opset.py 58.3 KB
Newer Older
1
# Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
C
update  
channingss 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14
#
# 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.

C
Channingss 已提交
15
from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode
C
update  
channingss 已提交
16 17 18
from x2paddle.core.graph import GraphNode
from x2paddle.core.fluid_code import Layer
from x2paddle.core.fluid_code import FluidCode
C
channingss 已提交
19
from x2paddle.core.util import string
C
Channingss 已提交
20
from x2paddle.op_mapper.onnx2paddle.opset9.custom_layer import *
C
Channingss 已提交
21
from functools import reduce
C
update  
channingss 已提交
22
import numpy as np
C
channingss 已提交
23
import onnx
C
channingss 已提交
24
import onnx.numpy_helper as numpy_helper
C
channingss 已提交
25
from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
C
update  
channingss 已提交
26
import logging as _logging
27
from collections import OrderedDict
C
channingss 已提交
28
import math
C
channingss 已提交
29 30
import os
import shutil
31

C
update  
channingss 已提交
32 33 34 35
_logger = _logging.getLogger(__name__)


def _const_weight_or_none(node):
C
channings 已提交
36
    if 'Constant' in node.layer_type:
C
channingss 已提交
37
        return node.value
C
update  
channingss 已提交
38 39 40 41 42
    if isinstance(node, ONNXGraphDataNode):
        return node.weight
    return None


C
Channingss 已提交
43
def _get_same_padding(in_size, kernel_size, stride):
C
channingss 已提交
44 45 46 47 48 49 50
    new_size = int(math.ceil(in_size * 1.0 / stride))
    pad_size = (new_size - 1) * stride + kernel_size - in_size
    pad0 = int(pad_size / 2)
    pad1 = pad_size - pad0
    return [pad0, pad1]


51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
def print_mapping_info(func):
    def run_mapping(*args, **kwargs):
        node = args[1]
        try:
            res = func(*args, **kwargs)
        except:
            print("convert failed node:{}, op_type is {}".format(
                node.layer_name[9:], node.layer_type))
            raise
        else:
            #print("convert successfully node:{}, op_type is {}".format(
            #    node.layer_name[9:], node.layer_type))
            return res

    return run_mapping


C
Channingss 已提交
68
class OpSet9():
69 70 71 72 73
    elementwise_ops = {
        'Add': 'elementwise_add',
        'Div': 'elementwise_div',
        'Sub': 'elementwise_sub',
        'Mul': 'elementwise_mul',
R
root 已提交
74 75
        'Pow': 'elementwise_pow',
    }
76

77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
    default_op_mapping_field_values = OrderedDict()
    default_op_mapping_field_values['FLUID_OP'] = ''
    default_op_mapping_field_values['FLUID_INPUT_ARGS'] = None
    default_op_mapping_field_values['FLUID_OUTPUT_ARGS'] = None
    default_op_mapping_field_values['ATTR_MAPPING'] = dict()
    default_op_mapping_field_values['DEFAULTS'] = dict()
    default_op_mapping_field_values['INPUT_PERM'] = None
    default_op_mapping_field_values['OUTPUT_PERM'] = None
    default_op_mapping_field_values['FILL_NAME_FIELD'] = True

    default_op_mapping = {
        'Shape': ['shape', ['X'], ['Out']],
        'Clip': [
            'clip', ['X'], ['Out'], dict(), dict(
                min=(np.asarray(
                    [255, 255, 127, 255], dtype=np.uint8).view(np.float32)[0]),
                max=(np.asarray(
                    [255, 255, 127, 127], dtype=np.uint8).view(np.float32)[0]),
            )
        ],
        'Erf': ['erf', ['X'], ['Out']],
        'Ceil': ['ceil', ['X'], ['Out']],
        'ReduceMean': [
            'reduce_mean', ['X'], ['Out'], dict(
                axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
        ],
        'ReduceSum': [
            'reduce_sum', ['X'], ['Out'], dict(
                axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
        ],
        'ReduceMin': [
            'reduce_min', ['X'], ['Out'], dict(
                axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
        ],
C
Channingss 已提交
111 112 113 114
        'ReduceMax': [
            'reduce_max', ['X'], ['Out'], dict(
                axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
        ],
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138
        #active function
        'Relu': ['relu', ['X'], ['Out']],
        'LeakyRelu': ['leaky_relu', ['X'], ['Out'], dict(), dict(alpha=.01)],
        'Elu': ['elu', ['X'], ['Out'], dict(), dict(alpha=1.)],
        'ThresholdedRelu': [
            'thresholded_relu', ['X'], ['Out'], dict(alpha='threshold'),
            dict(alpha=1.)
        ],
        'Tanh': ['tanh', ['X'], ['Out']],
        'Sigmoid': ['sigmoid', ['X'], ['Out']],
        'HardSigmoid': [
            'hard_sigmoid', ['X'], ['Out'], dict(
                alpha='slope', beta='offset'), dict(
                    slope=.2, offset=.5)
        ],
        'Softsign': ['softsign', ['X'], ['Out']],
        'Softplus': ['softplus', ['X'], ['Out']],
        'Exp': ['exp', ['X'], ['Out']],
        'Softmax': ['softmax', ['X'], ['Out'], dict(), dict(axis=1)],
        'Sqrt': ['sqrt', ['X'], ['Out']],
        'Floor': ['floor', ['X'], ['Out']],
        'Abs': ['abs', ['X'], ['Out']],
    }

C
Channingss 已提交
139
    default_ioa_constraint = {}
140 141

    def __init__(self, decoder):
C
Channingss 已提交
142
        super(OpSet9, self).__init__()
143
        self.graph = decoder.graph
C
update  
channingss 已提交
144 145 146
        self.input_shapes = []
        self.weights = dict()
        self.omit_nodes = list()
C
channingss 已提交
147
        self.used_custom_layers = dict()
R
root 已提交
148

149
    @print_mapping_info
C
channingss 已提交
150
    def directly_map(self, node, name='', *args, **kwargs):
C
update  
channingss 已提交
151 152 153 154
        inputs = node.layer.input
        outputs = node.layer.output
        op_type = node.layer_type
        attrs = node.attr_map
155 156 157
        info = self.default_op_mapping[op_type]
        info.extend(
            list(self.default_op_mapping_field_values.values())[len(info):])
C
update  
channingss 已提交
158 159 160 161 162 163 164 165
        (
            fluid_op,
            fluid_input_args,
            fluid_output_args,
            attr_mapping,
            default_attrs,
            input_perm,
            output_perm,
166
            fill_name_field, ) = info
C
update  
channingss 已提交
167

168 169
        if fluid_op in self.default_ioa_constraint:
            for predicate, message in self.default_ioa_constraint[fluid_op]:
C
update  
channingss 已提交
170 171 172 173 174 175 176 177 178 179 180 181
                assert predicate(inputs, outputs, attrs), message

        mapped_attrs = {
            attr_mapping.get(key, key): value
            for key, value in attrs.items()
        }
        if '' in mapped_attrs:
            mapped_attrs.pop('')
        if '_' in mapped_attrs:
            mapped_attrs.pop('_')
        fluid_attrs = default_attrs.copy()
        fluid_attrs.update(mapped_attrs)
C
channingss 已提交
182
        inputs = inputs if input_perm is None else list(
C
update  
channingss 已提交
183
            map(lambda i: inputs[i], input_perm))
C
channingss 已提交
184 185 186 187
        val_inps = []
        for idx, ipt in enumerate(inputs):
            val_inps.append(self.graph.get_input_node(node, idx=idx, copy=True))

C
update  
channingss 已提交
188 189 190
        val_outs = outputs if output_perm is None else list(
            map(lambda i: outputs[i], output_perm))
        attr = fluid_attrs
C
channingss 已提交
191
        assert len(val_inps) == 1, 'directly_map error with multi inputs'
192
        if fluid_op not in ['shape', 'erf']:
C
update  
channingss 已提交
193
            attr['name'] = string(node.layer_name)
194 195 196 197 198 199 200 201 202 203
        node.fluid_code.add_layer(
            fluid_op, inputs=val_inps[0], output=val_outs[0], param_attr=attr)
        if fluid_op in ['shape']:
            node.fluid_code.add_layer(
                'cast',
                inputs=val_outs[0],
                output=val_outs[0],
                param_attr={'dtype': string('int64')})

