# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License" # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode from x2paddle.core.graph import GraphNode from x2paddle.core.util import * from functools import reduce import numpy as np import onnx import onnx.numpy_helper as numpy_helper from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE import logging as _logging from collections import OrderedDict import math import os import copy import sys import shutil _logger = _logging.getLogger(__name__) def _const_weight_or_none(node, necessary=False): if 'Constant' in node.layer_type: return node.value if isinstance(node, ONNXGraphDataNode): return node.weight if necessary: assert '{} should be an initializer or Constant operator.'.format( node.name) return None def _rename_or_remove_weight(weights, origin_name, target_name=None, is_remove=True): ''' Rename parameters by Paddle's naming rule of parameters. Args: weights(dict[String:np.ndarray]): Dict stored paramters, the key in weights is name of parameter. origin_name(String): Name of parameter to rename or remove. target_name(String, optional): if target_name is not None, add new key-value pair {target_name:weights[origin_name]} to weights, and target_name must follow paddle's naming rule of parameters. Default: None. is_remove: if is_remove is True, remove origin key-value pair. Default: True. Returns: None ''' if origin_name not in weights: raise KeyError('{} not a key in {}'.format(origin_name, weights)) if is_remove: # remove weight data = weights.pop(origin_name) else: data = weights[origin_name] if target_name is not None: # rename weight weights[target_name] = data def _is_static_shape(shape): negtive_dims = 0 error_dims = 0 for dim in shape: if dim < 0: negtive_dims += 1 if dim < -1: error_dims += 1 if negtive_dims > 1: return False if error_dims > 0: return False return True def _get_same_padding(in_size, kernel_size, stride): new_size = int(math.ceil(in_size * 1.0 / stride)) pad_size = (new_size - 1) * stride + kernel_size - in_size pad0 = int(pad_size / 2) pad1 = pad_size - pad0 return [pad0, pad1] 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.name[9:], node.layer_type)) raise else: return res return run_mapping class OpSet9(): elementwise_ops = { 'Add': 'paddle.add', 'Div': 'paddle.divide', 'Sub': 'paddle.subtract', 'Mul': 'paddle.multiply', 'Pow': 'paddle.pow', } directly_map_ops = { 'Ceil': ['paddle.ceil'], # reduce function 'ReduceMean': ['paddle.mean', dict(axes='axis', keepdims='keepdim'), dict(axes=None, keepdims=1)], 'ReduceSum': ['paddle.sum', dict(axes='axis', keepdims='keepdim'), dict(axes=None, keepdims=1)], 'ReduceMin': ['paddle.min', dict(axes='axis', keepdims='keepdim'), dict(axes=None, keepdim=1)], 'ReduceMax': ['paddle.max', dict(axes='axis', keepdims='keepdim'), dict(axes=None, keepdim=1)], 'ReduceProd': ['paddle.prod', dict(axes='axis', keepdims='keepdim'), dict(axes=None, keepdim=1)], # active function 'Relu': ['paddle.nn.ReLU'], 'LeakyRelu': ['paddle.nn.LeakyReLU', dict(alpha='negative_slope'), dict(negative_slope=.01)], 'Elu': ['paddle.nn.functional.elu', dict(alpha='alpha'), dict(alpha=1.)], 'ThresholdedRelu': ['paddle.nn.functional.thresholded_relu', dict(alpha='threshold'), dict(alpha=1.)], 'Tanh': ['paddle.nn.Tanh'], 'Sigmoid': ['paddle.nn.Sigmoid'], 'Softsign': ['paddle.nn.Softsign'], 'Softplus': ['paddle.nn.Softplus', dict(threshold='threshold'), dict(threshold=float(sys.maxsize))], 'Exp': ['paddle.exp'], 'Log': ['paddle.log'], 'LogSoftmax': ['paddle.nn.functional.log_softmax', dict(axis='axis'), dict(axis=1)], 'Softmax': ['paddle.nn.Softmax', dict(axis='axis'), dict(axis=1)], 'Sqrt': ['paddle.sqrt'], 'Floor': ['paddle.floor'], 'Abs': ['paddle.abs'], 'Erf': ['paddle.erf'], } def __init__(self, decoder, paddle_graph): super(OpSet9, self).__init__() self.graph = decoder.graph self.paddle_graph = paddle_graph self.input_index = 0 self.inputs_info = dict() self.weights = dict() self.nn_name2id = dict() self.done_weight_list = list() @print_mapping_info def directly_map(self, node, *args, **kwargs): inputs = node.layer.input assert len(inputs) == 1, 'directly_map error with multi inputs' input = self.graph.get_input_node(node, idx=0, copy=True) onnx_attrs = node.attr_map if '' in onnx_attrs: onnx_attrs.pop('') if '_' in onnx_attrs: onnx_attrs.pop('_') op_info = self.directly_map_ops[node.layer_type] paddle_op = op_info[0] layer_attrs = dict() if len(op_info) > 1: attrs_name_map_dict = op_info[1] for onnx_attr_name, pd_attr_name in attrs_name_map_dict.items(): if onnx_attr_name in onnx_attrs: layer_attrs[pd_attr_name] = onnx_attrs[onnx_attr_name] else: layer_attrs[pd_attr_name] = op_info[2][onnx_attr_name] if paddle_op.startswith("paddle.nn") and 'functional' not in paddle_op: op_name = paddle_op[10:].lower() op_name = name_generator(op_name, self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] self.paddle_graph.add_layer( kernel=paddle_op, inputs={"x": input.name}, outputs=layer_outputs, **layer_attrs) else: self.paddle_graph.add_layer( kernel=paddle_op, inputs={"x": input.name}, outputs=[node.name], **layer_attrs) @print_mapping_info def elementwise_map(self, node): op_type = self.elementwise_ops[node.layer_type] val_x = self.graph.get_input_node(node, idx=0, copy=True) val_y = self.graph.get_input_node(node, idx=1, copy=True) inputs_dict = {'x': val_x.name, 'y': val_y.name} self.paddle_graph.add_layer( op_type, inputs=inputs_dict, outputs=[node.name]) @print_mapping_info def place_holder(self, node): shape = node.out_shapes[0] for i, dim_shape in enumerate(shape): if dim_shape == 0 and i == 0: shape[i] = 1 if dim_shape == 0 and i != 0: assert 'shape of input is not assigned' self.paddle_graph.add_layer( kernel="paddle.to_tensor", inputs={}, outputs=[node.name], data="x{}".format(self.input_index)) self.inputs_info["x{}".format(self.input_index)] = [shape, node.dtype] self.input_index += 1 @print_mapping_info def create_parameter(self, node, parameter=None): if parameter is not None: node = parameter dtype = node.dtype shape = node.out_shapes[0] if hasattr(node.weight, "shape") and len(node.weight.shape) == 0: self.paddle_graph.add_layer( "paddle.full", inputs={}, outputs=[node.name], dtype=string(dtype), shape=[1], fill_value=node.weight) else: self.weights[node.name] = node.weight self.paddle_graph.add_layer( "self.create_parameter", inputs={}, outputs=[node.name], shape=shape, attr=string(node.name), dtype=string(dtype), default_initializer="paddle.nn.initializer.Constant(value=0.0)") 