onnx_decoder.py 20.6 KB
Newer Older
C
update  
channingss 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
#   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 onnx.checker import ValidationError
from onnx.checker import check_model
C
Channingss 已提交
18
from onnx import helper, shape_inference
C
update  
channingss 已提交
19 20 21 22
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
C
channingss 已提交
23
from onnx import AttributeProto, TensorProto, GraphProto
C
update  
channingss 已提交
24 25
from collections import OrderedDict as Dict
import onnx
C
channingss 已提交
26
from onnx.helper import ValueInfoProto
C
update  
channingss 已提交
27 28
import numpy as np
from copy import deepcopy
C
channingss 已提交
29
import logging as _logging
C
channingss 已提交
30
import os
S
SunAhong1993 已提交
31
import copy
C
update  
channingss 已提交
32 33

default_op_domain = 'ai.onnx'
C
channingss 已提交
34
_logger = _logging.getLogger(__name__)
C
update  
channingss 已提交
35 36 37 38 39 40 41 42 43 44


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.attr_map = self.get_attr_map()
C
channingss 已提交
45
        self.out_shapes = list()
C
update  
channingss 已提交
46
        self.dtype = None
C
channingss 已提交
47
        self.which_child = {}
C
update  
channingss 已提交
48

Y
yeliang2258 已提交
49 50 51 52 53 54 55 56 57 58 59 60
    def get_input_index(self, input_name):
        """
        get the index of input_name in layer.input
        -1 means input_name is not in the input
        """
        index = -1
        for i in range(len(self.layer.input)):
            if input_name == self.layer.input[i]:
                index = i
                break
        return index

C
update  
channingss 已提交
61 62 63 64 65
    def get_attr_map(self):
        """
        convert ONNX node attributes to dict
        """
        return {
66
            attr.name: self.get_attribute_value(attr)
C
update  
channingss 已提交
67 68 69 70 71
            for attr in self.layer.attribute
        }

    @property
    def value(self):
C
channingss 已提交
72 73 74
        assert 'Constant' in self.layer_type, "Only Constant | ConstantOfShape node has value."
        if 'value' not in self.attr_map:
            return None
C
channingss 已提交
75
        return self.attr_map['value']
S
SunAhong1993 已提交
76

S
SunAhong1993 已提交
77 78 79 80 81
    @property
    def name(self):
        if hasattr(self, 'index'):
            return "{}_p{}".format(self.layer_name, self.index)
        return self.layer_name
C
update  
channingss 已提交
82

83
    def get_attribute_value(self, attr):
C
update  
channingss 已提交
84 85 86 87 88 89
        """
        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
90 91
            value = np.frombuffer(
                data, dtype=dtype, count=(len(data) // dtype.itemsize))
C
update  
channingss 已提交
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
        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]

C
Channingss 已提交
107
    def output(self, index=0):
S
SunAhong1993 已提交
108 109 110 111
        if index > 0 and len(self.layer.output) <= index:
            raise IndexError(
                'Output numbers of Node:{} is {} <= index:{}'.format(
                    self.layer_name, len(self.layer.output), index))
112
        return self.layer.output[index]
C
Channingss 已提交
113

C
update  
channingss 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127

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.weight = None
        self.embeded_as = None
C
channingss 已提交
128
        self.which_child = {}
C
update  
channingss 已提交
129 130 131

    @property
    def out_shapes(self):
C
channingss 已提交
132 133 134
        if isinstance(self.layer, ValueInfoProto):
            values = self.layer.type.tensor_type.shape.dim
            out_shapes = list()
S
SunAhong1993 已提交
135 136 137 138 139 140 141
            shape = list()
            for dim in values:
                if dim.dim_value == 0:
                    shape.append(-1)
                else:
                    shape.append(dim.dim_value)
            out_shapes.append(shape)
C
channingss 已提交
142
            return out_shapes
S
SunAhong1993 已提交
143 144 145 146 147 148 149 150 151 152 153
        elif isinstance(self.layer, TensorProto):
            values = self.layer.dims
            out_shapes = list()
            shape = list()
            for dim in values:
                if dim == 0:
                    shape.append(-1)
                else:
                    shape.append(dim)
            out_shapes.append(shape)
            return out_shapes
C
channingss 已提交
154 155 156 157 158
        else:
            values = self.layer.dims
            out_shapes = list()
            out_shapes.append(values)
            return out_shapes
S
SunAhong1993 已提交
159