    @print_mapping_info
C
channingss 已提交
204 205 206
    def deal_custom_layer(self, node):
        op = node.layer_type
        custom_code, func = make_custom_layer(node)
C
channingss 已提交
207
        child_func_code, child_func = make_custom_child_func(node)
C
channingss 已提交
208 209 210
        params = get_params(node.layer, node.layer_type)
        arg_names, kwargs = set_args(func, params)
        kwargs['name'] = string(node.layer_name)
211 212 213 214 215 216
        node.fluid_code.add_layer(
            func.__code__.co_name,
            inputs=node.inputs,
            output=node,
            param_attr=kwargs,
            is_custom_layer=True)
C
channingss 已提交
217 218
        if op not in self.used_custom_layers:
            self.used_custom_layers[op] = custom_code
C
channingss 已提交
219
            if op + '_child_func' not in self.used_custom_layers:
C
channingss 已提交
220 221 222
                if child_func_code is not None:
                    self.used_custom_layers[op +
                                            '_child_func'] = child_func_code
R
root 已提交
223

224
    @print_mapping_info
225 226 227
    def elementwise_map(self, node):
        assert node.layer_type in self.elementwise_ops
        op_type = self.elementwise_ops[node.layer_type]
R
root 已提交
228

229 230 231 232
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_y = self.graph.get_input_node(node, idx=1, copy=True)
        val_y_shape = val_y.out_shapes[0]
        val_x_shape = val_x.out_shapes[0]
R
root 已提交
233 234

        if len(val_x_shape) < len(val_y_shape):
235
            val_x, val_y = val_y, val_x
236
            val_y_shape, val_x_shape = val_x_shape, val_y_shape
237 238 239

        str_y_shape = ','.join(str(e) for e in val_y_shape)
        str_x_shape = ','.join(str(e) for e in val_x_shape)
240
        slice_idx = 0
241 242 243 244 245 246
        if str_y_shape not in str_x_shape:
            for dim in val_y_shape:
                if dim == 1:
                    slice_idx += 1
                else:
                    break
247 248 249 250 251 252 253 254
        attr = {"name": string(node.layer_name)}
        if slice_idx < len(val_y_shape) and slice_idx > 0:
            val_y_reshaped = val_y_shape[slice_idx:]
            var_y_reshaped = val_y.layer_name + '_reshaped'
            attr_reshaped = {
                'shape': val_y_reshaped,
                'name': string(var_y_reshaped)
            }
255 256 257 258 259
            node.fluid_code.add_layer(
                'reshape',
                inputs=val_y,
                output=var_y_reshaped,
                param_attr=attr_reshaped)
260
            inputs = {'x': val_x, 'y': var_y_reshaped}
261 262
            node.fluid_code.add_layer(
                op_type, inputs=inputs, output=node, param_attr=attr)
263 264
        else:
            inputs = {'x': val_x, 'y': val_y}
265 266
            node.fluid_code.add_layer(
                op_type, inputs=inputs, output=node, param_attr=attr)
C
channingss 已提交
267

268
    @print_mapping_info
C
update  
channingss 已提交
269
    def place_holder(self, node):
C
channingss 已提交
270
        self.input_shapes.append(node.out_shapes[0])
R
root 已提交
271

C
channings 已提交
272 273
        shape = node.out_shapes[0]
        for i, dim_shape in enumerate(shape):
R
root 已提交
274 275 276
            if dim_shape == 0 and i == 0:
                shape[i] = 1
            if dim_shape == 0 and i != 0:
C
channings 已提交
277
                assert 'shape of input is not assigned'
C
update  
channingss 已提交
278 279
        attr = {
            "dtype": string(node.dtype),
C
channings 已提交
280
            "shape": shape,
C
update  
channingss 已提交
281 282 283 284
            "name": string(node.layer_name),
            "append_batch_size": 'False'
        }

285 286
        node.fluid_code.add_layer(
            "data", inputs=None, output=node, param_attr=attr)
C
update  
channingss 已提交
287

288
    @print_mapping_info
C
update  
channingss 已提交
289 290 291 292
    def create_parameter(self, node, parameter=None):
        if parameter is not None:
            node = parameter
        dtype = node.dtype
C
channingss 已提交
293
        shape = node.out_shapes[0]
C
channingss 已提交
294 295
        if len(node.weight.shape) == 0:
            shape = [1]
C
update  
channingss 已提交
296 297 298 299 300 301 302
        self.weights[node.layer_name] = node.weight
        attr = {
            'dtype': string(dtype),
            'shape': shape,
            'name': string(node.layer_name),
            'default_initializer': 'Constant(0.0)'
        }
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
        if dtype == 'bool':
            attr['dtype'] = string('int64')
            node.fluid_code.add_layer(
                "create_parameter", inputs=None, output=node, param_attr=attr)
            node.fluid_code.add_layer(
                "cast",
                inputs=node,
                output=node,
                param_attr={'dtype': string('bool')})
        elif dtype == 'uint8':
            attr['dtype'] = string('float32')
            node.fluid_code.add_layer(
                "create_parameter", inputs=None, output=node, param_attr=attr)
        else:
            node.fluid_code.add_layer(
                "create_parameter", inputs=None, output=node, param_attr=attr)
C
update  
channingss 已提交
319 320 321 322 323 324 325 326 327 328 329 330 331 332

    def _pad_if_asymmetric(self, node, pads, val_name):  # pads: SSEE
        assert len(pads) & 1 == 0
        symmetric = True
        ndims = len(pads) // 2
        for idx_dim in range(ndims):
            if pads[idx_dim] != pads[ndims + idx_dim]:
                symmetric = False
                break
        if symmetric:
            return pads[:ndims], val_name
        val_padded = self.Pad(node, op_independent=False)
        return [0] * ndims, val_padded

C
channingss 已提交
333
    def _interpolate(self, node):
C
channingss 已提交
334
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
Channingss 已提交
335
        inputs = {'input': val_x}
336
        if node.layer_type == 'Resize':
C
Channingss 已提交
337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362
            if len(node.layer.input) == 2:
                # opset 10
                val_scales = self.graph.get_input_node(node, idx=1, copy=True)
                inputs['scale'] = val_scales
            elif len(node.layer.input) == 3:
                # opset 11
                val_scales = self.graph.get_input_node(node, idx=2, copy=True)
                inputs['scale'] = val_scales
            elif len(node.layer.input) == 4:
                # opset 11
                val_sizes = self.graph.get_input_node(node, idx=3, copy=True)
                var_nc, var_hw = val_sizes.layer_name + '_nc', val_sizes.layer_name + '_hw'
                node.fluid_code.add_layer(
                    'split',
                    inputs=val_sizes,
                    output=var_nc + ',' + var_hw,
                    param_attr={
                        'dim': 0,
                        'num_or_sections': [2, 2],
                    })
                node.fluid_code.add_layer(
                    "cast",
                    inputs=var_hw,
                    output=var_hw,
                    param_attr={'dtype': string('int32')})
                inputs['out_shape'] = var_hw
363 364
        elif node.layer_type == 'Upsample':
            val_scales = self.graph.get_input_node(node, idx=1, copy=True)
C
Channingss 已提交
365
            inputs['scale'] = val_scales
R
root 已提交
366 367

        attr = {'name': string(node.layer_name)}
C
channingss 已提交
368 369
        mode = node.get_attr('mode', 'nearest')
        fluid_op = 'resize_{}'.format(mode)
370
        if 'linear' in mode:
R
root 已提交
371 372 373
            print(
                'Warnning: paddle not support op:resize wiht mode: linear, we use bilinear replace linear'
            )
374
            fluid_op = 'resize_bilinear'
375
        node.fluid_code.add_layer(
C
Channingss 已提交
376
            fluid_op, inputs=inputs, output=node, param_attr=attr)
R
root 已提交
377

378
    @print_mapping_info
C
channings 已提交
379 380 381
    def RoiAlign(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_rois = self.graph.get_input_node(node, idx=1, copy=True)
R
root 已提交
382 383 384

        pooled_height = node.get_attr('output_height')
        pooled_width = node.get_attr('output_width')
C
channings 已提交
385 386 387
        spatial_scale = node.get_attr('spatial_scale')
        sampling_ratio = node.get_attr('sampling_ratio')
        attr = {
R
root 已提交
388 389 390 391 392
            'pooled_height': pooled_height,
            'pooled_width': pooled_width,
            'spatial_scale': spatial_scale,
            'sampling_ratio': sampling_ratio,
        }
393 394 395 396 397 398 399 400
        node.fluid_code.add_layer(
            'roi_align',
            inputs={'input': val_x,
                    'rois': val_rois},
            output=node,
            param_attr=attr)

    @print_mapping_info
C
channings 已提交
401 402 403
    def MaxRoiPool(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_rois = self.graph.get_input_node(node, idx=1, copy=True)
R
root 已提交
404