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 def _interpolate(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) inputs = {'x': val_x.name} attrs = dict() if node.layer_type == 'Resize': if len(node.layer.input) == 2: # opset 10 val_scales = self.graph.get_input_node(node, idx=1, copy=True) # TODO(syf): paddle.nn.functional.interpolate will support the length # which is the same as the rank of input. attrs['scale_factor'] = self.weights[val_scales.name].tolist()[2:] elif len(node.layer.input) == 3: # opset 11 val_scales = self.graph.get_input_node(node, idx=2, copy=True) # TODO(syf): paddle.nn.functional.interpolate will support the length # which is the same as the rank of input. attrs['scale_factor'] = self.weights[val_scales.name].tolist()[2:] 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.name + '_nc', val_sizes.name + '_hw' self.paddle_graph.add_layer( 'paddle.split', inputs={"x": val_sizes.name}, outputs=[var_nc, var_hw], num_or_sections=[2, 2], axis=0) self.paddle_graph.add_layer( "paddle.cast", inputs={"x": var_hw}, outputs=[var_hw], dtype=string('int32')) inputs['size'] = var_hw attrs = {"align_corners": False, "mode": string(node.get_attr('mode', 'nearest'))} self.paddle_graph.add_layer( kernel="paddle.nn.functional.interpolate", inputs=inputs, outputs=[node.name], **attrs) return elif node.layer_type == 'Upsample': val_scales = self.graph.get_input_node(node, idx=1, copy=True) self.paddle_graph.add_layer( "paddle.slice", inputs={"input": val_scales.name}, outputs=[val_scales.name], axes=[0], starts=[2], ends=[4]) inputs['scale_factor'] = val_scales.name mode = node.get_attr('mode', 'nearest') attrs.update({"align_corners": False, "mode": string(mode), "align_mode": 1}) val_x_shape = val_x.out_shapes[0] if mode == "linear" and len(val_x_shape) == 4: attrs["mode"] = string("bilinear") attrs["align_corners"] = True self.paddle_graph.add_layer( kernel="paddle.nn.functional.interpolate", inputs=inputs, outputs=[node.name], **attrs) @print_mapping_info def HardSigmoid(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) alpha = node.get_attr('alpha', 0.2) beta = node.get_attr('beta', 0.5) self.paddle_graph.add_layer( kernel="paddle.scale", inputs={"x": val_x.name}, outputs=[node.name + "_val"], scale=alpha, bias=beta) self.paddle_graph.add_layer( kernel="paddle.clip", inputs={"x": node.name + "_val"}, outputs=[node.name], min=0.0, max=1.0) @print_mapping_info def Shape(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) self.paddle_graph.add_layer( kernel="paddle.shape", inputs={"input": val_x.name}, outputs=[node.name]) self.paddle_graph.add_layer( 'paddle.cast', inputs={"x": node.name}, outputs=[node.name], dtype=string('int64')) @print_mapping_info 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) pooled_height = node.get_attr('output_height') pooled_width = node.get_attr('output_width') spatial_scale = node.get_attr('spatial_scale') sampling_ratio = node.get_attr('sampling_ratio') layer_attrs = { 'pooled_height': pooled_height, 'pooled_width': pooled_width, 'spatial_scale': spatial_scale, 'sampling_ratio': sampling_ratio, } self.paddle_graph.add_layer( 'paddle.fluid.layers.roi_align', inputs={'input': val_x.name, 'rois': val_rois.name}, outputs=[node.name], **layer_attrs) @print_mapping_info 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) spatial_scale = node.get_attr('spatial_scale') pooled_height, pooled_width = node.get_attr('pooled_shape') layer_attrs = { 'pooled_height': pooled_height, 'pooled_width': pooled_width, 'spatial_scale': spatial_scale, } self.paddle_graph.add_layer( 'paddle.fluid.layers.roi_pool', inputs={'input': val_x.name, 'rois': val_rois.name}, outputs=[node.name], **layer_attrs) @print_mapping_info def Pad(self, node, op_independent=True): val_x = self.graph.get_input_node(node, idx=0, copy=True) pads = node.get_attr('pads') is_pads_attr = True if pads is None: val_pad = self.graph.get_input_node(node, idx=1, copy=True) pad_shape = val_pad.out_shapes[0] is_pads_attr = False pads = _const_weight_or_none(val_pad) if pads is not None: is_pads_attr = True mode = node.get_attr('mode', 'constant') value = node.get_attr('value', 0.) data_shape = val_x.out_shapes[0] output_shape = node.out_shapes[0] assume_pad = False layer_attrs = {} layer_attrs['mode'] = string(mode) layer_attrs['value'] = value if not op_independent: output_name = node.name + '_paded' else: output_name = node.name nn_op_name = name_generator("pad", self.nn_name2id) layer_outputs = [nn_op_name, output_name] if is_pads_attr: paddings = [] if len(pads) in [2, 4, 6]: if data_shape: assume_pad |= data_shape and 2 * (len(data_shape) - 2) == len(pads) # NCHW if output_shape: assume_pad |= output_shape and 2 * (len(output_shape) - 2) == len(pads) # NCHW if assume_pad: paddle_op = 'paddle.nn.Pad{}D'.format(len(output_shape) - 2) paddings = np.array(pads).reshape( (2, -1)).transpose().astype("int32") paddings = np.flip(paddings, axis=0).flatten().tolist() layer_attrs['padding'] = paddings else: if data_shape: assume_pad |= data_shape and 2 * len(data_shape) == len(pads) # NCHW if output_shape: assume_pad |= output_shape and 2 * len(output_shape) == len(pads) # NCHW if assume_pad: paddle_op = 'paddle.nn.functional.pad' paddings = np.array(pads).reshape( (2, -1)).transpose().astype("int32").flatten().tolist() layer_attrs['pad'] = paddings else: raise Exception("The padding value {} is wrong!".format(pads)) elif len(pads) == 8: if data_shape: assume_pad |= data_shape and 2 * len(data_shape) == len(pads) # NCHW if output_shape: assume_pad |= output_shape and 2 * len(output_shape) == len(pads) # NCHW if assume_pad: paddle_op = 'paddle.nn.Pad2D' paddings = np.array(pads).reshape( (2, -1)).transpose().astype("int32") paddings = np.flip(paddings, axis=0).flatten().tolist() if sum(paddings[:4]) == 0: paddings = paddings[4:] layer_attrs['padding'] = paddings else: layer_attrs["pad"] = paddings paddle_op = "custom_layer:PadAllDim4WithOneInput" else: raise Exception("The padding value {} is wrong!".format(pads)) self.paddle_graph.add_layer( paddle_op, inputs={'x': val_x.name}, outputs=layer_outputs[1:] if paddle_op == 'paddle.nn.functional.pad' else layer_outputs, **layer_attrs) if not op_independent: return node.name + '_paded' else: pads_len = val_pad.out_shapes[0][0] if pads_len in [2, 4, 6]: if data_shape: assume_pad |= data_shape and 2 * (len(data_shape) - 2) == pads_len # NCHW if