S
SunAhong1993 已提交
160 161 162
    @property
    def name(self):
        return self.layer_name
C
update  
channingss 已提交
163 164 165

    @property
    def dtype(self):
C
channingss 已提交
166 167 168 169 170 171
        if isinstance(self.layer, ValueInfoProto):
            dtype = self.layer.type.tensor_type.elem_type
            return TENSOR_TYPE_TO_NP_TYPE[dtype]
        else:
            dtype = self.layer.data_type
            return TENSOR_TYPE_TO_NP_TYPE[dtype]
C
update  
channingss 已提交
172 173 174


class ONNXGraph(Graph):
175
    def __init__(self, onnx_model, input_shape_dict):
176 177
        super(ONNXGraph, self).__init__(onnx_model)
        self.fixed_input_shape = {}
178 179 180
        if input_shape_dict is not None:
            for k, v in eval(input_shape_dict).items():
                self.fixed_input_shape["x2paddle_" + k] = v
C
update  
channingss 已提交
181 182
        self.initializer = {}
        self.place_holder_nodes = list()
183 184
        self.value_infos = {}
        self.graph = onnx_model.graph
C
update  
channingss 已提交
185
        self.get_place_holder_nodes()
W
wjj19950828 已提交
186 187 188 189
        print("Shape inferencing ...")
        onnx_model = shape_inference.infer_shapes(onnx_model)
        self.graph = onnx_model.graph
        print("Shape inferenced.")
190 191 192
        self.build()
        self.collect_value_infos()
        self.allocate_shapes()
S
SunAhong1993 已提交
193
        self.graph_name = "ONNXModel"
C
update  
channingss 已提交
194 195 196 197 198 199

    def get_inner_nodes(self):
        """
        generate inner node of ONNX model
        """
        inner_nodes = []
200
        if not isinstance(self.graph, onnx.GraphProto):
C
update  
channingss 已提交
201 202
            logger.error('graph is not a GraphProto instance')
            return
203
        for initializer in self.graph.initializer:
C
update  
channingss 已提交
204 205 206 207
            name = initializer.name
            inner_nodes.append(name)
        return inner_nodes

208 209 210 211 212 213 214 215 216
    def get_symbolic_shape(self, dims):
        shape = []
        for dim in dims:
            if dim.HasField('dim_param'):
                shape.append(dim.dim_param)
            else:
                shape.append(dim.dim_value)
        return shape

C
update  
channingss 已提交
217 218 219 220 221
    def get_place_holder_nodes(self):
        """
        generate place_holder node of ONNX model
        """
        inner_nodes = self.get_inner_nodes()
222 223 224
        for ipt_vi in self.graph.input:
            if ipt_vi.name not in inner_nodes:
                self.place_holder_nodes.append(ipt_vi.name)
C
update  
channingss 已提交
225

C
channingss 已提交
226 227 228 229
    def get_output_nodes(self):
        """
        generate output_nodes node of ONNX model
        """
C
Channingss 已提交
230
        self.output_nodes = [value.name for value in self.graph.output]
C
channingss 已提交
231

C
update  
channingss 已提交
232 233 234 235 236 237 238 239 240 241 242 243
    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
        """
244
        for layer in self.graph.node:
C
channingss 已提交
245 246
            node = ONNXGraphNode(layer)
            self.node_map[layer.name] = node
C
update  
channingss 已提交
247