C
channings 已提交
405 406 407
        spatial_scale = node.get_attr('spatial_scale')
        pooled_height, pooled_width = node.get_attr('pooled_shape')
        attr = {
R
root 已提交
408 409 410 411
            'pooled_height': pooled_height,
            'pooled_width': pooled_width,
            'spatial_scale': spatial_scale,
        }
412 413 414 415 416 417 418 419
        node.fluid_code.add_layer(
            'roi_pool',
            inputs={'input': val_x,
                    'rois': val_rois},
            output=node,
            param_attr=attr)

    @print_mapping_info
C
update  
channingss 已提交
420
    def Pad(self, node, op_independent=True):
C
channingss 已提交
421
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
422 423 424
        pads = node.get_attr('pads')
        mode = node.get_attr('mode', 'constant')
        value = node.get_attr('value', 0.)
C
channingss 已提交
425 426
        data_shape = val_x.out_shapes[0]
        output_shape = node.out_shapes[0]
C
update  
channingss 已提交
427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447
        assume_pad2d = False
        attr = {}
        if len(pads) == 4:
            assume_pad2d |= mode != 'constant'
            if data_shape:
                assume_pad2d |= data_shape and len(data_shape) == 4  # NCHW
            if output_shape:
                assume_pad2d |= output_shape and len(output_shape) == 4  # NCHW
        if assume_pad2d:
            fluid_op = 'pad2d'
            attr['data_format'] = string('NCHW')
            attr['mode'] = string(mode)
        else:
            attr = {'pad_value': value}
            fluid_op = 'pad'
        if len(pads) == 4:
            paddings = np.array(pads).reshape(
                (-1, 2)).transpose().flatten().tolist()  # SSEE -> SESE
        elif len(pads) == 8:
            paddings = np.array(pads).reshape(
                (-1, 4)).transpose().flatten().tolist()  # SSEE -> SESE
C
channingss 已提交
448 449 450 451
            if sum(paddings[:4]) == 0:
                fluid_op = 'pad2d'
                paddings = paddings[4:]
                attr['mode'] = string(mode)
C
update  
channingss 已提交
452 453 454
        attr['paddings'] = paddings
        if op_independent:
            attr['name'] = string(node.layer_name)
455 456
            node.fluid_code.add_layer(
                fluid_op, inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
457 458
        else:
            attr['name'] = string(node.layer_name + '_paded')
459 460 461 462 463
            node.fluid_code.add_layer(
                fluid_op,
                inputs=val_x,
                output=node.layer_name + '_paded',
                param_attr=attr)
C
update  
channingss 已提交
464 465
            return node.layer_name + '_paded'

466
    @print_mapping_info
C
update  
channingss 已提交
467
    def Unsqueeze(self, node):
C
channingss 已提交
468
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
469
        axes = node.get_attr('axes')
470
        attr = {'axes': axes, 'name': string(node.layer_name)}
R
root 已提交
471
        if len(val_x.out_shapes[0]) == 0:
472 473 474 475 476 477
            if node.layer_name:
                node.fluid_code.add_layer(
                    'reshape',
                    inputs=val_x,
                    output=node,
                    param_attr={'shape': [1]})
478
        else:
479 480
            node.fluid_code.add_layer(
                'unsqueeze', inputs=val_x, output=node, param_attr=attr)
481

482
    @print_mapping_info
C
channingss 已提交
483
    def Shrink(self, node):
C
channingss 已提交
484
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
channingss 已提交
485 486 487 488
        bias = node.get_attr('bias')
        lambd = node.get_attr('lambd')
        assert bias == 0.0, 'not support bias!=0'
        attr = {'threshold': lambd, 'name': node.layer_name}
489 490
        node.fluid_code.add_layer(
            'hard_shrink', inputs=val_x, output=node, param_attr=attr)
C
channingss 已提交
491

492 493 494 495 496 497 498 499 500 501 502
    def Greater(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_y = self.graph.get_input_node(node, idx=1, copy=True)
        node.fluid_code.add_layer(
            'greater_than',
            inputs={'x': val_x,
                    'y': val_y},
            output=node,
            param_attr=None)

    @print_mapping_info
C
update  
channingss 已提交
503 504 505 506 507 508 509 510
    def Constant(self, node):
        val_output = self.graph.get_node(node.layer.output[0], copy=True)

        value = node.get_attr('value')
        dtype = np.dtype(value.dtype)
        output_dtype = val_output.dtype
        if output_dtype:
            assert dtype == output_dtype, 'tensor dtype unmatches storage dtype'
R
root 已提交
511

C
update  
channingss 已提交
512
        shape = node.get_attr('shape', None)
R
root 已提交
513

C
update  
channingss 已提交
514
        if shape is None:
C
channingss 已提交
515
            shape = val_output.out_shapes[0]
C
update  
channingss 已提交
516 517
        if shape is None:
            shape = list(value.shape)
518 519 520 521
            _logger.warning('in (Constant -> %s): '
                            'attribute "shape" of %s not inferred, '
                            'using value as 1-D tensor may lead to fails',
                            val_output.layer_name, val_output.layer_name)
522
        if len(value) == 1:
C
channingss 已提交
523
            value = value.tolist()
C
update  
channingss 已提交
524 525 526 527 528
            shape = [1]
            value = value[0]
            if dtype.name == 'int64':
                dtype = 'int32'
            attr = {'shape': shape, 'dtype': string(dtype), 'value': value}
529 530
            node.fluid_code.add_layer(
                'fill_constant', inputs=None, output=node, param_attr=attr)
C
channingss 已提交
531
        else:
532 533
            if dtype.name == 'uint8':
                dtype = 'int64'
C
channingss 已提交
534 535 536 537 538 539 540 541
            value = np.reshape(value, shape)
            self.weights[node.layer_name] = value
            attr = {
                'dtype': string(dtype),
                'shape': shape,
                'name': string(node.layer_name),
                'default_initializer': 'Constant(0.0)'
            }
542 543
            node.fluid_code.add_layer(
                "create_parameter", inputs=None, output=node, param_attr=attr)
C
update  
channingss 已提交
544

545
    @print_mapping_info
C
update  
channingss 已提交
546
    def Resize(self, node):
547 548
        self._interpolate(node)

549
    @print_mapping_info
550 551 552
    def Upsample(self, node):
        self._interpolate(node)

553 554 555 556 557 558 559 560 561 562 563 564 565 566 567
    @print_mapping_info
    def InstanceNormalization(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_scale = self.graph.get_input_node(node, idx=1, copy=True)
        val_b = self.graph.get_input_node(node, idx=2, copy=True)
        epsilon = node.get_attr('epsilon', 1e-5)
        attr = {
            'epsilon': epsilon,
            'param_attr': string(val_scale.layer_name),
            'bias_attr': string(val_b.layer_name)
        }
        node.fluid_code.add_layer(
            "instance_norm", inputs=val_x, output=node, param_attr=attr)

    @print_mapping_info
568
    def Expand(self, node):
C
channingss 已提交
569
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
570
        val_shape = self.graph.get_input_node(node, idx=1, copy=True)
R
root 已提交
571 572

        if len(val_shape.outputs) == 1:
573 574
            self.omit_nodes.append(val_shape.layer_name)

C
channingss 已提交
575
        val_y = self.graph.get_node(node.layer.output[0], copy=True)
576
        out_shape = node.out_shapes[0]
577
        val_x_dtype = val_x.dtype
R
root 已提交
578 579 580

        name_ones = node.layer_name + '_ones'
        attr_ones = {'shape': out_shape, 'dtype': string(val_x_dtype)}
581 582
        node.fluid_code.add_layer(
            'ones', inputs=None, output=name_ones, param_attr=attr_ones)
R
root 已提交
583 584
        inputs = {'x': name_ones, 'y': val_x}
        attr = {'name': string(node.layer_name)}
585 586 587 588 589
        node.fluid_code.add_layer(
            'elementwise_mul',
            inputs=inputs,
            output=node.layer_name,
            param_attr=attr)
C
update  
channingss 已提交
590

