=k_size[0:2],
            stride=strides[2:4],
            dilation=dilations[2:4],
            groups=k_size[3] * in_shape[1],
            padding=string(pad_mode),
            param_attr=string(kernel.layer_name),
            bias_attr=False)

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def AvgPool(self, node):
        input = self.graph.get_node(node.layer.input[0])

        k_size = node.get_attr("ksize")
        strides = node.get_attr("strides")
        data_format = node.get_attr("data_format").decode()
        pad_mode = node.get_attr("padding").decode()

        input_name = input.name
        if data_format == "NHWC":
            transpose_name = gen_name("avg_pool", "transpose")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            strides = [strides[i] for i in [0, 3, 1, 2]]
            k_size = [k_size[i] for i in [0, 3, 1, 2]]
            input_name = transpose_name

        program.add_layer(
            kernel="fluid.layers.pool2d",
            inputs={"input": input_name},
            outputs=[node.name],
            pool_size=k_size[2:4],
            pool_type=string("avg"),
            pool_stride=strides[2:4],
            pool_padding=string(pad_mode))

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def Pack(self, node):
        inputs = [self.graph.get_node(name) for name in node.layer.input]
        axis = node.get_attr("axis")
        program.add_layer(
            kernel="fluid.layers.stack",
            inputs={"x": [i.name for i in inputs]},
            outputs=[node.name],
            axis=axis)

    def ConcatV2(self, node):
        inputs = [self.graph.get_node(name) for name in node.layer.input[:-1]]
        axis = self.graph.get_node(node.layer.input[-1])
        assert axis.layer_type == "Const", "axis for ConcatV2 must be type Const"
        axis = axis.value
        if axis < 0:
            axis += len(inputs[0].out_shapes[0])
        program.add_layer(
            kernel="fluid.layers.concat",
            inputs={"input": [i.name for i in inputs]},
            outputs=[node.name],
            axis=axis)

    def StridedSlice(self, node):
        input = self.graph.get_node(node.layer.input[0])
        begin = self.graph.get_node(node.layer.input[1])
        end = self.graph.get_node(node.layer.input[2])
        strides = self.graph.get_node(node.layer.input[3])
        assert begin.layer_type == "Const"
        assert end.layer_type == "Const"
        assert strides.layer_type == "Const"
        strides = strides.value.tolist()
        assert len(set(strides)) == 1 and strides[
            0] == 1, "Only support strides be 1 in StridedSlice OP"

        begin = begin.value.tolist()
        end = end.value.tolist()

        for i in range(len(end)):
            if end[i] == 0:
                end[i] = 999999

        begin_mask = node.get_attr('begin_mask')
        end_mask = node.get_attr('end_mask')
        ellipsis_mask = node.get_attr('ellipsis_mask')
        new_axis_mask = node.get_attr('new_axis_mask')
        shrink_axis_mask = node.get_attr('shrink_axis_mask')

        assert ellipsis_mask == 0, "(OP:{} Name:{})Only support ellipsis_mask be 0[now: {}] n StridedSlice OP".format(
            node.layer_type, node.layer.name, ellipsis_mask)

        # TODO codes without validation
        # Use it carefully
        new_begin = list()
        new_end = list()
        new_axes = list()
        shrink_axes = list()
        for i, item in enumerate(begin):
            mask = (new_axis_mask >> i) & 1
            if mask != 0:
                new_axes.append(i)
                continue

            mask = (shrink_axis_mask >> i) & 1
            if mask != 0:
                shrink_axes.append(i)

            mask = (begin_mask >> i) & 1
            if mask != 0:
                new_begin.append(0)
            else:
                new_begin.append(item)

            mask = (end_mask >> i) & 1
            if mask != 0:
                new_end.append(999999)
            else:
                new_end.append(end[i])

        program.add_layer(
            kernel="fluid.layers.slice",
            inputs={"input": input.name},
            outputs=[node.name],
            axes=[i for i in range(len(new_begin))],
            starts=new_begin,
            ends=new_end)
        if len(new_axes) > 0:
            program.add_layer(
                kernel="fluid.layers.unsqueeze",
                inputs={"input": node.name},
                outputs=[node.name],
                axes=new_axes)
        if len(shrink_axes) > 0:
            if len(input.out_shapes[0]) + len(new_axes) <= 1:
                pass
            else:
                program.add_layer(
                    kernel="fluid.layers.unsqueeze",
                    inputs={"input": node.name},
                    outputs=[node.name],
                    axes=new_axes)

    def Split(self, node):
        dim = self.graph.get_node(node.layer.input[0])
        input = self.graph.get_node(node.layer.input[1])
        assert dim.layer_type == "Const"
        num_split = node.get_attr('num_split')
        dim = dim.value

