onnx_decoder.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.core.graph import GraphNode, Graph
from x2paddle.core.fluid_code import FluidCode
from onnx.checker import ValidationError
from onnx.checker import check_model
from onnx.utils import polish_model
from onnx.version_converter import convert_version
from onnx import helper
from onnx.helper import get_attribute_value, make_attribute
from onnx.shape_inference import infer_shapes
from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
from onnx.numpy_helper import to_array
from onnx import AttributeProto, TensorProto, GraphProto
from collections import OrderedDict as Dict
import onnx
import numpy as np
from copy import deepcopy
import logging as _logging

default_op_domain = 'ai.onnx'
_logger = _logging.getLogger(__name__)


class ONNXGraphNode(GraphNode):
    def __init__(self, layer, layer_name=None):
        if layer_name is None:
            super(ONNXGraphNode, self).__init__(layer, layer.name)
        else:
            super(ONNXGraphNode, self).__init__(layer, layer_name)
        self.layer_type = layer.op_type
        self.fluid_code = FluidCode()
        self.attr_map = self.get_attr_map()
        self.dtype_map = {1: "float32", 3: "int32", 9: "int64"}
        self.weight_inputs = list()
        self.out_shapes = list()
        self.dtype = None

    def get_attr_map(self):
        """
        convert ONNX node attributes to dict
        """
        return {
            attr.name: self.get_attribute_value2(attr)
            for attr in self.layer.attribute
        }

    @property
    def value(self):
        assert 'Constant' in self.layer_type, "Only Constant | ConstantOfShape node has value."
        attr = self.layer.attribute['value']
        if 'value' not in self.attr_map:
            return None
        return self.attr_map[name]

    def get_attribute_value2(self, attr):
        """
        get_attribute_value enhanced
        """
        if attr.type == onnx.AttributeProto.TENSOR:
            dtype = np.dtype(TENSOR_TYPE_TO_NP_TYPE[attr.t.data_type])
            data = attr.t.raw_data
            value = np.frombuffer(data,
                                  dtype=dtype,
                                  count=(len(data) // dtype.itemsize))
        elif attr.type == onnx.AttributeProto.STRING:
            value = attr.s
            value = value.decode() if isinstance(value, bytes) else value
        else:
            value = get_attribute_value(attr)
        return value

    def get_attr(self, name, default=None):
        """
        get_attribute_value from attr_map
        """
        if name not in self.attr_map:
            return default
        return self.attr_map[name]


class ONNXGraphDataNode(GraphNode):
    def __init__(self, layer, layer_name=None, is_global_input=False):
        if layer_name is None:
            super(ONNXGraphDataNode, self).__init__(layer, layer.name)
        else:
            super(ONNXGraphDataNode, self).__init__(layer, layer_name)
        if is_global_input:
            self.layer_type = 'place_holder'
        else:
            self.layer_type = 'create_parameter'
        self.layer_name = layer_name
        self.fluid_code = FluidCode()
        self.weight = None
        self.embeded_as = None

    @property
    def out_shapes(self):
        values = self.layer.type.tensor_type.shape.dim
        out_shapes = list()
        out_shapes.append([dim.dim_value for dim in values])
        return out_shapes

    @property
    def dtype(self):
        dtype = self.layer.type.tensor_type.elem_type
        return TENSOR_TYPE_TO_NP_TYPE[dtype]


class ONNXGraph(Graph):
    def __init__(self, graph, onnx_model):
        super(ONNXGraph, self).__init__(graph)
        self.onnx_model = onnx_model
        self.initializer = {}
        self.place_holder_nodes = list()
        self.get_place_holder_nodes()

        self.value_infos = self.inferred_model_value_info(graph)
        self.results_of_inference = dict()

    def get_inner_nodes(self):
        """
        generate inner node of ONNX model
        """
        inner_nodes = []
        if not isinstance(self.model, onnx.GraphProto):
            logger.error('graph is not a GraphProto instance')
            return
        for initializer in self.model.initializer:
            name = initializer.name
            inner_nodes.append(name)
        return inner_nodes