output_shape: assume_pad |= output_shape and 2 * (len(output_shape) - 2) == pads_len # NCHW if assume_pad: if pads_len == 2: data_format = "NCL" elif pads_len == 4: data_format = "NCHW" else: data_format = "NCDHW" self.paddle_graph.add_layer( "custom_layer:PadWithTwoInput", inputs={'x': val_x.name, 'pad': val_pad.name}, outputs=layer_outputs, value=value, mode=string(mode), data_format=string(data_format)) else: if data_shape: assume_pad |= data_shape and 2 * len(data_shape) == pads_len # NCHW if output_shape: assume_pad |= output_shape and 2 * len(output_shape) == pads_len # NCHW if assume_pad: if pads_len == 4: self.paddle_graph.add_layer( "custom_layer:PadAllDim2", inputs={'x': val_x.name, 'pad': val_pad.name}, outputs=layer_outputs, value=value, mode=string(mode)) else: raise Exception("The padding value is wrong!") elif pads_len == 8: if data_shape: assume_pad |= data_shape and 2 * len(data_shape) == pads_len # NCHW if output_shape: assume_pad |= output_shape and 2 * len(output_shape) == pads_len # NCHW if assume_pad: self.paddle_graph.add_layer( "custom_layer:PadAllDim4", inputs={'x': val_x.name, 'pad': val_pad.name}, outputs=layer_outputs, value=value, mode=string(mode)) else: print(pads_len) raise Exception("The padding value is wrong!") if not op_independent: return node.name + '_paded' @print_mapping_info def Unsqueeze(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) axes = node.get_attr('axes') layer_attrs = {'axis': axes} if len(val_x.out_shapes[0]) == 0: if node.name: self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": val_x.name}, outputs=[node.name], shape=[1]) else: self.paddle_graph.add_layer( 'paddle.unsqueeze', inputs={"x": val_x.name}, outputs=[node.name], **layer_attrs) @print_mapping_info def Shrink(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) bias = node.get_attr('bias') lambd = node.get_attr('lambd') assert bias == 0.0, 'not support bias!=0' self.paddle_graph.add_layer( 'paddle.nn.functional.hardshrink', inputs={"x": val_x.name}, outputs=[node.name], threshold=lambd) @print_mapping_info 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' shape = node.get_attr('shape', None) if shape is None: shape = val_output.out_shapes[0] if shape is None: shape = list(value.shape) _logger.warning('in (Constant -> %s): ' 'attribute "shape" of %s not inferred, ' 'using value as 1-D tensor may lead to fails', val_output.name, val_output.name) if len(value) == 1: value = value.tolist() value = value[0] self.paddle_graph.add_layer( "paddle.full", inputs={}, outputs=[node.name], dtype=string(dtype), shape=[1], fill_value=value) else: value = np.reshape(value, shape) self.weights[node.name] = value self.paddle_graph.add_layer( "self.create_parameter", inputs={}, outputs=[node.name], shape=shape, attr=string(node.name), dtype=string(dtype), default_initializer="paddle.nn.initializer.Constant(value=0.0)") @print_mapping_info def Resize(self, node): self._interpolate(node) @print_mapping_info def Upsample(self, node): self._interpolate(node) @print_mapping_info def InstanceNormalization(self, node): op_name = name_generator("instanse_norm", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] 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) self.weights[op_name+'.scale'] = self.weights[val_scale.name] self.weights[op_name+'.bias'] = self.weights[val_b.name] layer_attrs = { 'num_features': node.out_shapes[0][1], 'epsilon': epsilon, } dim = len(val_x.out_shapes[0]) if dim == 3: paddle_op = "paddle.nn.InstanceNorm1D" elif dim == 4: paddle_op = "paddle.nn.InstanceNorm2D" elif dim == 5: paddle_op = "paddle.nn.InstanceNorm3D" else: raise Exception("The paddle only support 2D, 3D, 4D or 5D input in InstanceNormalization.") self.paddle_graph.add_layer( paddle_op, inputs={"x": val_x.name}, outputs=layer_outputs, **layer_attrs) @print_mapping_info def Expand(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) val_shape = self.graph.get_input_node(node, idx=1, copy=True) val_x_dtype = val_x.dtype name_ones = node.name + '_ones' attr_ones = { 'shape': val_shape.name, 'dtype': string(val_x_dtype), 'fill_value': 1 } self.paddle_graph.add_layer( 'paddle.full', inputs={}, outputs=[name_ones], **attr_ones) inputs_dict = {'x': name_ones, 'y': val_x.name} self.paddle_graph.add_layer( 'paddle.multiply', inputs=inputs_dict, outputs=[node.name]) @print_mapping_info 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] axis = node.get_attr('axis', 0) #assert len( # indices_shape) <= 2, "Gather op don't support dim of indice >2 " if axis == 0 and len(indices_shape) <= 1: if len(val_x.out_shapes[0]) <= 1: self.paddle_graph.add_layer( 'paddle.gather', inputs={'x': val_x.name, 'index': indices.name}, outputs=[node.name]) elif len(val_x.out_shapes[0]) > 1: if len(indices_shape) == 0: gather_ = node.name + '_1' self.paddle_graph.add_layer( 'paddle.gather', inputs={'x': val_x.name, 'index': indices.name}, outputs=[gather_]) self.paddle_graph.add_layer( 'paddle.squeeze', inputs={'x': gather_}, outputs=[node.name], axis=[0]) else: self.paddle_graph.add_layer( 'paddle.gather', inputs={'x': val_x.name, 'index': indices.name}, outputs=[node.name]) elif axis > 0 and len(indices_shape) <= 1: perm = list(range(len(val_x.out_shapes[0]))) perm = [axis] + perm[:axis] + perm[axis + 1:] name_trans = val_x.name + '_trans' self.paddle_graph.add_layer( 'paddle.transpose', inputs={"x": val_x.name}, outputs=[name_trans], perm=perm) self.paddle_graph.add_layer( 'paddle.gather', inputs={'x': name_trans, 'index': indices.name}, outputs=[node.name]) self.paddle_graph.add_layer( 'paddle.transpose', inputs={"x": node.name}, outputs=[node.name], perm=perm) if len(indices_shape) < 1: self.paddle_graph.add_layer( 'paddle.squeeze', inputs={'x': node.name}, outputs=[node.name], axis=[axis]) elif axis == 0 and len(indices_shape) > 1: if val_x.out_shapes[0] is not None and isinstance( val_x, ONNXGraphDataNode): indices_cast = indices.name + '_cast' self.paddle_graph.add_layer( 'paddle.cast', inputs={"x": indices.name}, outputs=[indices_cast], dtype=string('int64')) op_name = name_generator("embedding", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] self.weights[op_name + '.weight'] = _const_weight_or_none(val_x) self.paddle_graph.add_layer( 'paddle.nn.Embedding', inputs={"x": indices_cast}, outputs=layer_outputs, num_embeddings=val_x.out_shapes[0][0], embedding_dim=val_x.out_shapes[0][1]) else: from functools import reduce reshape_shape = reduce(lambda x, y: x * y, indices_shape) indices_reshape = indices.name + '_shape' self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": indices.name}, outputs=[indices_reshape], shape=[reshape_shape, ]) perm = list(range(len(val_x.out_shapes[0]))) self.paddle_graph.add_layer( 'paddle.gather', inputs={'x': val_x.name, 'index': indices_reshape}, outputs=[node.name]) 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) self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": node.name}, outputs=[node.name], shape=reshaped_shape) elif axis > 0 and len(indices_shape) > 1: from functools import reduce reshape_shape = reduce(lambda x, y: x * y, indices_shape) indices_reshape = indices.name + '_shape' self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": indices.name}, outputs=[indices_reshape], shape=[reshape_shape, ]) perm = list(range(len(val_x.out_shapes[0]))) perm = [axis] + perm[:axis] + perm[axis + 1:] name_trans = val_x.name + '_transpose' self.paddle_graph.add_layer( 'paddle.transpose', inputs={"x": val_x.name}, outputs=[name_trans], perm=perm) self.paddle_graph.add_layer( 'paddle.gather', inputs={'x': name_trans, 'index': indices_reshape}, outputs=[node.name]) input_transpose = node.name + '_transpose' self.paddle_graph.add_layer( 'paddle.transpose', inputs={"x": node.name}, outputs=[input_transpose], perm=perm) 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) self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": input_transpose}, outputs=[node.name], shape=reshaped_shape) @print_mapping_info def ScatterND(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) updates = self.graph.get_input_node(node, idx=2, copy=True) if len(indices.out_shapes[0]) == 1: self.paddle_graph.add_layer( 'paddle.scatter', inputs={'x': val_x.name, 'index': indices.name, 'updates': updates.name}, outputs=[node.name]) else: input_inner_indices = node.name + '_input_inner_indices' shape = val_x.out_shapes[0] self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": indices.name}, outputs=[indices.name], shape=indices.out_shapes[0]) zeros_like_val_x = val_x.name + '_zeros' self.paddle_graph.add_layer( 'paddle.zeros_like', inputs={"x": val_x.name}, outputs=[zeros_like_val_x]) self.paddle_graph.add_layer( 'paddle.scatter_nd_add', inputs={ 'x': zeros_like_val_x, 'index': indices.name, 'updates': updates.name }, outputs=[input_inner_indices]) indices_mask = node.name + '_indices_mask' constant_minus_one = node.name + '_constant_minus_one' # full_like support create tensor shape like input tensor self.paddle_graph.add_layer( 'paddle.full_like', inputs={"x": updates.name}, outputs=[constant_minus_one], dtype=string(updates.dtype), fill_value=-1) self.paddle_graph.add_layer( 'paddle.scatter_nd_add', inputs={ 'x': zeros_like_val_x, 'index': indices.name, 'updates': constant_minus_one }, outputs=[indices_mask]) constant_one = node.name + '_constant_1' # full_like support create tensor shape like input tensor self.paddle_graph.add_layer( 'paddle.full_like', inputs={"x": val_x.name}, outputs=[constant_one], dtype=string(val_x.dtype), fill_value=1) input_out_indices_mask = node.name + '_input_out_indices_mask' self.paddle_graph.add_layer( "paddle.add", inputs={"x": indices_mask, "y": constant_one}, outputs=[input_out_indices_mask]) input_out_indices = node.name + '_input_out_indices' self.paddle_graph.add_layer( "paddle.multiply", inputs={"x": val_x.name, "y": input_out_indices_mask}, outputs=[input_out_indices]) self.paddle_graph.add_layer( "paddle.add", inputs={"x": input_inner_indices, "y": input_out_indices}, outputs=[node.name]) @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.name, 'end': val_limit.name, 'step': val_delta.name} self.paddle_graph.add_layer( 'paddle.arange', inputs=inputs, outputs=[node.name], dtype=string(dtype)) @print_mapping_info def Slice(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) starts, ends, axes, steps = None, None, None, None layer_attrs = {} 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) starts_value = _const_weight_or_none(starts) if starts_value is not None: starts_value = starts_value.tolist() ends_value = _const_weight_or_none(ends) if ends_value is not None: ends_value = ends_value.tolist() if len(node.inputs) > 2: s_len = len(val_x.out_shapes[0]) axes = list(range(s_len)) if len(node.inputs) > 3: axes_node = self.graph.get_input_node(node, idx=3, copy=True) axes = _const_weight_or_none(axes_node, necessary=True).tolist() if len(node.inputs) > 4: steps = self.graph.get_input_node(node, idx=4, copy=True) steps = _const_weight_or_none(steps).tolist() layer_attrs = { "axes": axes, "starts": starts.name, "ends": ends.name } if starts_value is not None and ends_value is not None and axes is not None: starts_value = starts_value.copy() ends_value = ends_value.copy() for idx in range(len(ends_value)): if starts_value[idx] >= val_x.out_shapes[0][axes[idx]]: starts_value[idx] = val_x.out_shapes[0][axes[idx]] - 1 ends_value[idx] = val_x.out_shapes[0][axes[idx]] starts_value[idx] = val_x.out_shapes[0][axes[idx]] - 1 elif ends_value[idx] > 2**31 - 1: ends_value[idx] = 2**31 - 1 layer_attrs = { "axes": axes, "starts": starts_value, "ends": ends_value } else: if starts.dtype != 'int32': starts_cast = starts.name + '_cast' self.paddle_graph.add_layer( 'paddle.cast', inputs={"x": starts.name}, outputs=[starts_cast], dtype=string('int32')) layer_attrs['starts'] = starts_cast if ends.dtype != 'int32': ends_cast = ends.name + '_cast' else: ends_cast = ends.name self.paddle_graph.add_layer( 'paddle.cast', inputs={"x": ends.name}, outputs=[ends_cast], dtype=string('int32')) layer_attrs['ends'] = ends_cast else: starts = node.get_attr('starts') ends = node.get_attr('ends') axes = node.get_attr('axes') for idx in range(len(ends)): if ends[idx] > 2**31 - 1: ends[idx] = 2**31 - 1 layer_attrs = {"axes": axes, "starts": starts, "ends": ends} if steps is not None: layer_attrs['strides'] = steps self.paddle_graph.add_layer( 'paddle.strided_slice', inputs={"x": val_x.name}, outputs=[node.name], **layer_attrs) else: self.paddle_graph.add_layer( 'paddle.slice', inputs={"input": val_x.name}, outputs=[node.name], **layer_attrs) @print_mapping_info def ConstantOfShape(self, node): val_shape = self.graph.get_input_node(node, idx=0, copy=True) val_y = self.graph.get_node(node.layer.output[0], copy=True) value = node.get_attr('value') dtype = value.dtype value = value.tolist() assert len(value) == 1, ('given value not Scalar, shape of value > 1, ' 'this is not