591
    @print_mapping_info
C
channingss 已提交
592 593 594 595
    def Gather(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        indices = self.graph.get_input_node(node, idx=1, copy=True)
        indices_shape = indices.out_shapes[0]
C
Channingss 已提交
596
        axis = node.get_attr('axis', 0)
597 598
        #assert len(
        #    indices_shape) <= 2, "Gather op don't support dim of indice >2 "
R
root 已提交
599
        if axis == 0 and len(indices_shape) <= 1:
600 601 602 603 604 605
            node.fluid_code.add_layer(
                'gather',
                inputs={'input': val_x,
                        'index': indices},
                output=node,
                param_attr=None)
C
channingss 已提交
606 607
        elif axis > 0 and len(indices_shape) <= 1:
            perm = list(range(len(val_x.out_shapes[0])))
C
channingss 已提交
608 609 610
            perm = [axis] + perm[:axis] + perm[axis + 1:]
            attr_trans = {'perm': perm}
            name_trans = val_x.layer_name + '_trans'
611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665
            node.fluid_code.add_layer(
                'transpose',
                inputs=val_x,
                output=name_trans,
                param_attr=attr_trans)
            node.fluid_code.add_layer(
                'gather',
                inputs={'input': name_trans,
                        'index': indices},
                output=node,
                param_attr=None)
            node.fluid_code.add_layer(
                'transpose', inputs=node, output=node, param_attr=attr_trans)
        elif axis == 0 and len(indices_shape) > 1:
            if val_x.out_shapes[0] is not None and isinstance(
                    val_x, ONNXGraphDataNode):
                node.fluid_code.add_layer(
                    'embedding',
                    inputs=indices,
                    output=node,
                    use_fluid=True,
                    param_attr={
                        'param_attr': string(val_x.layer_name),
                        'size': val_x.out_shapes[0]
                    })
            else:
                from functools import reduce
                #indices_shape = [1,7]
                reshape_shape = reduce(lambda x, y: x * y, indices_shape)
                indices_reshape = indices.layer_name + '_shape'
                node.fluid_code.add_layer(
                    'reshape',
                    inputs=indices,
                    output=indices_reshape,
                    param_attr={'shape': [reshape_shape, ]})

                perm = list(range(len(val_x.out_shapes[0])))
                node.fluid_code.add_layer(
                    'gather',
                    inputs={'input': val_x,
                            'index': indices_reshape},
                    output=node,
                    param_attr=None)
                val_x_shape = val_x.out_shapes[0]
                reshaped_shape = []
                for i in perm:
                    reshaped_shape.append(indices_shape[i])
                for i in val_x_shape[:axis] + val_x_shape[axis + 1:]:
                    reshaped_shape.append(i)
                node.fluid_code.add_layer(
                    'reshape',
                    inputs=node,
                    output=node,
                    param_attr={'shape': reshaped_shape})
        elif axis > 0 and len(indices_shape) > 1:
C
Channingss 已提交
666
            from functools import reduce
R
root 已提交
667
            reshape_shape = reduce(lambda x, y: x * y, indices_shape)
668 669 670 671 672 673
            indices_reshape = indices.layer_name + '_shape'
            node.fluid_code.add_layer(
                'reshape',
                inputs=indices,
                output=indices_reshape,
                param_attr={'shape': [reshape_shape, ]})
R
root 已提交
674

C
Channingss 已提交
675 676 677 678
            perm = list(range(len(val_x.out_shapes[0])))
            perm = [axis] + perm[:axis] + perm[axis + 1:]
            attr_trans = {'perm': perm}
            name_trans = val_x.layer_name + '_trans'
679 680 681 682 683 684 685 686 687 688 689 690 691
            node.fluid_code.add_layer(
                'transpose',
                inputs=val_x,
                output=name_trans,
                param_attr=attr_trans)
            node.fluid_code.add_layer(
                'gather',
                inputs={'input': name_trans,
                        'index': indices_reshape},
                output=node,
                param_attr=None)
            node.fluid_code.add_layer(
                'transpose', inputs=node, output=node, param_attr=attr_trans)
C
Channingss 已提交
692 693 694 695 696 697
            val_x_shape = val_x.out_shapes[0]
            reshaped_shape = []
            for i in perm:
                reshaped_shape.append(indices_shape[i])
            for i in val_x_shape[:axis] + val_x_shape[axis + 1:]:
                reshaped_shape.append(i)
698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
            node.fluid_code.add_layer(
                'reshape',
                inputs=node,
                output=node,
                param_attr={'shape': reshaped_shape})

    @print_mapping_info
    def Range(self, node):
        val_start = self.graph.get_input_node(node, idx=0, copy=True)
        val_limit = self.graph.get_input_node(node, idx=1, copy=True)
        val_delta = self.graph.get_input_node(node, idx=2, copy=True)
        dtype = val_start.dtype
        inputs = {'start': val_start, 'end': val_limit, 'step': val_delta}
        node.fluid_code.add_layer(
            'range',
            inputs=inputs,
            output=node,
            param_attr={'dtype': string(dtype)})

    @print_mapping_info
C
channingss 已提交
718
    def Slice(self, node):
C
channingss 已提交
719
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
channings 已提交
720
        starts, ends, axes, steps = None, None, None, None
721
        attr = {}
C
channingss 已提交
722 723 724
        if len(node.inputs) > 1:
            starts = self.graph.get_input_node(node, idx=1, copy=True)
            ends = self.graph.get_input_node(node, idx=2, copy=True)
R
root 已提交
725
            if len(node.inputs) > 3:
C
channings 已提交
726 727
                axes = self.graph.get_input_node(node, idx=3, copy=True)
                axes = _const_weight_or_none(axes)
R
root 已提交
728
            if len(node.inputs) > 4:
C
channings 已提交
729 730
                steps = self.graph.get_input_node(node, idx=4, copy=True)
                steps = _const_weight_or_none(steps)
731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
                if steps is not None:
                    assert steps == 1, "Only support convert op:Slice, which attribute:steps == 1"
            attr = {
                "axes": axes,
                "starts": starts.layer_name,
                "ends": ends.layer_name
            }
            starts_value = _const_weight_or_none(starts)
            ends_value = _const_weight_or_none(ends)
            if starts_value is not None and ends_value is not None:
                self.omit_nodes.append(starts.layer_name)
                self.omit_nodes.append(ends.layer_name)
                ends_value = ends_value.copy()
                for idx in range(len(ends_value)):
                    if ends_value[idx] > 2**31 - 1:
                        ends_value[idx] = 2**31 - 1
                attr = {
                    "axes": axes,
                    "starts": starts_value,
                    "ends": ends_value
                }
            else:
                if starts.dtype != 'int32':
                    node.fluid_code.add_layer(
                        'cast',
                        inputs=starts,
                        output=starts,
                        param_attr={'dtype': string('int32')})
                if ends.dtype != 'int32':
                    node.fluid_code.add_layer(
                        'cast',
                        inputs=ends,
                        output=ends,
                        param_attr={'dtype': string('int32')})
C
channingss 已提交
765 766 767 768
        else:
            starts = node.get_attr('starts')
            ends = node.get_attr('ends')
            axes = node.get_attr('axes')
769 770 771 772
            for idx in range(len(ends)):
                if ends[idx] > 2**31 - 1:
                    ends[idx] = 2**31 - 1
            attr = {"axes": axes, "starts": starts, "ends": ends}
C
channingss 已提交
773

774 775
        node.fluid_code.add_layer(
            'slice', inputs=val_x, output=node, param_attr=attr)
C
channingss 已提交
776

777
    @print_mapping_info
C
update  
channingss 已提交
778
    def ConstantOfShape(self, node):
C
channingss 已提交
779
        val_shape = self.graph.get_input_node(node, idx=0, copy=True)
C
channingss 已提交
780
        val_y = self.graph.get_node(node.layer.output[0], copy=True)
C
update  
channingss 已提交
781 782 783 784

        value = node.get_attr('value')
        dtype = value.dtype
        value = value.tolist()
785 786
        assert len(value) == 1, ('given value not Scalar, shape of value > 1, '
                                 'this is not supported')
C
update  
channingss 已提交
787 788 789 790
        if len(value) == 1:
            value = value[0]
            if dtype.name == 'int64':
                dtype = 'int32'
791 792 793 794 795 796 797
            attr = {
                'shape': val_shape.layer_name,
                'dtype': string(dtype),
                'value': value
            }
            node.fluid_code.add_layer(
                'fill_constant', inputs=None, output=node, param_attr=attr)
C
update  
channingss 已提交
798

799
    @print_mapping_info
C
update  
channingss 已提交
800
    def Split(self, node):
C
channingss 已提交
801 802
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_y = self.graph.get_node(node.layer.output[0], copy=True)
C
update  
channingss 已提交
803 804

        fluid_op = 'split'
C
channingss 已提交
805
        split = node.get_attr('split')
C
update  
channingss 已提交
806
        axis = node.get_attr('axis', 0)
C
channingss 已提交
807 808 809 810 811
        attr = {
            'num_or_sections': split,
            'dim': axis,
            'name': string(node.layer_name)
        }
R
root 已提交
812