    def get_place_holder_nodes(self):
        """
        generate place_holder node of ONNX model
        """
        inner_nodes = self.get_inner_nodes()
        input_nodes = [value.name for value in self.model.input]
        for ipt_data in input_nodes:
            if ipt_data not in inner_nodes:
                self.place_holder_nodes.append(ipt_data)

    def is_place_holder_nodes(self, layer):
        """
        return layer is or not place_holder node
        """
        if layer in self.place_holder_nodes:
            return True
        return False

    def build(self):
        """
        build topo_sort of ONNX model
        """
        data_node = self.place_holder_nodes[0]
        value_info = self.value_infos[data_node]
        input_shape = value_info['shape']
        self.get_results_of_inference(self.onnx_model, input_shape)
        for layer in self.model.node:
            node = ONNXGraphNode(layer)
            self.node_map[layer.name] = node
            for opt in layer.output:
                if opt in self.value_infos:
                    value_info = self.value_infos[opt]
                    node.dtype = value_info['dtype']
                    node.out_shapes.append(value_info['shape'])
                else:
                    _, dtype, shape = self.get_dynamic_shape(opt)
                    node.dtype = dtype
                    node.out_shapes.append(shape)

        for layer in self.model.input:
            if layer.name not in self.node_map:
                is_place_holder = self.is_place_holder_nodes(layer.name)
                self.node_map[layer.name] = ONNXGraphDataNode(
                    layer,
                    layer_name=layer.name,
                    is_global_input=is_place_holder)

        #set data node's weight
        for name, weight in self.graph_weights(self.model):
            if name in self.node_map:
                if isinstance(self.node_map[name], ONNXGraphDataNode):
                    self.node_map[name].weight = weight
                    self.node_map[name].embeded_as = []

        #generate connection between nodes for topo
        for layer_name, node in self.node_map.items():
            if isinstance(node, ONNXGraphNode):
                for idx, in_node in enumerate(node.layer.input):
                    if in_node not in self.node_map:
                        raise Exception(
                            'input[{}] of node[{}] does not exist in node_map'.
                            format(in_node, layer_name))
                    else:
                        self.connect(in_node, layer_name)
        #generate topo
        super(ONNXGraph, self).build()

        self.input_nodes = self.place_holder_nodes

    def get_nodes(self, names, copy=False):
        """
        get nodes by more than one name
        """
        nodes = []
        for name in names:
            nodes.add(self.get_node(name, copy=copy))

    def graph_weights(self, graph):
        """
        generator for weights
        """

        if not isinstance(graph, onnx.GraphProto):
            logger.error('graph is not a GraphProto instance')
            return

        for initializer in graph.initializer:
            name = initializer.name
            weight = to_array(initializer)
            yield name, weight

    def inferred_model_value_info(self, graph):
        """
        collect value/type info for an ONNX model
        """
        assert isinstance(graph,
                          onnx.GraphProto), 'model is not a ModelProto instance'

        value_info = Dict()
        for item in graph.value_info:
            value_info[item.name] = {
                'dtype':
                TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
                'shape':
                [dim.dim_value for dim in item.type.tensor_type.shape.dim],
                'external': False
            }
        for item in graph.input:
            assert item.name not in value_info
            value_info[item.name] = {
                'dtype':
                TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
                'shape':
                [dim.dim_value for dim in item.type.tensor_type.shape.dim],
                'external': True
            }
        for item in graph.output:
            assert item.name not in value_info
            value_info[item.name] = {
                'dtype':
                TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
                'shape':
                [dim.dim_value for dim in item.type.tensor_type.shape.dim],
                'external': True
            }
        return value_info

    def get_results_of_inference(self, model, shape):
        try:
            import torch
            version = torch.__version__
            if '1.1.0' not in version:
                print("your model have dynamic graph, torch==1.1.0 is required")
                return
        except:
            print(
                "your model have dynamic graph, we use caff2 to inference graph, please use \"pip install torch==1.1.0\"."
            )
            return
        from x2paddle.decoder.onnx_backend import prepare

        np_images = np.random.rand(shape[0], shape[1], shape[2],
                                   shape[3]).astype('float32')