        program.add_layer(
            kernel="fluid.layers.split",
            inputs={"input": input.name},
            outputs=[
                "{}_p{}".format(node.layer_name, i) for i in range(num_split)
            ],
            num_or_sections=num_split,
            dim=dim)

    def Slice(self, node):
        input = self.graph.get_node(node.layer.input[0])
        begin = self.graph.get_node(node.layer.input[1])
        size = self.graph.get_node(node.layer.input[2])

        inputs = {"x": input.name}
        attrs = {}
        if begin.layer_type == "Const":
            begin = begin.value.tolist()
            attrs['offsets'] = begin
        else:
            #             shape = begin.out_shapes[0]
            #             reshape_name = gen_name("slice", "reshape")
            #             program.add_layer(
            #                 kernel="fluid.layers.reshape",
            #                 inputs={"x": begin.name},
            #                 outputs=[reshape_name],
            #                 shape=shape)
            #             inputs['offsets'] = reshape_name
            begin = self.decoder.infer_tensor(begin).tolist()
            attrs['offsets'] = begin
        if size.layer_type == "Const":
            size = size.value.tolist()
            attrs['shape'] = size
        else:
            shape = size.out_shapes[0]
            reshape_name = gen_name("slice", "reshape")
            program.add_layer(
                kernel="fluid.layers.reshape",
                inputs={"x": size.name},
                outputs=[reshape_name],
                shape=shape)
            inputs['shape'] = reshape_name
        program.add_layer(
            kernel="fluid.layers.crop_tensor",
            inputs=inputs,
            outputs=[node.name],
            **attrs)

    def ResizeNearestNeighbor(self, node):
        input = self.graph.get_node(node.layer.input[0])
        resize_shape = self.graph.get_node(node.layer.input[1])
        data_format = "NHWC"
        inputs = {"input": input.name}
        attrs = {"align_corners": node.get_attr("align_corners")}

        if resize_shape.layer_type == "Const":
            resize_shape = resize_shape.value.tolist()
            attrs["out_shape"] = resize_shape
        else:
            shape = resize_shape.out_shapes[0]
            reshape_name = gen_name("resize_nearest", "reshape")
            program.add_layer(
                kernel="fluid.layers.reshape",
                inputs={"x": resize_shape.name},
                outputs=[reshape_name],
                shape=shape)
            inputs["out_shape"] = reshape_name

        if data_format == "NHWC":
            transpose_name = gen_name("resize_nearest", "reshape")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            inputs["input"] = transpose_name

        program.add_layer(
            kernel="fluid.layers.resize_nearest",
            inputs=inputs,
            outputs=[node.name],
            **attrs)

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def ResizeBilinear(self, node):
        input = self.graph.get_node(node.layer.input[0])
        resize_shape = self.graph.get_node(node.layer.input[1])
        data_format = "NHWC"
        inputs = {"input": input.name}
        attrs = {"align_corners": node.get_attr("align_corners")}

        if resize_shape.layer_type == "Const":
            resize_shape = resize_shape.value.tolist()
            attrs["out_shape"] = resize_shape
        else:
            shape = resize_shape.out_shapes[0]
            reshape_name = gen_name("resize_bilinear", "reshape")
            program.add_layer(
                kernel="fluid.layers.reshape",
                inputs={"x": resize_shape.name},
                outputs=[reshape_name],
                shape=shape)
            inputs["out_shape"] = reshape_name

        if data_format == "NHWC":
            transpose_name = gen_name("resize_bilinear", "reshape")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            inputs["input"] = transpose_name

        program.add_layer(
            kernel="fluid.layers.resize_bilinear",
            inputs=inputs,
            outputs=[node.name],
            **attrs)

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def Cast(self, node):
        input = self.graph.get_node(node.layer.input[0])
        dtype = node.dtype
        program.add_layer(
            kernel="fluid.layers.cast",
            inputs={"x": input.name},
            outputs=[node.name],
            dtype=string(dtype))

    def Sum(self, node):
        input = self.graph.get_node(node.layer.input[0])
        reduce_idx = self.graph.get_node(node.layer.input[1])
        assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
        keep_dims = node.get_attr("keep_dims")
        dim = reduce_idx.value.tolist()

        program.add_layer(
            kernel="fluid.layers.reduce_sum",
            inputs={"input": input.name},
            outputs=[node.name],
            dim=dim,
            keep_dim=keep_dims)