248
        for layer in self.graph.input:
C
update  
channingss 已提交
249 250 251 252 253 254
            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)
C
channingss 已提交
255

C
update  
channingss 已提交
256
        #set data node's weight
257
        for initializer in self.graph.initializer:
C
channingss 已提交
258 259
            name = initializer.name
            weight = to_array(initializer)
C
update  
channingss 已提交
260 261 262 263
            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 = []
C
channingss 已提交
264
            else:
265 266
                self.node_map[name] = ONNXGraphDataNode(
                    initializer, layer_name=name, is_global_input=False)
C
channingss 已提交
267 268
                self.node_map[name].weight = weight
                self.node_map[name].embeded_as = []
C
update  
channingss 已提交
269 270 271 272

        #generate connection between nodes for topo
        for layer_name, node in self.node_map.items():
            if isinstance(node, ONNXGraphNode):
273
                self.build_connection(layer_name, node)
C
channingss 已提交
274
        #generate topo
C
update  
channingss 已提交
275 276
        super(ONNXGraph, self).build()

S
SunAhong1993 已提交
277
        self.input_nodes = copy.deepcopy(self.place_holder_nodes)
C
update  
channingss 已提交
278

279 280 281 282 283 284 285 286 287
    def build_connection(self, layer_name, node):
        """
        find connection for nodes
        """
        for idx, in_node in enumerate(node.layer.input):
            if in_node == '':
                continue
            if in_node not in self.node_map:
                flag = 0
288
                for nd in self.graph.node:
289 290 291 292
                    for idx, opt in enumerate(nd.output):
                        if opt == in_node:
                            self.connect(nd.name, layer_name)
                            flag = 1
S
fix  
SunAhong1993 已提交
293 294 295 296 297
                            if nd.name in node.which_child:
                                for n_i, n_ipt in enumerate(node.inputs):
                                    if first_i == n_i:
                                        continue
                                    if n_ipt == nd.name:
S
SunAhong1993 已提交
298 299
                                        new_nd_name = "{}/{}".format(nd.name,
                                                                     n_i)
S
fix  
SunAhong1993 已提交
300 301 302 303 304 305
                                        if new_nd_name not in node.which_child:
                                            node.which_child[new_nd_name] = idx
                                            break
                            else:
                                first_i = node.inputs.index(nd.name)
                                node.which_child[nd.name] = idx
306 307 308 309 310 311 312 313 314 315 316
                            self.node_map[nd.name].index = 0
                            break
                    if flag == 1:
                        break
                if flag == 0:
                    raise Exception(
                        'input[{}] of node[{}] does not exist in node_map'.
                        format(in_node, layer_name))
            else:
                self.connect(in_node, layer_name)

C
channingss 已提交
317 318
    def get_input_node(self, node, idx=0, copy=False):
        if len(node.which_child) == 0:
C
channingss 已提交
319 320
            ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
            return ipt_node
C
channingss 已提交
321 322
        else:
            ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
S
fix  
SunAhong1993 已提交
323 324 325 326 327 328
            new_ipt_name = "{}/{}".format(ipt_node.layer_name, idx)
            if new_ipt_name in node.which_child:
                ipt_node.index = node.which_child[new_ipt_name]
            else:
                if ipt_node.layer_name in node.which_child:
                    ipt_node.index = node.which_child[ipt_node.layer_name]
S
SunAhong1993 已提交
329

C
channingss 已提交
330
            return ipt_node
C
update  
channingss 已提交
331

332
    def graph_weights(self):
C
update  
channingss 已提交
333 334 335 336
        """
        generator for weights
        """

337
        if not isinstance(self.graph, onnx.GraphProto):
C
update  
channingss 已提交
338 339 340
            logger.error('graph is not a GraphProto instance')
            return

341
        for initializer in self.graph.initializer:
C
update  
channingss 已提交
342 343 344 345
            name = initializer.name
            weight = to_array(initializer)
            yield name, weight