        outputs = []
        for node in model.graph.node:
            value_info = helper.make_tensor_value_info(node.name,
                                                       TensorProto.UNDEFINED,
                                                       [])
            outputs.append(value_info)

        while len(outputs) > 0:
            tmp_outputs = outputs[:254]
            model.graph.ClearField('output')
            model.graph.output.MergeFrom(tmp_outputs)
            prepared_backend = prepare(model,
                                       device='CPU',
                                       no_check_UNSAFE=True)
            res = prepared_backend.run(inputs=np_images)
            for idx, info in enumerate(tmp_outputs):
                self.results_of_inference[info.name] = res[idx]
            outputs = outputs[254:]
        return

    def get_dynamic_shape(self, layer):
        """
        get dynamic shape from caffe2.backend
        """
        output = self.results_of_inference[layer]
        return output.tolist(), output.dtype, output.shape


class ONNXDecoder(object):
    def __init__(self, onnx_model):
        model = onnx.load(onnx_model)
        print('model ir_version: {}, op version: {}'.format(
            model.ir_version, model.opset_import[0].version))
        if model.opset_import[0].version < 9:
            _logger.warning(
                'Now, onnx2paddle main support convert onnx model opset_verison == 9,'
                'opset_verison of your onnx model is %d < 9,'
                'some operator may cannot convert.',
                model.opset_import[0].version)

        check_model(model)
        model = onnx.shape_inference.infer_shapes(model)
        model = self.optimize_model_skip_op_for_inference(model)
        model = self.optimize_model_strip_initializer(model)
        self.standardize_variable_name(model.graph)

        self.model = model
        graph_def = model.graph
        self.onnx_graph = ONNXGraph(graph_def, model)
        self.onnx_graph.build()

    def build_value_refs(self, nodes):
        """
        build op reference of inputs and outputs
        """
        input_refs = Dict()
        output_refs = Dict()
        for idx, node in enumerate(nodes):
            for val_name in node.input:
                input_refs.setdefault(val_name, set()).add(idx)
            for val_name in node.output:
                output_refs.setdefault(val_name, set()).add(idx)
        return input_refs, output_refs

    def skip_node_forward(self, nodes, src_output_name, dst_input_name,
                          input_refs):
        """
        skip nodes between src_output_name -> dst_input_name and connect this pair
        """
        processed = 0
        for next_idx in input_refs[src_output_name]:
            next_node = nodes[next_idx]
            for val_idx, next_input_name in enumerate(next_node.input):
                if next_input_name == src_output_name:
                    next_node.input[val_idx] = dst_input_name
                    processed += 1
        return processed

    def skip_node_backward(self, nodes, src_input_name, dst_output_name,
                           output_refs):
        """
        skip nodes between dst_output_name -> src_input_name and connect this pair
        """
        processed = 0
        for prev_idx in output_refs[src_input_name]:
            prev_node = nodes[prev_idx]
            for val_idx, prev_output_name in enumerate(prev_node.output):
                if prev_output_name == src_input_name:
                    prev_node.output[val_idx] = dst_output_name
                    processed += 1
        return processed

    def optimize_model_skip_op_for_inference(self, model, op_list=None):
        """
        skip ops can be bypassed for inference
        """
        if op_list is None:
            op_list = ['Dropout']

        nodes = model.graph.node
        input_refs, output_refs = self.build_value_refs(nodes)
        ret = type(model)()
        ret.CopyFrom(model)
        ret_nodes = ret.graph.node
        nodes_to_remove = []
        for node_idx, node in enumerate(nodes):
            if not (node.domain == default_op_domain or node.domain == ''):
                continue
            op_type = node.op_type
            if not (op_type in op_list):
                continue
            if op_type in ['Dropout']:
                input_name = node.input[0]
                output_name = node.output[0]
            elif not (len(node.input) == 1 and len(node.output) == 1):
                print(
                    'currently only 1-input-1-output op supported, skip required %d: %s',
                    node_idx, node.op_type)
                continue
            else:
                input_name = node.input[0]
                output_name = node.output[0]