813 814
        node.fluid_code.add_layer(
            'split', inputs=val_x, output=val_y, param_attr=attr)
C
update  
channingss 已提交
815

816
    @print_mapping_info
C
update  
channingss 已提交
817
    def Reshape(self, node):
C
channingss 已提交
818 819
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_shape = self.graph.get_input_node(node, idx=1, copy=True)
C
update  
channingss 已提交
820
        val_reshaped = self.graph.get_node(node.layer.output[0], copy=True)
821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837
        attr = {}
        shape_value = _const_weight_or_none(val_shape)
        shape_dims = len(val_shape.out_shapes[0])

        if shape_value is not None:
            node.fluid_code.add_layer(
                'reshape',
                inputs={'x': val_x},
                output=node,
                param_attr={'shape': shape_value.tolist()})
        elif val_shape.dtype == 'int64':
            val_shape_cast = val_shape.layer_name + '_cast'
            node.fluid_code.add_layer(
                'cast',
                inputs=val_shape,
                output=val_shape_cast,
                param_attr={'dtype': string('int32')})
838 839 840 841 842 843 844
            # shape may be [], come form Gather by scalar indices
            if len(val_shape.out_shapes[0]) > 0:
                node.fluid_code.add_layer(
                    'reshape',
                    inputs=val_shape_cast,
                    output=val_shape_cast,
                    param_attr={'shape': val_shape.out_shapes[0]})
845 846 847 848 849 850 851
            node.fluid_code.add_layer(
                'reshape',
                inputs={'x': val_x,
                        'shape': val_shape_cast},
                output=node,
                param_attr=attr)
        else:
852 853 854 855 856 857 858
            # shape may be [], come form Gather by scalar indices
            if len(val_shape.out_shapes[0]) > 0:
                node.fluid_code.add_layer(
                    'reshape',
                    inputs=val_shape,
                    output=val_shape,
                    param_attr={'shape': val_shape.out_shapes[0]})
859 860 861 862 863 864 865 866
            node.fluid_code.add_layer(
                'reshape',
                inputs={'x': val_x,
                        'shape': val_shape},
                output=node,
                param_attr=attr)

    @print_mapping_info
C
update  
channingss 已提交
867
    def Cast(self, node):
C
channingss 已提交
868
        val_input = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
869 870 871 872 873 874 875 876 877 878
        val_output = self.graph.get_node(node.layer.output[0], copy=True)

        dtype = node.get_attr('to')
        if not isinstance(dtype, np.dtype):
            dtype = TENSOR_TYPE_TO_NP_TYPE[dtype]

        output_dtype = val_output.dtype
        if output_dtype:
            assert dtype == output_dtype, 'dtype of to unmatches output'
        attr = {'dtype': string(dtype)}
879 880
        node.fluid_code.add_layer(
            'cast', inputs=val_input, output=node, param_attr=attr)
C
update  
channingss 已提交
881

882
    @print_mapping_info
C
update  
channingss 已提交
883
    def AveragePool(self, node):
C
channingss 已提交
884
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
channingss 已提交
885 886

        auto_pad = node.get_attr('auto_pad', 'NOTSET')
C
update  
channingss 已提交
887 888 889 890 891 892 893 894
        kernel_shape = node.get_attr("kernel_shape")
        poolnd = len(kernel_shape)
        strides = node.get_attr("strides")
        pad_mode = node.get_attr("pads")
        ceil_mode = bool(node.get_attr('ceil_mode', 0))
        pads = node.get_attr('pads', [0] * (poolnd * 2))
        fluid_op = 'pool{}d'.format(poolnd)
        assert 2 <= poolnd <= 3, 'only pool2d and pool3d is supported'
C
channingss 已提交
895

C
channingss 已提交
896 897
        paddings, val_x = self._pad_if_asymmetric(node, pads, val_x)

C
channingss 已提交
898
        if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER":
C
channingss 已提交
899
            input_shape = val_x.out_shapes[0]
C
channingss 已提交
900 901 902 903 904 905
            pad_h = get_same_padding(input_shape[2], kernel_shape[0],
                                     strides[0])
            pad_w = get_same_padding(input_shape[3], kernel_shape[1],
                                     strides[1])
            attr = {"paddings": pad_h + pad_w, "pad_value": 0.0}

C
update  
channingss 已提交
906 907 908 909 910 911 912 913 914 915
        attr = {
            "pool_size": kernel_shape,
            "pool_type": string('avg'),
            "pool_stride": strides,
            "pool_padding": paddings,
            "ceil_mode": ceil_mode,
            "exclusive": 'True',
            "name": string(node.layer_name)
        }

916 917
        node.fluid_code.add_layer(
            fluid_op, inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
918

919
    @print_mapping_info
C
update  
channingss 已提交
920 921 922
    def Concat(self, node):
        inputs = []
        for i in range(len(node.layer.input)):
C
channingss 已提交
923
            ipt = self.graph.get_input_node(node, idx=i, copy=True)
C
update  
channingss 已提交
924 925 926 927 928 929
            if isinstance(ipt, str):
                inputs.append(ipt)
            else:
                inputs.append(ipt.layer_name)
        axis = node.get_attr('axis')
        attr = {'axis': axis}
930 931
        node.fluid_code.add_layer(
            'concat', inputs=inputs, output=node, param_attr=attr)
C
update  
channingss 已提交
932

933
    @print_mapping_info
C
update  
channingss 已提交
934
    def Flatten(self, node):
C
channingss 已提交
935
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
936 937
        axis = node.get_attr('axis', 1)
        attr = {"axis": str(axis), "name": string(node.layer_name)}
938 939
        node.fluid_code.add_layer(
            'flatten', inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
940

941
    @print_mapping_info
C
update  
channingss 已提交
942
    def Gemm(self, node):
C
channingss 已提交
943 944 945
        val_a = self.graph.get_input_node(node, idx=0, copy=True)
        val_b = self.graph.get_input_node(node, idx=1, copy=True)
        val_c = self.graph.get_input_node(node, idx=2, copy=True)
C
update  
channingss 已提交
946 947 948 949 950 951 952 953 954 955 956 957 958

        alpha = node.get_attr('alpha', 1.)  # optional
        beta = node.get_attr('beta', 1.)  # optional
        trans_a = bool(node.get_attr('transA', 0))  # optional
        trans_b = bool(node.get_attr('transB', 0))  # optional
        val_mm = node.layer_name + '_mm'
        matmul_inputs = {"x": val_a, "y": val_b}
        attr_matmul = {
            "transpose_x": trans_a,
            "transpose_y": trans_b,
            "alpha": alpha,
            "name": string(val_mm)
        }
959 960 961 962 963
        node.fluid_code.add_layer(
            'matmul',
            inputs=matmul_inputs,
            output=val_mm,
            param_attr=attr_matmul)
C
channingss 已提交
964

C
update  
channingss 已提交
965 966 967 968
        if beta != 0:
            if beta == 1.:
                add_inputs = {"x": val_mm, "y": val_c}
                attr = {"name": string(node.layer_name)}
969 970 971 972 973
                node.fluid_code.add_layer(
                    "elementwise_add",
                    inputs=add_inputs,
                    output=node,
                    param_attr=attr)
C
update  
channingss 已提交
974
            else:
C
channingss 已提交
975 976
                var_beta = node.layer_name + '_beta'
                matmul_beta_inputs = {"x": val_c, "y": var_beta}
977 978 979 980 981
                node.fluid_code.add_layer(
                    "Constant",
                    inputs=matmul_beta_inputs,
                    output=var_beta,
                    param_attr={'value': beta})
C
channingss 已提交
982 983 984

                add_inputs = {"x": val_mm, "y": var_beta}
                attr = {"name": string(node.layer_name)}
985 986 987 988 989
                node.fluid_code.add_layer(
                    "elementwise_add",
                    inputs=add_inputs,
                    output=node,
                    param_attr=attr)
C
update  
channingss 已提交
990

991
    @print_mapping_info
C
update  
channingss 已提交
992
    def Sum(self, node):
993
        val_inps = node.layer.input
994
        inputs = {
995 996 997 998
            "x": self.graph.get_input_node(
                node, idx=0, copy=True),
            "y": self.graph.get_input_node(
                node, idx=1, copy=True),
999 1000
        }
        node.fluid_code.add_layer("elementwise_add", inputs=inputs, output=node)
1001

C
channingss 已提交
1002 1003
        for idx, ipt in enumerate(val_inps[2:]):
            y = self.graph.get_input_node(node, idx=idx, copy=True)
1004 1005
            inputs = {
                "x": node.layer_name,
C
channingss 已提交
1006
                "y": y,
1007
            }
1008 1009
            node.fluid_code.add_layer(
                "elementwise_add", inputs=inputs, output=node)
C
update  
channingss 已提交
1010