346
    def collect_value_infos(self):
C
channingss 已提交
347 348 349
        """
        collect value/type info for an ONNX model
        """
350
        assert isinstance(self.graph,
C
channingss 已提交
351 352
                          onnx.GraphProto), 'model is not a ModelProto instance'

353 354
        for item in self.graph.value_info:
            self.value_infos[item.name] = {
C
channingss 已提交
355 356 357 358 359 360
                '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
            }
361 362 363 364 365 366 367 368 369 370 371 372 373

    def allocate_shapes(self):
        """
        run shape inference
        """
        for layer in self.graph.node:
            node = self.node_map[layer.name]
            for opt in layer.output:
                if opt in self.value_infos:
                    value_info = self.value_infos[opt]
                    #if len(value_info['shape']) == 0 or value_info[
                    #        'dtype'] is None or 0 in value_info['shape']:
                    #    #TODO add node shape inference
374 375 376 377 378
                    shape = value_info['shape']
                    for idx in range(len(shape)):
                        if shape[idx] == 0:
                            shape[idx] = -1
                    node.out_shapes.append(shape)
379 380 381
                    node.dtype = value_info['dtype']
                else:
                    node.out_shapes.append([])
C
channingss 已提交
382

C
update  
channingss 已提交
383 384

class ONNXDecoder(object):
385
    def __init__(self, onnx_model, input_shape_dict, enable_onnx_checker):
386
        onnx_model = onnx.load(onnx_model)
C
update  
channingss 已提交
387
        print('model ir_version: {}, op version: {}'.format(
388 389 390
            onnx_model.ir_version, onnx_model.opset_import[0].version))
        self.op_set = onnx_model.opset_import[0].version

W
wjj19950828 已提交
391 392
        if enable_onnx_checker:
            check_model(onnx_model)
393 394 395

        onnx_model = self.optimize_model_skip_op(onnx_model)
        onnx_model = self.optimize_node_name(onnx_model)
396
        self.graph = ONNXGraph(onnx_model, input_shape_dict)
C
update  
channingss 已提交
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438

    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

439
    def optimize_model_skip_op(self, model, op_list=None):
C
update  
channingss 已提交
440 441 442
        """
        skip ops can be bypassed for inference
        """
443
        nodes = model.graph.node
C
update  
channingss 已提交
444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478
        if op_list is None:
            op_list = ['Dropout']
        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)
C
channingss 已提交
479 480 481 482 483
                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)

C
update  
channingss 已提交
484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533
                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')
W
WJJ1995 已提交
534
        for s in ' .*?\\/-:;':
C
update  
channingss 已提交
535
            name = name.replace(s, '_')
536 537
        return 'x2paddle_' + name

538
    def optimize_node_name(self, model):
C
update  
channingss 已提交
539 540 541
        """
        standardize variable name for paddle's code
        """
542
        graph = model.graph
C
update  
channingss 已提交
543 544 545 546 547 548 549 550 551
        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:
C
channingss 已提交
552
            node.name = node.output[0]
W
wjj19950828 已提交
553 554 555
            # Avoid topological sort errors caused by :: in the name
            if "::" in node.name and len(node.output) > 1:
                node.name = node.name.replace('::', '_')
W
wjj19950828 已提交
556 557
            if ":" in node.name and len(
                    node.output) > 1 and node.op_type != "LSTM":
W
WJJ1995 已提交
558
                node.name = node.name.split(':')[0]
C
update  
channingss 已提交
559 560
            node.name = self.make_variable_name(node.name)
            for i in range(len(node.input)):
561 562 563 564
                if node.input[i] == '':
                    continue
                else:
                    node.input[i] = self.make_variable_name(node.input[i])
C
update  
channingss 已提交
565 566
            for i in range(len(node.output)):
                node.output[i] = self.make_variable_name(node.output[i])
S
fix  
SunAhong1993 已提交
567
        return model