            if output_name in input_refs:
                processed = self.skip_node_forward(ret_nodes, output_name,
                                                   input_name, input_refs)
            elif input_name in output_refs:
                processed = self.skip_node_backward(ret_nodes, input_name,
                                                    output_name, output_refs)
            else:
                processed = -1
            if processed > 0:
                nodes_to_remove.append(node_idx)
                for value_info in ret.graph.value_info:
                    for output in node.output:
                        if value_info.name == output:
                            ret.graph.value_info.remove(value_info)

                print('skip op {}: {} -> {} -> {}'.format(
                    node_idx, input_name, node.op_type, output_name))
            elif processed == 0:
                print('weird, no node processed')
            else:
                print('standalone op {}: {} -> {} -> {} not skipped'.format(
                    node_idx, input_name, node.op_type, output_name))

        nodes_to_remove.sort(reverse=True)
        for node_idx in nodes_to_remove:
            ret_nodes.pop(node_idx)
        return ret

    def optimize_model_strip_initializer(self, model, keep_input_only=True):
        """
        strip weights for inference
        """
        nodes = model.graph.node
        input_refs, output_refs = self.build_value_refs(nodes)
        out_names = [val.name for val in model.graph.output]

        ret = type(model)()
        ret.CopyFrom(model)
        # strip initializers
        ret.graph.ClearField('initializer')
        ret_initializers = ret.graph.initializer
        for initializer in model.graph.initializer:
            name = initializer.name
            if name in input_refs:
                ret_initializers.add().CopyFrom(initializer)
            elif not keep_input_only and name in output_refs:
                ret_initializers.add().CopyFrom(initializer)
            else:
                dtype = TENSOR_TYPE_TO_NP_TYPE[initializer.data_type]

        # strip inputs
        ret.graph.ClearField('input')
        ret_inputs = ret.graph.input
        for item in model.graph.input:
            name = item.name
            if name in input_refs or name in out_names:
                ret_inputs.add().CopyFrom(item)
        return ret

    def make_variable_name(self, name):
        """
        make a valid code name for ParamAttr
        """

        if name == '':
            raise ValueError('name should not be empty')
        for s in ' .*?\\/-:':  #
            name = name.replace(s, '_')
        return '_' + name

    def standardize_variable_name(self, graph):
        """
        standardize variable name for paddle's code
        """
        for initializer in graph.initializer:
            initializer.name = self.make_variable_name(initializer.name)
        for ipt in graph.input:
            ipt.name = self.make_variable_name(ipt.name)
        for output in graph.output:
            output.name = self.make_variable_name(output.name)
        for item in graph.value_info:
            item.name = self.make_variable_name(item.name)
        for node in graph.node:
            node.name = node.output[0]
            node.name = self.make_variable_name(node.name)
            for i in range(len(node.input)):
                node.input[i] = self.make_variable_name(node.input[i])
            for i in range(len(node.output)):
                node.output[i] = self.make_variable_name(node.output[i])

    def split_model(self, model, outputs=None):
        """
        Takes a model and changes its outputs.
        """
        if outputs is None:
            raise RuntimeError("outputs is None")
        if outputs == model.graph.output[0].name:
            return model
        nodes = model.graph.node
        keep_nodes = []

        # all the nodes we need to keep.
        for node in nodes:
            if outputs in node.output:
                keep_nodes.append(node)
                break
            keep_nodes.append(node)

        infer_shapes = onnx.shape_inference.infer_shapes(model)

        var_out = []
        for value_info in infer_shapes.graph.value_info:
            if value_info.name == outputs:
                var_out.append(value_info)
                break

        graph = helper.make_graph(keep_nodes, model.graph.name,
                                  model.graph.input, var_out,
                                  model.graph.initializer)

        onnx_model = helper.make_model(graph)
        onnx_model.ir_version = model.ir_version
        onnx_model.producer_name = model.producer_name
        onnx_model.producer_version = model.producer_version
        onnx_model.domain = model.domain
        onnx_model.model_version = model.model_version
        onnx_model.doc_string = model.doc_string

        if len(onnx_model.graph.input) != len(model.graph.input):
            raise RuntimeError("Input mismatch {} != {}".format(
                len(onnx_model.input), len(model.input)))
        return onnx_model