1011
    @print_mapping_info
C
update  
channingss 已提交
1012
    def MatMul(self, node):
C
channingss 已提交
1013 1014
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_y = self.graph.get_input_node(node, idx=1, copy=True)
C
update  
channingss 已提交
1015 1016
        inputs = {"x": val_x, "y": val_y}
        attr = {"name": string(node.layer_name)}
1017 1018
        node.fluid_code.add_layer(
            "matmul", inputs=inputs, output=node, param_attr=attr)
C
update  
channingss 已提交
1019

1020
    @print_mapping_info
C
update  
channingss 已提交
1021
    def BatchNormalization(self, node):
C
channingss 已提交
1022 1023 1024 1025 1026
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_scale = self.graph.get_input_node(node, idx=1, copy=True)
        val_b = self.graph.get_input_node(node, idx=2, copy=True)
        val_mean = self.graph.get_input_node(node, idx=3, copy=True)
        val_var = self.graph.get_input_node(node, idx=4, copy=True)
C
update  
channingss 已提交
1027 1028 1029 1030 1031 1032 1033 1034 1035

        self.omit_nodes.append(val_scale.layer_name)
        self.omit_nodes.append(val_b.layer_name)
        self.omit_nodes.append(val_mean.layer_name)
        self.omit_nodes.append(val_var.layer_name)

        momentum = node.get_attr('momentum', .9)
        epsilon = node.get_attr('epsilon', 1e-5)

C
channingss 已提交
1036 1037
        # Attribute: spatial is used in BatchNormalization-1,6,7
        spatial = bool(node.get_attr('spatial'))
C
update  
channingss 已提交
1038 1039 1040 1041
        attr = {
            "momentum": momentum,
            "epsilon": epsilon,
            "data_layout": string('NCHW'),
C
channingss 已提交
1042
            "is_test": True,
C
update  
channingss 已提交
1043 1044 1045 1046
            "param_attr": string(val_scale.layer_name),
            "bias_attr": string(val_b.layer_name),
            "moving_mean_name": string(val_mean.layer_name),
            "moving_variance_name": string(val_var.layer_name),
C
channingss 已提交
1047
            "use_global_stats": spatial,
C
update  
channingss 已提交
1048 1049
            "name": string(node.layer_name)
        }
1050 1051
        node.fluid_code.add_layer(
            "batch_norm", inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
1052

1053
    @print_mapping_info
C
update  
channingss 已提交
1054
    def Transpose(self, node):
C
channingss 已提交
1055
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
1056 1057
        perm = node.get_attr('perm')
        attr = {'perm': perm, "name": string(node.layer_name)}
1058 1059
        node.fluid_code.add_layer(
            "transpose", inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
1060

1061
    @print_mapping_info
C
update  
channingss 已提交
1062
    def Relu(self, node):
C
channingss 已提交
1063
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
1064
        attr = {"name": string(node.layer_name)}
1065 1066
        node.fluid_code.add_layer(
            "relu", inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
1067

1068
    @print_mapping_info
C
update  
channingss 已提交
1069
    def PRelu(self, node):
C
channingss 已提交
1070 1071
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_slope = self.graph.get_input_node(node, idx=1, copy=True)
C
update  
channingss 已提交
1072

C
channingss 已提交
1073 1074 1075 1076 1077 1078 1079 1080 1081 1082
        mode = 'channel'
        shape_slope = val_slope.out_shapes[0]
        if len(shape_slope) == 1:
            mode = 'all'
        elif len(shape_slope) > 2:
            mode = 'element'
        attr = {
            "param_attr": string(val_slope.layer_name),
            'mode': string(mode)
        }
1083 1084
        node.fluid_code.add_layer(
            "prelu", inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
1085

1086
    @print_mapping_info
C
update  
channingss 已提交
1087
    def Squeeze(self, node):
C
channingss 已提交
1088 1089 1090
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        axes = node.get_attr('axes')
        attr = {'axes': axes, "name": string(node.layer_name)}
1091 1092 1093 1094 1095 1096 1097 1098 1099
        if len(val_x.out_shapes[0]) == 1:
            node.fluid_code.add_layer(
                "cast",
                inputs=val_x,
                output=node,
                param_attr={'dtype': string(val_x.dtype)})
        else:
            node.fluid_code.add_layer(
                "squeeze", inputs=val_x, output=node, param_attr=attr)
R
root 已提交
1100

1101
    @print_mapping_info
C
channings 已提交
1102 1103 1104
    def Equal(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_y = self.graph.get_input_node(node, idx=1, copy=True)
1105 1106 1107 1108 1109 1110 1111
        node.fluid_code.add_layer(
            "equal",
            inputs={'x': val_x,
                    'y': val_y},
            output=node,
            param_attr=None)

C
Channingss 已提交
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122
    @print_mapping_info
    def Greater(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_y = self.graph.get_input_node(node, idx=1, copy=True)
        node.fluid_code.add_layer(
            "greater_than",
            inputs={'x': val_x,
                    'y': val_y},
            output=node,
            param_attr=None)

1123
    @print_mapping_info
C
channings 已提交
1124 1125 1126 1127
    def Where(self, node):
        condition = self.graph.get_input_node(node, idx=0, copy=True)
        val_x = self.graph.get_input_node(node, idx=1, copy=True)
        val_y = self.graph.get_input_node(node, idx=2, copy=True)
R
root 已提交
1128

C
channings 已提交
1129
        not_condition = condition.layer_name + '_not'
1130 1131 1132 1133 1134
        node.fluid_code.add_layer(
            "logical_not",
            inputs=condition,
            output=not_condition,
            param_attr=None)
R
root 已提交
1135
        cast_not_condition = not_condition + '_cast'
1136 1137 1138 1139 1140
        node.fluid_code.add_layer(
            "cast",
            inputs=not_condition,
            output=cast_not_condition,
            param_attr={'dtype': string(val_x.dtype)})
C
channings 已提交
1141
        cast_condition = condition.layer_name + '_cast'
1142 1143 1144 1145 1146
        node.fluid_code.add_layer(
            "cast",
            inputs=condition,
            output=cast_condition,
            param_attr={'dtype': string(val_x.dtype)})
R
root 已提交
1147
        mul_val_x = val_x.layer_name + '_mul'
1148 1149 1150 1151 1152 1153
        node.fluid_code.add_layer(
            "elementwise_mul",
            inputs={'x': val_x,
                    'y': cast_condition},
            output=mul_val_x,
            param_attr=None)
R
root 已提交
1154

C
channings 已提交
1155
        mul_val_y = val_y.layer_name + '_mul'
1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170
        node.fluid_code.add_layer(
            "elementwise_mul",
            inputs={'x': val_y,
                    'y': cast_not_condition},
            output=mul_val_y,
            param_attr=None)

        node.fluid_code.add_layer(
            "elementwise_add",
            inputs={'x': mul_val_x,
                    'y': mul_val_y},
            output=node,
            param_attr=None)

    @print_mapping_info
R
root 已提交
1171 1172
    def NonZero(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192
        val_x_dim = len(val_x.out_shapes[0])
        print(val_x.layer_name, val_x.out_shapes[0])
        if val_x_dim == 1:
            node.fluid_code.add_layer("nonzero", inputs=val_x, output=val_x)
            node.fluid_code.add_layer(
                "transpose",
                inputs=val_x,
                output=node,
                param_attr={'perm': [1, 0]})
        if val_x_dim > 1:
            node.fluid_code.add_layer("nonzero", inputs=val_x, output=val_x)
            node.fluid_code.add_layer(
                "split",
                inputs=val_x,
                output=val_x,
                param_attr={'num_or_sections': 1,
                            'dim': val_x_dim})
            node.fluid_code.add_layer("concat", inputs=val_x, output=node)

    @print_mapping_info
C
update  
channingss 已提交
1193
    def Identity(self, node):
C
channingss 已提交
1194
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
1195
        node.fluid_code.add_layer("assign", inputs=val_x, output=node)
R
root 已提交
1196

1197
    @print_mapping_info
C
channings 已提交
1198 1199 1200 1201
    def Tile(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_repeats = self.graph.get_input_node(node, idx=1, copy=True)
        repeats = _const_weight_or_none(val_repeats)
R
root 已提交
1202

1203 1204 1205
        if repeats is None:
            repeats = val_repeats.layer_name
        elif isinstance(repeats, int):
C
channings 已提交
1206
            repeats = [repeats]
R
root 已提交
1207