supported') if len(value) == 1: value = value[0] layer_attrs = { 'dtype': string(dtype), 'fill_value': value } self.paddle_graph.add_layer( "paddle.full", inputs={'shape': val_shape.name}, outputs=[node.name], **layer_attrs) @print_mapping_info def Clip(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) val_y = self.graph.get_node(node.layer.output[0], copy=True) max_value, min_value = None, None if len(node.inputs) == 1: max_value = node.get_attr('max') min_value = node.get_attr('min') layer_attrs = { 'max': max_value, 'min': min_value, } self.paddle_graph.add_layer( 'paddle.clip', inputs={"x": val_x.name}, outputs=[node.name], **layer_attrs) else: min_ipt = self.graph.get_input_node(node, idx=1, copy=True) max_ipt = self.graph.get_input_node(node, idx=2, copy=True) min_value = _const_weight_or_none(min_ipt) max_value = _const_weight_or_none(max_ipt) if max_value.shape == (1, ): max_value = max_value[0] if min_value.shape == (1, ): min_value = min_value[0] if max_value is not None and min_value is not None: layer_attrs = {'max': max_value, 'min': min_value} self.paddle_graph.add_layer( 'paddle.clip', inputs={"x": val_x.name}, outputs=[node.name], **layer_attrs) else: raise @print_mapping_info def Split(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) paddle_op = 'split' split = node.get_attr('split') axis = node.get_attr('axis', 0) layer_attrs = { 'num_or_sections': split, 'axis': axis, } outputs_list = list() if isinstance(split, list) or isinstance(split, tuple): if len(split) == 1: outputs_list.append(node.name) else: for i in range(len(split)): outputs_list.append("{}_p{}".format(node.layer_name, i)) else: outputs_list.append(node.name) self.paddle_graph.add_layer( 'paddle.split', inputs={"x": val_x.name}, outputs=outputs_list, **layer_attrs) @print_mapping_info def Reshape(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) val_shape = self.graph.get_input_node(node, idx=1, copy=True) val_reshaped = self.graph.get_node(node.layer.output[0], copy=True) shape_value = _const_weight_or_none(val_shape) shape_dims = len(val_shape.out_shapes[0]) if shape_value is not None: self.paddle_graph.add_layer( 'paddle.reshape', inputs={'x': val_x.name}, outputs=[node.name], shape=shape_value.tolist()) elif len(node.out_shapes[0]) > 0 and _is_static_shape(node.out_shapes[ 0]): self.paddle_graph.add_layer( 'paddle.reshape', inputs={'x': val_x.name}, outputs=[node.name], shape=node.out_shapes[0]) else: # shape may be [], come form Gather by scalar indices if len(val_shape.out_shapes[0]) > 0: self.paddle_graph.add_layer( 'paddle.reshape', inputs={'x': val_shape.name}, outputs=[val_shape.name], shape=val_shape.out_shapes[0]) if val_shape.dtype != "int32": self.paddle_graph.add_layer( 'paddle.cast', inputs={'x': val_shape.name}, outputs=[val_shape.name], dtype=string("int32")) self.paddle_graph.add_layer( 'paddle.reshape', inputs={'x': val_x.name, 'shape': val_shape.name}, outputs=[node.name]) @print_mapping_info def Cast(self, node): val_input = self.graph.get_input_node(node, idx=0, copy=True) 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' self.paddle_graph.add_layer( 'paddle.cast', inputs={'x': val_input.name}, outputs=[node.name], dtype=string(dtype)) @print_mapping_info def Not(self, node): val_input = self.graph.get_input_node(node, idx=0, copy=True) self.paddle_graph.add_layer('paddle.logical_not', inputs={'x': val_input.name}, outputs=[node.name]) @print_mapping_info def AveragePool(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) auto_pad = node.get_attr('auto_pad', 'NOTSET') 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)) paddings, val_x = self._pad_if_asymmetric(node, pads, val_x) if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER": input_shape = val_x.out_shapes[0] 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 op_name = name_generator("pool", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] paddle_op = 'paddle.nn.AvgPool{}D'.format(poolnd) assert 1 <= poolnd <= 3, 'only Pool1D, Pool2D and Pool3D are supported' layer_attrs = { "kernel_size": kernel_shape, "stride": strides, "padding": paddings, "ceil_mode": ceil_mode, "exclusive": 'True', } self.paddle_graph.add_layer( paddle_op, inputs={'x': val_x.name}, outputs=layer_outputs, **layer_attrs) @print_mapping_info def Concat(self, node): inputs_list = [] dtypes = set() for i in range(len(node.layer.input)): ipt = self.graph.get_input_node(node, idx=i, copy=True) inputs_list.append(ipt.name) dtypes.add(ipt.dtype) if len(dtypes) > 1: assert 'Unspported situation happened, please create issue on https://github.com/PaddlePaddle/X2Paddle/issues.' axis = node.get_attr('axis') self.paddle_graph.add_layer( 'paddle.concat', inputs={"x": inputs_list}, outputs=[node.name], axis=axis) @print_mapping_info def Flatten(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) output_shape = node.out_shapes[0] axis = node.get_attr('axis', 1) shape_list = [1, 1] if axis == 0: for s in output_shape: shape_list[1] *= s else: for s in output_shape[:axis]: shape_list[0] *= s for s in output_shape[axis:]: shape_list[1] *= s self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": val_x.name}, outputs=[node.name], shape=shape_list) @print_mapping_info def Gemm(self, node): 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) 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.name + '_mm' matmul_inputs = {"x": val_a.name, "y": val_b.name} attr_matmul = { "transpose_x": trans_a, "transpose_y": trans_b, } self.paddle_graph.add_layer( 'paddle.matmul', inputs=matmul_inputs, outputs=[val_mm], **attr_matmul) self.paddle_graph.add_layer( "paddle.scale", inputs={"x": val_mm}, outputs=[val_mm], scale=alpha) if beta != 0: if beta == 1.: add_inputs = {"x": val_mm, "y": val_c.name} self.paddle_graph.add_layer( "paddle.add", inputs=add_inputs, outputs=[node.name]) else: var_beta = node.name + '_beta' self.paddle_graph.add_layer( "paddle.scale", inputs={"x": val_c.name}, outputs=[var_beta], scale=beta) add_inputs = {"x": val_mm, "y": var_beta} self.paddle_graph.add_layer( "paddle.add", inputs=add_inputs, outputs=[node.name]) @print_mapping_info def Sum(self, node): val_inps = node.layer.input inputs_dict = { "x": self.graph.get_input_node( node, idx=0, copy=True).name, "y": self.graph.get_input_node( node, idx=1, copy=True).name, } self.paddle_graph.add_layer("paddle.add", inputs=inputs_dict, outputs=[node.name]) for idx, ipt in enumerate(val_inps[2:]): y = self.graph.get_input_node(node, idx=idx, copy=True) inputs_dict = { "x": node.name, "y": y.name, } self.paddle_graph.add_layer( "paddle.add", inputs=inputs_dict, outputs=[node.name]) @print_mapping_info def MatMul(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) x_shape = val_x.out_shapes[0] y_shape = val_y.out_shapes[0] inputs_dict = {"x": val_x.name, "y": val_y.name} if y_shape[0] == 1 and x_shape[-1] != 1 and x_shape[0] != 1: y_squeeze = val_y.name + '_squeeze' self.paddle_graph.add_layer( "paddle.squeeze", inputs={"x": val_y.name}, outputs=[y_squeeze], axis=[0]) inputs_dict['y'] = y_squeeze self.paddle_graph.add_layer( "paddle.matmul", inputs=inputs_dict, outputs=[node.name]) else: self.paddle_graph.add_layer( "paddle.matmul", inputs=inputs_dict, outputs=[node.name]) @print_mapping_info def BatchNormalization(self, node): op_name = name_generator("batchnorm", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] 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) momentum = node.get_attr('momentum', .9) epsilon = node.get_attr('epsilon', 1e-5) c = val_x.out_shapes[0][1] _rename_or_remove_weight(self.weights, val_scale.name, op_name+'.weight') _rename_or_remove_weight(self.weights, val_b.name, op_name+'.bias') _rename_or_remove_weight(self.weights, val_var.name, op_name+'._variance') _rename_or_remove_weight(self.weights, val_mean.name, op_name+'._mean') # Attribute: spatial is used in BatchNormalization-1,6,7 spatial = bool(node.get_attr('spatial')) layer_attrs = { "num_channels": c, "momentum": momentum, "epsilon": epsilon, "is_test": True, "use_global_stats": False, } self.paddle_graph.add_layer( "paddle.nn.BatchNorm", inputs={"x": val_x.name}, outputs=layer_outputs, **layer_attrs) @print_mapping_info def Transpose(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) s_len = len(val_x.out_shapes[0]) perm_default = list(range(s_len)) perm_default.reverse() perm = node.get_attr('perm', perm_default) self.paddle_graph.add_layer( "paddle.transpose", inputs={"x": val_x.name}, outputs=[node.name], perm=perm) @print_mapping_info def PRelu(self, node): op_name = name_generator("prelu", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] val_x = self.graph.get_input_node(node, idx=0, copy=True) val_slope = self.graph.get_input_node(node, idx=1, copy=True) mode = 'channel' shape_slope = val_slope.out_shapes[0] if shape_slope == [1] * len(shape_slope): mode = 'all' if mode == "element": self.paddle_graph.add_layer( "paddle.zeros", inputs={}, outputs=[output_name + "__zeros"], shape=shape_slope, dtype=string(node.dtype)) self.paddle_graph.add_layer( "paddle.maximum", inputs={"x": val_x.name, "y": output_name + "__zeros"}, outputs=[output_name + "__max"]) self.paddle_graph.add_layer( "paddle.minimum", inputs={"x": val_x.name, "y": output_name + "__zeros"}, outputs=[output_name + "__max"]) self.paddle_graph.add_layer( "paddle.multiply", inputs={"x": val_slope.name, "y": output_name + "__min"}, outputs=[output_name + "__mul"]) self.paddle_graph.add_layer( "paddle.add", inputs={"x": output_name + "__max", "y": output_name + "__mul"}, outputs=[output_name]) else: if mode == 'channel': slope_data = _const_weight_or_none(val_slope) if slope_data is None: self.paddle_graph.add_layer( "paddle.reshape", inputs={"x": val_slope.name}, outputs=[val_slope.name], shape=[shape_slope[0]]) self.paddle_graph.add_layer( "paddle.nn.functional.prelu", inputs={"x": val_x.name, "weight": val_slope.name}, outputs=[node.name]) return _rename_or_remove_weight(self.weights, val_slope.name) if len(shape_slope) > 1: self.weights[op_name+'._weight'] = np.reshape(slope_data, shape_slope[0]) num_parameters = val_x.out_shapes[0][1] else: num_parameters = 1 _rename_or_remove_weight(self.weights, val_slope.name) self.weights[op_name+'._weight'] = np.reshape(self.weights[val_slope.name], [1]) self.paddle_graph.add_layer( "paddle.nn.PReLU", inputs={"x": val_x.name}, outputs=layer_outputs, num_parameters=num_parameters) @print_mapping_info def Squeeze(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) axes = node.get_attr('axes') if len(val_x.out_shapes[0]) == 1: self.paddle_graph.add_layer( "paddle.cast", inputs={"x": val_x.name}, outputs=[node.name], dtype=string(val_x.dtype)) else: self.paddle_graph.add_layer( "paddle.squeeze", inputs={"x": val_x.name}, outputs=[node.name], axis=axes) @print_mapping_info 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) self.paddle_graph.add_layer( "paddle.equal", inputs={'x': val_x.name, 'y': val_y.name}, outputs=[node.name]) @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) self.paddle_graph.add_layer( "paddle.greater_than", inputs={'x': val_x.name, 'y': val_y.name}, outputs=[node.name], param_attr=None) @print_mapping_info 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) not_condition = condition.name + '_not' self.paddle_graph.add_layer( "paddle.logical_not", inputs={"x": condition.name}, outputs=[not_condition]) cast_not_condition = not_condition + '_cast' self.paddle_graph.add_layer( "paddle.cast", inputs={"x": not_condition}, outputs=[cast_not_condition], dtype=string(val_x.dtype)) cast_condition = condition.name + '_cast' self.paddle_graph.add_layer( "paddle.cast", inputs={"x": condition.name}, outputs=[cast_condition], dtype=string(val_x.dtype)) mul_val_x = val_x.name + '_mul' self.paddle_graph.add_layer( "paddle.multiply", inputs={'x': val_x.name, 'y': cast_condition}, outputs=[mul_val_x]) mul_val_y = val_y.name + '_mul' self.paddle_graph.add_layer( "paddle.multiply", inputs={'x': val_y.name, 'y': cast_not_condition}, outputs=[mul_val_y]) self.paddle_graph.add_layer( "paddle.add", inputs={'x': mul_val_x, 'y': mul_val_y}, outputs=[node.name]) @print_mapping_info def NonZero(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) val_x_dim = len(val_x.out_shapes[0]) if val_x_dim == 1: self.paddle_graph.add_layer( "paddle.nonzero", inputs={"x": val_x.name}, outputs=[val_x.name]) self.paddle_graph.add_layer( "paddle.transpose", inputs={"x": val_x.name}, outputs=[node.layer_name], perm=[1, 0]) if val_x_dim > 1: self.paddle_graph.add_layer( "paddle.nonzero", inputs={"x": val_x.name}, outputs=[val_x.name]) self.paddle_graph.add_layer( "paddle.split", inputs={"x": val_x.name}, outputs=[val_x.name], num_or_sections=1, axis=val_x_dim) self.paddle_graph.add_layer( "paddle.concat", inputs={"x": val_x.name}, outputs=[node.name]) @print_mapping_info def Identity(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) self.paddle_graph.add_layer( "paddle.assign", inputs={"x": val_x.name}, outputs=[node.name]) @print_mapping_info 