    def Max(self, node):
        input = self.graph.get_node(node.layer.input[0])
        reduce_idx = self.graph.get_node(node.layer.input[1])
        assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
        keep_dims = node.get_attr("keep_dims")
        dim = reduce_idx.value.tolist()
        program.add_layer(
            kernel="fluid.layers.reduce_max",
            inputs={"input": input.name},
            outputs=[node.name],
            dim=dim,
            keep_dim=keep_dims)

    def RandomUniform(self, node):
        shape = self.graph.get_node(node.layer.input[0])
        if shape.layer_type == "Const":
            shape = shape.value.tolist()
            program.add_layer(
                kernel="fluid.layers.uniform_random",
                inputs={},
                outputs=[node.name],
                shape=shape,
                min=0.0,
                max=0.9999)
        else:
            program.add_layer(
                kernel="fluid.layers.uniform_random",
                inputs={'shape': shape.name},
                outputs=[node.name],
                min=0.0,
                max=0.9999)

    def Conv2DBackpropInput(self, node):
        out_shape = self.graph.get_node(node.layer.input[0])
        kernel = self.graph.get_node(node.layer.input[1])
        input = self.graph.get_node(node.layer.input[2])

        assert kernel.layer_type == "Const", "Kernel of Conv2DBackpropInput should be Const"

        if out_shape.layer_type == "Const":
            out_shape = out_shape.value.tolist()
        else:
            out_shape = self.decoder.infer_shape_tensor(out_shape,
                                                        node.out_shapes[0])

        in_shape = input.out_shapes[0]
        if in_shape.count(-1) > 2:
            in_shape = self.decoder.infer_tensor(input).shape
        k_size = kernel.out_shapes[0]
        if k_size.count(-1) > 2:
            k_size = self.decoder.infer_tensor(kernel).shape

        pad_mode = node.get_attr("padding").decode()
        strides = node.get_attr("strides")
        dilations = node.get_attr("dilations")
        data_format = node.get_attr("data_format").decode()

        program.parameters[kernel.layer_name.replace(
            '/', '_')] = numpy.transpose(kernel.value, (3, 2, 0, 1))

        input_name = input.name
        if data_format == "NHWC":
            in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
            strides = [strides[i] for i in [0, 3, 1, 2]]
            dilations = [dilations[i] for i in [0, 3, 1, 2]]
            transpose_name = gen_name("conv2dbackpropinput", "transpose")
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": input.name},
                outputs=[transpose_name],
                perm=[0, 3, 1, 2])
            input_name = transpose_name

        program.add_layer(
            kernel="fluid.layers.conv2d_transpose",
            inputs={"input": input_name},
            outputs=[node.name],
            bias_attr=False,
            param_attr=string(kernel.layer_name),
            num_filters=k_size[2],
            filter_size=k_size[0:2],
            stride=strides[2:4],
            dilation=dilations[2:4],
            padding=string(pad_mode),
            output_size=out_shape[1:3])

        if data_format == "NHWC":
            program.add_layer(
                kernel="fluid.layers.transpose",
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])

    def Tile(self, node):
        input = self.graph.get_node(node.layer.input[0])
        expand_times = self.graph.get_node(node.layer.input[1])
        inputs = {"x": input.name}
        attr = dict()
        if expand_times.layer_type == "Const":
            expand_times = expand_times.value.tolist()
            attr["expand_times"] = expand_times
        else:
            inputs["expand_times"] = expand_times.name

        program.add_layer(
            kernel="fluid.layers.expand",
            inputs=inputs,
            outputs=[node.name],
            **attr)

    def Range(self, node):
        start = self.graph.get_node(node.layer.input[0])
        limit = self.graph.get_node(node.layer.input[1])
        delta = self.graph.get_node(node.layer.input[2])
        inputs = dict()
        attr = dict()

        if start.layer_type == "Const":
            attr["start"] = start.value
        else:
            inputs["start"] = start.name
        if limit.layer_type == "Const":
            attr["end"] = limit.value
        else:
            inputs["end"] = limit.name
        if delta.layer_type == "Const":
            attr["step"] = delta.value
        else:
            inputs["step"] = delta.name
        attr["dtype"] = string(node.dtype)

        program.add_layer(
            kernel="fluid.layers.range",
            inputs=inputs,
            outputs=[node.name],
            **attr)

    def SquaredDifference(self, node):
        x = self.graph.get_node(node.layer.input[0])
        y = self.graph.get_node(node.layer.input[1])
        inputs = {"x": x.name, "y": y.name}
        program.add_layer(
            "fluid.layers.elementwise_sub", inputs=inputs, outputs=[node.name])
        inputs = {"x": node.name, "y": node.name}
        program.add_layer(
            "fluid.layers.elementwise_mul", inputs=inputs, outputs=[node.name])