C
channings 已提交
1208
        attr = {
R
root 已提交
1209
            'expand_times': repeats,
C
channings 已提交
1210 1211
            "name": string(node.layer_name),
        }
1212 1213
        node.fluid_code.add_layer(
            "expand", inputs=val_x, output=node, param_attr=attr)
R
root 已提交
1214

1215
    @print_mapping_info
C
update  
channingss 已提交
1216
    def MaxPool(self, node):
C
channingss 已提交
1217
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
channingss 已提交
1218
        auto_pad = node.get_attr('auto_pad', 'NOTSET')
C
update  
channingss 已提交
1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229
        assert node.get_attr(
            "dilations") is None, 'only dilations = 0 is supported'  # optional

        kernel_shape = node.get_attr("kernel_shape")
        poolnd = len(kernel_shape)
        strides = node.get_attr("strides")
        pad_mode = node.get_attr("pads")
        ceil_mode = bool(node.get_attr('ceil_mode', 0))  # optional
        pads = node.get_attr('pads', [0] * (poolnd * 2))  # optional
        fluid_op = 'pool{}d'.format(poolnd)
        assert 2 <= poolnd <= 3, 'only pool2d and pool3d is supported'
C
channingss 已提交
1230

C
channingss 已提交
1231 1232
        paddings, val_x = self._pad_if_asymmetric(node, pads, val_x)

C
channingss 已提交
1233
        if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER":
C
channingss 已提交
1234
            input_shape = val_x.out_shapes[0]
C
channingss 已提交
1235 1236 1237 1238 1239 1240
            pad_h = get_same_padding(input_shape[2], kernel_shape[0],
                                     strides[0])
            pad_w = get_same_padding(input_shape[3], kernel_shape[1],
                                     strides[1])
            attr = {"paddings": pad_h + pad_w, "pad_value": 0.0}

C
update  
channingss 已提交
1241 1242 1243 1244 1245 1246 1247 1248 1249
        attr = {
            "pool_size": kernel_shape,
            "pool_type": string("max"),
            "pool_stride": strides,
            "pool_padding": paddings,
            "ceil_mode": ceil_mode,
            "name": string(node.layer_name),
            "exclusive": False
        }
1250 1251
        node.fluid_code.add_layer(
            fluid_op, inputs=val_x, output=node, param_attr=attr)
C
update  
channingss 已提交
1252

C
channings 已提交
1253
    def _global_pool(self, node):
C
channingss 已提交
1254
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
C
update  
channingss 已提交
1255
        val_y = self.graph.get_node(node.layer.output[0], copy=True)
1256
        fluid_op = 'pool2d'
C
channings 已提交
1257 1258 1259 1260 1261 1262
        pool_type = None
        if node.layer.op_type == 'GlobalMaxPool':
            pool_type = 'max'
        elif node.layer.op_type == 'GlobalAveragePool':
            pool_type = 'avg'

C
update  
channingss 已提交
1263
        attr = {
C
channings 已提交
1264
            "pool_type": string(pool_type),
C
update  
channingss 已提交
1265 1266 1267
            "global_pooling": True,
            "name": string(node.layer_name)
        }
1268 1269
        node.fluid_code.add_layer(
            fluid_op, inputs=val_x, output=node, param_attr=attr)
R
root 已提交
1270

1271
    @print_mapping_info
C
channings 已提交
1272 1273
    def GlobalMaxPool(self, node):
        self._global_pool(node)
R
root 已提交
1274

1275
    @print_mapping_info
C
channings 已提交
1276 1277
    def GlobalAveragePool(self, node):
        self._global_pool(node)
R
root 已提交
1278

1279
    @print_mapping_info
C
update  
channingss 已提交
1280
    def Conv(self, node):
C
channingss 已提交
1281 1282
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_w = self.graph.get_input_node(node, idx=1, copy=True)
C
update  
channingss 已提交
1283 1284 1285 1286 1287 1288
        val_y = self.graph.get_node(node.layer.output[0], copy=True)

        self.omit_nodes.append(val_w.layer_name)

        has_bias = len(node.layer.input) == 3
        if has_bias:
C
channingss 已提交
1289
            val_b = self.graph.get_input_node(node, idx=2, copy=True)
C
update  
channingss 已提交
1290 1291 1292
            self.omit_nodes.append(val_b.layer_name)
        auto_pad = node.get_attr('auto_pad', 'NOTSET')

C
channingss 已提交
1293
        kernel_shape = node.get_attr('kernel_shape')
C
update  
channingss 已提交
1294 1295
        convnd = len(kernel_shape)
        assert 2 <= convnd <= 3, 'only conv2d and conv3d is supported'
C
channingss 已提交
1296
        num_out_channels = val_w.out_shapes[0][0]  # OI...
C
update  
channingss 已提交
1297 1298 1299 1300 1301 1302 1303
        fluid_op = 'conv{}d'.format(convnd)

        num_groups = node.get_attr('group', 1)
        strides = node.get_attr('strides', [1] * convnd)  # optional
        dilations = node.get_attr('dilations', [1] * convnd)  # optional
        pads = node.get_attr('pads', [0] * (convnd * 2))  # optional

C
channingss 已提交
1304
        input_shape = val_x.out_shapes[0]
C
update  
channingss 已提交
1305 1306
        paddings, val_x = self._pad_if_asymmetric(node, pads, val_x)

C
channingss 已提交
1307
        if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER":
C
Channingss 已提交
1308 1309 1310 1311 1312
            pad_h = _get_same_padding(input_shape[2], kernel_shape[0],
                                      strides[0])
            pad_w = _get_same_padding(input_shape[3], kernel_shape[1],
                                      strides[1])
            paddings = pad_h + pad_w
C
update  
channingss 已提交
1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327

        attr = {
            "num_filters": num_out_channels,
            "filter_size": kernel_shape,
            "stride": strides,
            "padding": paddings,
            "dilation": dilations,
            "groups": num_groups,
            'param_attr': string(val_w.layer_name),
            "name": string(node.layer_name)
        }
        if has_bias:
            attr["bias_attr"] = string(val_b.layer_name)
        else:
            attr["bias_attr"] = False
1328 1329
        node.fluid_code.add_layer(
            fluid_op, inputs=val_x, output=node, param_attr=attr)
C
channingss 已提交
1330

1331
    @print_mapping_info
C
channingss 已提交
1332
    def ConvTranspose(self, node):
C
channingss 已提交
1333 1334
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_w = self.graph.get_input_node(node, idx=1, copy=True)
C
channingss 已提交
1335
        val_b = None
R
root 已提交
1336
        if len(node.layer.input) > 2:
C
channingss 已提交
1337 1338
            val_b = self.graph.get_input_node(node, idx=2, copy=True)
            self.omit_nodes.append(val_b.layer_name)
C
channingss 已提交
1339 1340 1341 1342 1343 1344
        self.omit_nodes.append(val_w.layer_name)

        val_y = self.graph.get_node(node.layer.output[0], copy=True)

        auto_pad = node.get_attr('auto_pad', 'NOTSET')
        out_padding = node.get_attr('output_padding', [0, 0])
C
channingss 已提交
1345
        kernel_shape = node.get_attr('kernel_shape')
C
channingss 已提交
1346 1347 1348
        assert kernel_shape, 'kernel_shape not inferred'
        convnd = len(kernel_shape)
        assert 2 <= convnd <= 3, 'only conv2d_transpose and conv3d_transpose supported'
C
channingss 已提交
1349
        num_out_channels = val_w.out_shapes[0][1]
C
channingss 已提交
1350 1351
        fluid_op = 'conv{}d_transpose'.format(convnd)

C
channingss 已提交
1352 1353 1354 1355 1356
        num_groups = node.get_attr('group', 1)
        strides = node.get_attr('strides', [1] * convnd)
        dilations = node.get_attr('dilations', [1] * convnd)
        output_size = node.get_attr('output_shape', [])
        pads = node.get_attr('pads', [0] * (convnd * 2))
C
channingss 已提交
1357 1358 1359 1360

        paddings, var_x = self._pad_if_asymmetric(node, pads, val_x)

        output_size = [0, 0]
C
channingss 已提交
1361

1362 1363
        output_size[0] = (val_x.out_shapes[0][2] - 1
                          ) * strides[0] - 2 * paddings[0] + dilations[0] * (
C
channingss 已提交
1364
                              kernel_shape[0] - 1) + 1 + out_padding[0]
1365 1366
        output_size[1] = (val_x.out_shapes[0][3] - 1
                          ) * strides[1] - 2 * paddings[1] + dilations[1] * (
C
channingss 已提交
1367 1368 1369 1370 1371 1372 1373 1374 1375 1376
                              kernel_shape[1] - 1) + 1 + out_padding[1]
        attr = {
            'num_filters': num_out_channels,
            'output_size': output_size or None,
            'filter_size': kernel_shape,
            'padding': paddings,
            'stride': strides,
            'dilation': dilations,
            'groups': num_groups,
            'param_attr': string(val_w.layer_name),
C
channingss 已提交
1377
            'bias_attr': None if val_b is None else string(val_b.layer_name),
C
channingss 已提交
1378 1379
            'name': string(node.layer_name),
        }
1380 1381
        node.fluid_code.add_layer(
            fluid_op, inputs=val_x, output=node, param_attr=attr)
C
channings 已提交
1382