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) if repeats is None: repeats = val_repeats.name if val_repeats.dtype != 'int32': self.paddle_graph.add_layer( "paddle.cast", inputs={"x": repeats}, outputs=["{}.tmp".format(repeats)], dtype=string("int32")) repeats = "{}.tmp".format(repeats) elif isinstance(repeats, int): repeats = [repeats] attr = { 'expand_times': repeats, "name": string(node.name), } self.paddle_graph.add_layer( "paddle.tile", inputs={"x": val_x.name}, outputs=[node.name], repeat_times=repeats) @print_mapping_info def MaxPool(self, node): op_name = name_generator("pool", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] val_x = self.graph.get_input_node(node, idx=0, copy=True) auto_pad = node.get_attr('auto_pad', 'NOTSET') 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 paddle_op = 'paddle.nn.MaxPool{}D'.format(poolnd) assert 1 <= poolnd <= 3, 'only Pool1D, Pool2D and Pool3D are supported' paddings, val_x = self._pad_if_asymmetric(node, pads, val_x) if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER": input_shape = val_x.out_shapes[0] 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 layer_attrs = { "kernel_size": kernel_shape, "stride": strides, "padding": paddings, "ceil_mode": ceil_mode, } self.paddle_graph.add_layer( paddle_op, inputs={'x': val_x if isinstance(val_x, str) else val_x.name}, outputs=layer_outputs, **layer_attrs) @print_mapping_info def GlobalMaxPool(self, node): op_name = name_generator("pool", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] val_x = self.graph.get_input_node(node, idx=0, copy=True) input_shape = val_x.out_shapes[0] if len(input_shape) == 4: poolnd = 2 elif len(input_shape) == 5: poolnd = 3 elif len(input_shape) == 3: poolnd = 1 paddle_op = 'paddle.nn.AdaptiveMaxPool{}D'.format(poolnd) assert 1 <= poolnd <= 3, 'only Pool1D, Pool2D and Pool3D are supported' output_shape = node.out_shapes[0] self.paddle_graph.add_layer( paddle_op, inputs={'x': val_x.name}, outputs=layer_outputs, output_size=output_shape[2:]) @print_mapping_info def GlobalAveragePool(self, node): op_name = name_generator("pool", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] val_x = self.graph.get_input_node(node, idx=0, copy=True) input_shape = val_x.out_shapes[0] if len(input_shape) == 4: poolnd = 2 elif len(input_shape) == 5: poolnd = 3 elif len(input_shape) == 3: poolnd = 1 paddle_op = 'paddle.nn.AdaptiveAvgPool{}D'.format(poolnd) assert 1 <= poolnd <= 3, 'only Pool1D, Pool2D and Pool3D are supported' output_shape = node.out_shapes[0] self.paddle_graph.add_layer( paddle_op, inputs={'x': val_x.name}, outputs=layer_outputs, output_size=output_shape[2:]) @print_mapping_info def Conv(self, node): op_name = name_generator("conv", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] val_x = self.graph.get_input_node(node, idx=0, copy=True) val_w = self.graph.get_input_node(node, idx=1, copy=True) has_bias = len(node.layer.input) == 3 if has_bias: val_b = self.graph.get_input_node(node, idx=2, copy=True) auto_pad = node.get_attr('auto_pad', 'NOTSET') kernel_shape = node.get_attr('kernel_shape') convnd = len(kernel_shape) assert 2 <= convnd <= 3, 'only Conv2D and Conv3D is supported' num_out_channels = val_w.out_shapes[0][0] num_in_channels = val_w.out_shapes[0][1] paddle_op = 'paddle.nn.Conv{}D'.format(convnd) num_groups = node.get_attr('group', 1) strides = node.get_attr('strides', [1] * convnd) dilations = node.get_attr('dilations', [1] * convnd) pads = node.get_attr('pads', [0] * (convnd * 2)) input_shape = val_x.out_shapes[0] paddings, val_x = self._pad_if_asymmetric(node, pads, val_x) if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER": 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 layer_inputs = {'x': val_x if isinstance(val_x, str) else val_x.name} layer_attrs = { "in_channels": num_in_channels * num_groups, "out_channels": num_out_channels, "kernel_size": kernel_shape, "stride": strides, "padding": paddings, "dilation": dilations, "groups": num_groups, } remove_weight = True if val_w.name in self.done_weight_list else False if remove_weight: self.done_weight_list.append(val_w.name) _rename_or_remove_weight(self.weights, val_w.name, op_name+'.weight', remove_weight) if has_bias: remove_bias = True if val_b.name in self.done_weight_list else False if remove_bias: self.done_weight_list.append(val_b_name) _rename_or_remove_weight(self.weights, val_b.name, op_name+'.bias', remove_bias) else: layer_attrs["bias_attr"] = False if reduce(lambda x,y:x*y, input_shape) in [1, -1] and 1 not in input_shape: input_shape[1] = num_in_channels * num_groups input_shape[0] = 0 input_shape[2] = 0 self.paddle_graph.add_layer( "paddle.reshape", inputs=layer_inputs, outputs=[layer_inputs["x"]], shape=input_shape) self.paddle_graph.add_layer( paddle_op, inputs=layer_inputs, outputs=layer_outputs, **layer_attrs) @print_mapping_info def ConvTranspose(self, node): op_name = name_generator("conv_trans", self.nn_name2id) output_name = node.name layer_outputs = [op_name, output_name] 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_b = None if len(node.layer.input) > 2: val_b = self.graph.get_input_node(node, idx=2, copy=True) auto_pad = node.get_attr('auto_pad', 'NOTSET') out_padding = node.get_attr('output_padding', [0, 0]) kernel_shape = node.get_attr('kernel_shape') assert kernel_shape, 'kernel_shape not inferred' convnd = len(kernel_shape) assert 2 <= convnd <= 3, 'only Conv2DTranspose and Conv3DTranspose supported' num_in_channels = val_w.out_shapes[0][0] num_out_channels = val_w.out_shapes[0][1] paddle_op = 'paddle.nn.Conv{}DTranspose'.format(convnd) 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)) paddings, var_x = self._pad_if_asymmetric(node, pads, val_x) output_size = [0, 0] output_size[0] = (val_x.out_shapes[0][2] - 1 ) * strides[0] - 2 * paddings[0] + dilations[0] * ( kernel_shape[0] - 1) + 1 + out_padding[0] output_size[1] = (val_x.out_shapes[0][3] - 1 ) * strides[1] - 2 * paddings[1] + dilations[1] * ( kernel_shape[1] - 1) + 1 + out_padding[1] # Conv2DTranspose缺少output_size，只能在forward里头传进output_size inputs_dict = {'x': val_x if isinstance(val_x, str) else val_x.name} layer_attrs = { "in_channels": num_in_channels, "out_channels": num_out_channels, "kernel_size": kernel_shape, "stride": strides, "dilation": dilations, "padding": paddings, "groups": num_groups, "output_padding":out_padding} _rename_or_remove_weight(self.weights, val_w.name, op_name+'.weight',) if