# Copyright (c) 2020 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. import numpy as np from x2paddle.optimizer.pattern_matcher import FuseBase from x2paddle.core.program import PaddleGraph, PaddleLayer from x2paddle.core.util import * class TraceFcFuser(FuseBase): def __init__(self): self.linear_index = 0 super(TraceFcFuser, self).__init__() self.patterns = list() def build_pattern(self): """ 描述需要替换的fc图结构。 fc层模式python实现代码示例: 模式一: encoder_layer_8_attention_self_key_weight = self.encoder_layer_8_attention_self_key_weight x748 = paddle.transpose(x=encoder_layer_8_attention_self_key_weight, perm=[1, 0]) x749 = paddle.matmul(x=x732, y=x748) encoder_layer_8_attention_self_key_bias = self.encoder_layer_8_attention_self_key_bias x750 = x749 + 1 * encoder_layer_8_attention_self_key_bias 模式二: x13 = self.x13 x14 = paddle.transpose(x=x13, perm=[1, 0]) x15 = self.x15 x16 = paddle.addmm(input=x15, x=x12, y=x14, beta=1, alpha=1) """ def gen_name(id): return "x" + str(id) pattern = PaddleGraph() pattern.add_layer( "self.create_parameter", inputs={}, outputs=[gen_name(0)]) pattern.add_layer( "paddle.transpose", inputs={"x": gen_name(0)}, outputs=[gen_name(1)], perm=[1, 0]) pattern.add_layer( "paddle.matmul", inputs={"x": "fc-input-0", "y": gen_name(1)}, outputs=[gen_name(2)]) pattern.add_layer( "self.create_parameter", inputs={}, outputs=[gen_name(3)]) pattern.add_layer( "prim.add_", inputs={"x": gen_name(2), "y": gen_name(3)}, outputs=[gen_name(4)], alpha=1) pattern.build(inputs={"input-0": "fc-input-0"}) self.patterns.append(pattern) pattern = PaddleGraph() pattern.add_layer( "self.create_parameter", inputs={}, outputs=[gen_name(0)]) pattern.add_layer( "paddle.transpose", inputs={"x": gen_name(0)}, outputs=[gen_name(1)], perm=[1, 0]) pattern.add_layer( "self.create_parameter", inputs={}, outputs=[gen_name(2)]) pattern.add_layer( "paddle.addmm", inputs={"input": gen_name(2), "x": "fc-input-0", "y": gen_name(1)}, outputs=[gen_name(4)], alpha=1, beta=1) pattern.build(inputs={"input-0": "fc-input-0"}) self.patterns.append(pattern) def insert_new_layer(self, graph, parameters, matches): new_layer = self.gen_new_layer(parameters, matches) new_layer_id = list(matches.keys())[0] graph.layers[new_layer_id] = new_layer matches.pop(new_layer_id) def gen_new_layer(self, parameters, matches): layers_id = list(matches.keys()) if len(layers_id) == 5: layer = matches[layers_id[2]] else: layer = matches[layers_id[-1]] input_name = layer.inputs["x"] scope_name = layer.scope_name layer = matches[layers_id[-1]] output_name = layer.outputs[0] layer = matches[layers_id[0]] weight_name = layer.outputs[0] layer = matches[layers_id[-2]] bias_name = layer.outputs[0] attrs = dict() attrs["in_features"] = parameters[weight_name].shape[1] attrs["out_features"] = parameters[weight_name].shape[0] linear_name = "linear{}".format(self.linear_index) self.linear_index += 1 weight_numpy = parameters[weight_name] parameters["{}.weight".format(linear_name)] = weight_numpy.transpose( (1, 0)) self.rm_params.add(weight_name) bias_numpy = parameters[bias_name] if len(bias_numpy.shape) == 2: bias_numpy = np.squeeze(bias_numpy) parameters["{}.bias".format(linear_name)] = bias_numpy self.rm_params.add(bias_name) new_layer = PaddleLayer( layers_id[0], "paddle.nn.Linear", inputs={"input": input_name}, outputs=[linear_name, output_name], scope_name=scope_name, **attrs) return new_layer