1383
    @print_mapping_info
C
channings 已提交
1384 1385 1386 1387
    def GRU(self, node):
        val_x = self.graph.get_input_node(node, idx=0, copy=True)
        val_w = self.graph.get_input_node(node, idx=1, copy=True)
        val_r = self.graph.get_input_node(node, idx=2, copy=True)
R
root 已提交
1388

C
channings 已提交
1389 1390 1391 1392 1393
        val_b = None
        val_len = None
        val_xh = None
        miss_arg_num = 0
        num_ipt = len(node.layer.input)
R
root 已提交
1394
        if num_ipt > 3 and node.layer.input[3] != '':
C
channings 已提交
1395 1396 1397
            val_b = self.graph.get_input_node(node, idx=3, copy=True)
        else:
            miss_arg_num += 1
R
root 已提交
1398
        if num_ipt > 4 and node.layer.input[4] != '':
1399 1400
            val_len = self.graph.get_input_node(
                node, idx=4 - miss_arg_num, copy=True)
C
channings 已提交
1401 1402
        else:
            miss_arg_num += 1
R
root 已提交
1403
        if num_ipt > 5 and node.layer.input[5] != '':
1404 1405
            val_xh = self.graph.get_input_node(
                node, idx=5 - miss_arg_num, copy=True)
R
root 已提交
1406

C
channings 已提交
1407
        x_shape = val_x.out_shapes[0]
C
Channingss 已提交
1408 1409

        assert x_shape[1] == 1, 'only X with batch_size = 1 supported'
C
channings 已提交
1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428
        assert node.get_attr('clip', None) is None, 'clipping not supported'

        hidden_size = node.get_attr('hidden_size', None)
        if hidden_size is None:
            r_shape = val_r.out_shapes[0]
            if r_shape:
                hidden_size = r_shape[-1]
        if hidden_size is None:
            w_shape = var_w.out_shapes[0]
            if w_shape:
                hidden_size = w_shape[-2] // 3
        if hidden_size is None and val_b:
            b_shape = val_b.out_shapes[0]
            if b_shape:
                hidden_size = b_shape[-1] // 6
        if hidden_size is None and val_xh:
            xh_shape = val_xh.out_shapes[0]
            if xh_shape:
                hidden_size = xh_shape[-1]
R
root 已提交
1429 1430

        direction = node.get_attr('direction', 'forward')
C
channings 已提交
1431
        assert direction != 'bidirectional', 'direction = bidirectional not supported'
R
root 已提交
1432

C
channings 已提交
1433 1434
        activations = node.get_attr('activations', ['Sigmoid', 'Tanh'])
        assert len(activations) == 2, 'bidirectional operation not supported'
R
root 已提交
1435 1436 1437 1438

        assert node.get_attr('linear_before_reset',
                             0) == 0, 'only linear_before_reset = 0 supported'

C
channings 已提交
1439 1440 1441
        activations = [s.lower() for s in activations]
        gate_activation, candidate_activation = activations
        is_reverse = direction == 'reverse'
R
root 已提交
1442

C
channings 已提交
1443
        var_x0 = node.layer_name + '_x0'
1444 1445 1446 1447 1448 1449
        node.fluid_code.add_layer(
            'squeeze',
            inputs=val_x,
            output=var_x0,
            param_attr={'axes': [1],
                        'name': string(var_x0)})
R
root 已提交
1450

C
channings 已提交
1451
        var_w0 = node.layer_name + '_w0'
1452 1453 1454 1455 1456 1457
        node.fluid_code.add_layer(
            'squeeze',
            inputs=val_w,
            output=var_w0,
            param_attr={'axes': [0],
                        'name': string(var_w0)})
R
root 已提交
1458

C
channings 已提交
1459 1460
        var_fc = node.layer_name + '_fc'
        var_mm = (node.layer_name + '_mm') if val_b else var_fc
1461 1462 1463 1464 1465 1466 1467 1468 1469 1470
        node.fluid_code.add_layer(
            'matmul',
            inputs={'x': var_x0,
                    'y': var_w0},
            output=var_mm,
            param_attr={
                'transpose_x': 0,
                'transpose_y': 1,
                'name': string(var_mm)
            })
R
root 已提交
1471

C
channings 已提交
1472
        var_r0 = node.layer_name + '_r0'
1473 1474 1475 1476 1477 1478
        node.fluid_code.add_layer(
            'squeeze',
            inputs=val_r,
            output=var_r0,
            param_attr={'axes': [0],
                        'name': string(var_r0)})
R
root 已提交
1479 1480 1481

        var_r0t = node.layer_name + '_r0t'

1482 1483 1484 1485 1486 1487
        node.fluid_code.add_layer(
            'transpose',
            inputs=var_r0,
            output=var_r0t,
            param_attr={'perm': [1, 0],
                        'name': string(var_r0t)})
C
channings 已提交
1488 1489 1490
        if val_b:
            var_bi = node.layer_name + '_bi'
            var_bh = node.layer_name + '_bh'
1491 1492 1493 1494 1495
            node.fluid_code.add_layer(
                'split',
                inputs=val_b,
                output=var_bi + ',' + var_bh,
                param_attr={
C
Channingss 已提交
1496 1497
                    'dim': 1,
                    'num_or_sections': [hidden_size * 3, hidden_size * 3],
1498 1499
                    'name': string(node.layer_name + '.b/split')
                })
C
channings 已提交
1500
            var_bi0 = node.layer_name + '_bi0'
1501 1502 1503 1504 1505 1506 1507 1508
            node.fluid_code.add_layer(
                'squeeze',
                inputs=var_bi,
                output=var_bi0,
                param_attr={'axes': [0],
                            'name': string(var_bi0)})

            node.fluid_code.add_layer(
C
Channingss 已提交
1509
                'elementwise_add',
1510 1511 1512 1513 1514 1515
                inputs=[var_mm, var_bi0],
                output=var_fc,
                param_attr={
                    'axes': 1,
                    'name': string(node.layer_name + '.i/bias')
                })
C
channings 已提交
1516 1517 1518

        if val_xh:
            var_xh0 = node.layer_name + '_xh0'
1519 1520 1521 1522 1523 1524
            node.fluid_code.add_layer(
                'squeeze',
                inputs=val_xh,
                output=var_xh0,
                param_attr={'axes': [1],
                            'name': string(var_xh0)})
C
channings 已提交
1525
        var_y00 = node.layer_name + '_y00'
R
root 已提交
1526 1527 1528

        attr = {
            'origin_mode': True,
C
channings 已提交
1529
            'h_0': var_xh0 if val_xh else None,
R
root 已提交
1530 1531 1532 1533 1534
            'is_reverse': is_reverse,
            'gate_activation': string(gate_activation),
            'candidate_activation': string(candidate_activation),
            'param_attr': string(var_r0t),
            'bias_attr': string(var_bh) if val_b else False,
C
channings 已提交
1535
        }
1536 1537 1538 1539 1540
        node.fluid_code.add_layer(
            'dynamic_gru',
            inputs=var_fc + ',' + str(hidden_size),
            output=var_y00,
            param_attr=attr)
R
root 已提交
1541

C
channings 已提交
1542
        num_opt = len(node.layer.output)
R
root 已提交
1543 1544

        if num_opt > 0 and node.layer.output[0] != '':
1545 1546 1547 1548 1549 1550 1551 1552
            node.fluid_code.add_layer(
                'unsqueeze',
                inputs=var_y00,
                output=node.layer.output[0],
                param_attr={
                    'axes': [1, 1],
                    'name': string(node.layer.output[0])
                })
R
root 已提交
1553
        if num_opt > 1 and node.layer.output[1] != '':
1554 1555 1556 1557 1558 1559 1560 1561
            node.fluid_code.add_layer(
                'unsqueeze',
                inputs=var_y00,
                output=node.layer.output[1],
                param_attr={
                    'axes': [1, 1],
                    'name': string(node.layer.output[1])
                })