val_b is not None: _rename_or_remove_weight(self.weights, val_b.name, op_name+'.bias') self.paddle_graph.add_layer( kernel=paddle_op, inputs=inputs_dict, outputs=layer_outputs, **layer_attrs) @print_mapping_info def ArgMax(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) axis = node.get_attr('axis') keepdims = False if node.get_attr('keepdims') == 0 else True layer_attrs = {'axis': axis, 'keepdim': keepdims} self.paddle_graph.add_layer( 'paddle.argmax', inputs={"x": val_x.name}, outputs=[node.name], **layer_attrs) @print_mapping_info def Size(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) self.paddle_graph.add_layer( "paddle.shape", inputs={"input": val_x.name}, outputs=[node.name]) self.paddle_graph.add_layer( 'paddle.cast', inputs={"x": node.name}, outputs=[node.name], dtype=string('int64')) self.paddle_graph.add_layer( "paddle.prod", inputs={"x": node.name}, outputs=[node.name]) @print_mapping_info def Sign(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) if node.dtype not in ["float16", "float32", "float64"]: self.paddle_graph.add_layer( "paddle.cast", inputs={"x": val_x.name}, outputs=[val_x.name], dtype=string("float32")) self.paddle_graph.add_layer( "paddle.sign", inputs={"x": val_x.name}, outputs=[node.name]) if node.dtype not in ["float16", "float32", "float64"]: self.paddle_graph.add_layer( "paddle.cast", inputs={"x": node.name}, outputs=[node.name], dtype=string(node.dtype)) @print_mapping_info def OneHot(self, node): nn_op_name = name_generator("onehot", self.nn_name2id) output_name = node.name layer_outputs = [nn_op_name, output_name] indices = self.graph.get_input_node(node, idx=0, copy=True) depth = self.graph.get_input_node(node, idx=1, copy=True) values = self.graph.get_input_node(node, idx=2, copy=True) axis = node.get_attr('axis', -1) self.paddle_graph.add_layer( "custom_layer:OneHot", inputs={"indices": indices.name, "depth": depth.name, "values": values.name}, outputs=layer_outputs, axis=axis) @print_mapping_info def Reciprocal(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) self.paddle_graph.add_layer( "paddle.reciprocal", inputs={"x": val_x.name}, outputs=[node.name]) @print_mapping_info def LSTM(self, node): x = self.graph.get_input_node(node, idx=0, copy=True) input_weight = self.graph.get_input_node(node, idx=1, copy=True) hidden_weight = self.graph.get_input_node(node, idx=2, copy=True) input_nums = len(node.layer.input) exist_input_nums = 3 have_bias = False if input_nums > 3 and node.layer.input[3] != '': bias = self.graph.get_input_node(node, idx=exist_input_nums, copy=True) have_bias = True exist_input_nums += 1 if input_nums > 4 and node.layer.input[4] != '': sequence_lens = self.graph.get_input_node(node, idx=exist_input_nums, copy=True) exist_input_nums += 1 if input_nums > 5 and node.layer.input[5] != '': init_h = self.graph.get_input_node(node, idx=exist_input_nums, copy=True) self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": init_h.name}, outputs=[init_h.name], shape=init_h.out_shapes[0] ) exist_input_nums += 1 if input_nums > 6 and node.layer.input[6] != '': init_c = self.graph.get_input_node(node, idx=exist_input_nums, copy=True) self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": init_c.name}, outputs=[init_c.name], shape=init_c.out_shapes[0] ) input_weight_np = _const_weight_or_none(input_weight) _rename_or_remove_weight(self.weights, input_weight.name) hidden_size = node.get_attr('hidden_size', input_weight_np.shape[1]/4) input_size = input_weight_np.shape[2] hidden_weight_np = _const_weight_or_none(hidden_weight) _rename_or_remove_weight(self.weights, hidden_weight.name) bias_np = _const_weight_or_none(bias) _rename_or_remove_weight(self.weights, bias.name) input_bias_np = bias_np[:, :4*hidden_size] hidden_bias_np = bias_np[:, 4*hidden_size:] # parameters order in paddle:lstm: # 1. gate order in paddle is: input, forget, cell, output. # 2. gate orfer in onnx is: input, output, forget, cell. def reform_weights(w, n, intervals): slices = [w[:,x * n: y * n] for x, y in intervals] return np.concatenate(slices, axis=1) def transform_weight_with_bias(weights, n, intervals): return [reform_weights(w, n, intervals) for w in weights] reform_permutation = [(0, 1), (2, 4), (1, 2)] weights = transform_weight_with_bias( [input_weight_np, hidden_weight_np, input_bias_np, hidden_bias_np], hidden_size, reform_permutation) op_name = name_generator("lstm", self.nn_name2id) y_out = node.output(0) yh_out = node.output(1) yc_out = node.output(2) direction = node.get_attr('direction', 'forward') def generate_paddle_param_names(op_name, suffix=''): param_names = [] param_names.extend(['{}.weight_ih_l0{}', '{}.weight_hh_l0{}']) if have_bias != False: param_names.append('{}.bias_ih_l0{}') if have_bias != False: param_names.append('{}.bias_hh_l0{}') param_names = [x.format(op_name, suffix) for x in param_names] return param_names def assign_params(op_name, weights, weight_idx=0, suffix=''): param_names = generate_paddle_param_names(op_name, suffix) print(param_names) for param_name, weight in zip(param_names, weights): self.weights[param_name] = weight[weight_idx] if direction == 'backward': raise Exception("LSTM support 'forward' or 'bidirectional', except '{}'.".format(direction)) else: assign_params(op_name, weights) if direction == 'bidirectional': assign_params(op_name, weights, 1, '_reverse') self.paddle_graph.add_layer( 'paddle.nn.LSTM', inputs={'input': x.name, 'initial_states': (init_h.name, init_c.name)}, outputs=[op_name, y_out, yh_out, yc_out], input_size=input_size, hidden_size=hidden_size, num_layers=1, direction=string(direction), time_major=True) self.paddle_graph.add_layer( 'paddle.reshape', inputs={"x": y_out}, outputs=[y_out], shape=[0, 0, -1, hidden_size] ) self.paddle_graph.add_layer( 'paddle.transpose', inputs={"x": y_out}, outputs=[y_out], perm=[0,2,1,3] ) @print_mapping_info def TopK(self, node): val_x = self.graph.get_input_node(node, idx=0, copy=True) val_k = self.graph.get_input_node(node, idx=1, copy=True) layer_attrs = dict() layer_attrs["axis"] = node.get_attr('axis', -1) layer_attrs["largest"] = True if node.get_attr('largest', 1) == 1 else False layer_attrs["sorted"] = True if node.get_attr('sorted', 1) == 1 else False self.paddle_graph.add_layer( "paddle.topk", inputs={"x": val_x.name, "k": val_k.name}, outputs=["{}_p{}".format(node.layer_name, 0), "{}_p{}".format(node.layer_name, 1)], **layer_attrs)