Skip to content

X
X2Paddle

2301_76417823 / X2Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / X2Paddle

通知 1
Star 0
Fork 0
X
X2Paddle

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

未验证 提交 f08b1f3f 编写于 作者: J Jason 提交者： GitHub
浏览文件
操作

Merge pull request #472 from SunAhong1993/lstm 

Lstm
上级 a94afe1a 95519af0
隐藏空白更改
内联 并排
Showing with 543 addition and 340 deletion
README.md
浏览文件 @ f08b1f3f
...@@ -10,7 +10,7 @@ X2Paddle在多个主流的CV模型上，测试过TensorFlow/Caffe/ONNX/PyTorch ...@@ -10,7 +10,7 @@ X2Paddle在多个主流的CV模型上，测试过TensorFlow/Caffe/ONNX/PyTorch
## 环境依赖 ## 环境依赖
python == 2.7 | python >= 3.5 python == 2.7 | python >= 3.5
paddlepaddle 2.0-rc 或者 develop paddlepaddle 2.0.0-rc1 或者 develop
**按需安装以下依赖** **按需安装以下依赖**
tensorflow ： tensorflow == 1.14.0 tensorflow ： tensorflow == 1.14.0
...@@ -93,12 +93,6 @@ X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README ...@@ -93,12 +93,6 @@ X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README
6. [X2Paddle添加内置的Caffe自定义层](./docs/user_guides/add_caffe_custom_layer.md) 6. [X2Paddle添加内置的Caffe自定义层](./docs/user_guides/add_caffe_custom_layer.md)
## 更新历史 ## 更新历史
2019.08.05
1. 统一tensorflow/caffe/onnx模型转换代码和对外接口
2. 解决上一版caffe2fluid无法转换多分支模型的问题
3. 解决Windows上保存模型无法加载的问题
4. 新增optimizer，优化代码结构，合并conv、batch_norm的bias和激活函数
2020.12.09 2020.12.09
1. 新增PyTorch2Paddle转换方式，转换得到Paddle动态图代码，并动转静获得inference_model。 1. 新增PyTorch2Paddle转换方式，转换得到Paddle动态图代码，并动转静获得inference_model。
方式一：trace方式，转换后的代码有模块划分，每个模块的功能与PyTorch相同。 方式一：trace方式，转换后的代码有模块划分，每个模块的功能与PyTorch相同。
...@@ -107,8 +101,6 @@ X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README ...@@ -107,8 +101,6 @@ X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README
3. 新增TensorFlow op（14个）：Neg、Greater、FloorMod、LogicalAdd、Prd、Equal、Conv3D、Ceil、AddN、DivNoNan、Where、MirrorPad、Size、TopKv2 3. 新增TensorFlow op（14个）：Neg、Greater、FloorMod、LogicalAdd、Prd、Equal、Conv3D、Ceil、AddN、DivNoNan、Where、MirrorPad、Size、TopKv2
4. 新增Optimizer模块，主要包括op融合、op消除功能，转换后的代码可读性更强，进行预测时耗时更短。 4. 新增Optimizer模块，主要包括op融合、op消除功能，转换后的代码可读性更强，进行预测时耗时更短。
**如果你需要之前版本的tensorflow2fluid/caffe2fluid/onnx2fluid，可以继续访问release-0.9分支，获取之前版本的代码使用。**
## Acknowledgements ## Acknowledgements
......
docs/introduction/op_list.md
浏览文件 @ f08b1f3f
...@@ -61,7 +61,7 @@ ...@@ -61,7 +61,7 @@
| 41 | MatMul | 42 | Sum | 43 | Transpose | 44 | BatchNormalization | | 41 | MatMul | 42 | Sum | 43 | Transpose | 44 | BatchNormalization |
| 45 | Squeeze | 46 | Equal | 47 | Identity | 48 | GlobalAveragePool | | 45 | Squeeze | 46 | Equal | 47 | Identity | 48 | GlobalAveragePool |
| 49 | MaxPool | 50 | Conv | 51 | Gemm | 52 | NonZero | | 49 | MaxPool | 50 | Conv | 51 | Gemm | 52 | NonZero |
| 53 | Abs | 54 | Floor | | 53 | Abs | 54 | Floor | 52 | ArgMax |
## PyTorch ## PyTorch
Aten: Aten:
...@@ -93,7 +93,8 @@ Aten: ...@@ -93,7 +93,8 @@ Aten:
| 93 | aten::sub | 94 | aten::t |95|aten::tanh|96|aten::split| | 93 | aten::sub | 94 | aten::t |95|aten::tanh|96|aten::split|
| 97 | aten::transpose | 98 | aten::to |99|aten::type\_as|100|aten::unsqueeze| | 97 | aten::transpose | 98 | aten::to |99|aten::type\_as|100|aten::unsqueeze|
| 101 | aten::upsample\_bilinear2d | 102 | aten::values |103|aten::view|104|aten::warn| | 101 | aten::upsample\_bilinear2d | 102 | aten::values |103|aten::view|104|aten::warn|
| 105 | aten::where | 106 | aten::zeros |107|aten::zeros\_like||| | 105 | aten::where | 106 | aten::zeros |107|aten::zeros\_like|108|aten::bmm|
| 109 | aten::sub\_ | 110 | aten:erf |111|aten::lstm|112|aten::gather|
Prim: Prim:
| 序号 | OP | 序号 | OP | 序号 | OP | 序号 | OP | | 序号 | OP | 序号 | OP | 序号 | OP | 序号 | OP |
......
docs/introduction/x2paddle_model_zoo.md
浏览文件 @ f08b1f3f
...@@ -5,28 +5,28 @@ ...@@ -5,28 +5,28 @@
## TensorFlow ## TensorFlow
| 模型 | 代码 | 备注 | | 模型 | 代码 |
|------|----------|------| |------|----------|
| SqueezeNet | [code](https://github.com/tensorflow/tpu/blob/master/models/official/squeezenet/squeezenet_model.py)|-| | SqueezeNet | [code](https://github.com/tensorflow/tpu/blob/master/models/official/squeezenet/squeezenet_model.py)|
| MobileNet_V1 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-| | MobileNet_V1 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |
| MobileNet_V2 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-| | MobileNet_V2 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |
| ShuffleNet | [code](https://github.com/TropComplique/shufflenet-v2-tensorflow) |-| | ShuffleNet | [code](https://github.com/TropComplique/shufflenet-v2-tensorflow) |
| mNASNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet) |-| | mNASNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet) |
| EfficientNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet) |-| | EfficientNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet) |
| Inception_V3 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v3.py) |-| | Inception_V3 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v3.py) |
| Inception_V4 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py) |-| | Inception_V4 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py) |
| Inception_ResNet_V2 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py) |-| | Inception_ResNet_V2 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py) |
| VGG16 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-| | VGG16 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |
| ResNet_V1_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-| | ResNet_V1_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |
| ResNet_V2_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-| | ResNet_V2_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |
| UNet | [code1](https://github.com/jakeret/tf_unet )/[code2](https://github.com/lyatdawn/Unet-Tensorflow) |-| | UNet | [code1](https://github.com/jakeret/tf_unet )/[code2](https://github.com/lyatdawn/Unet-Tensorflow) |
| MTCNN | [code](https://github.com/AITTSMD/MTCNN-Tensorflow) |-| | MTCNN | [code](https://github.com/AITTSMD/MTCNN-Tensorflow) |
| YOLO-V3| [code](https://github.com/YunYang1994/tensorflow-yolov3) | -| | YOLO-V3| [code](https://github.com/YunYang1994/tensorflow-yolov3) |
| FALSR | [code](https://github.com/xiaomi-automl/FALSR) | 需使用参数without_data_format_optimization | | FALSR | [code](https://github.com/xiaomi-automl/FALSR) |
| DCSCN | [code](https://modelzoo.co/model/dcscn-super-resolution) | 需使用参数without_data_format_optimization | | DCSCN | [code](https://modelzoo.co/model/dcscn-super-resolution) |
| Bert（albert） | [code](https://github.com/google-research/albert#pre-trained-models) | 需使用参数without_data_format_optimization | | Bert（albert） | [code](https://github.com/google-research/albert#pre-trained-models) |
| Bert（chinese_L-12_H-768_A-12） | [code](https://github.com/google-research/bert#pre-trained-models) | 需使用参数without_data_format_optimization | | Bert（chinese_L-12_H-768_A-12） | [code](https://github.com/google-research/bert#pre-trained-models) |
| Bert（multi_cased_L-12_H-768_A-12） | [code](https://github.com/google-research/bert#pre-trained-models) | 需使用参数without_data_format_optimization | | Bert（multi_cased_L-12_H-768_A-12） | [code](https://github.com/google-research/bert#pre-trained-models) |
## Caffe ## Caffe
...@@ -72,8 +72,8 @@ ...@@ -72,8 +72,8 @@
| EfficientNet | [pytorch(personal practice)](https://github.com/rwightman/gen-efficientnet-pytorch) |9| | EfficientNet | [pytorch(personal practice)](https://github.com/rwightman/gen-efficientnet-pytorch) |9|
| SqueezeNet | [onnx official](https://s3.amazonaws.com/download.onnx/models/opset_9/squeezenet.tar.gz) |9| | SqueezeNet | [onnx official](https://s3.amazonaws.com/download.onnx/models/opset_9/squeezenet.tar.gz) |9|
|Ultra-Light-Fast-Generic-Face-Detector-1MB| [onnx_model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/onnx)|9 | |Ultra-Light-Fast-Generic-Face-Detector-1MB| [onnx_model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/onnx)|9 |
|BERT| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](FAQ.md)| |BERT| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](../user_guides/FAQ.md)|
|GPT2| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](FAQ.md)| |GPT2| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](../user_guides/FAQ.md)|
## PyTorch ## PyTorch
...@@ -96,3 +96,6 @@ ...@@ -96,3 +96,6 @@
| FlaubertModel | [code](https://huggingface.co/transformers/model_doc/flaubert.html) |只支持trace模式| | FlaubertModel | [code](https://huggingface.co/transformers/model_doc/flaubert.html) |只支持trace模式|
| Roberta| [code](https://huggingface.co/transformers/model_doc/roberta.html) |只支持trace模式| | Roberta| [code](https://huggingface.co/transformers/model_doc/roberta.html) |只支持trace模式|
| XLMRobertaForTokenClassification|[code](https://huggingface.co/transformers/model_doc/xlmroberta.html) |只支持trace模式| | XLMRobertaForTokenClassification|[code](https://huggingface.co/transformers/model_doc/xlmroberta.html) |只支持trace模式|
| EasyOCR_detector|[code](https://github.com/JaidedAI/EasyOCR/blob/master/easyocr/detection.py) |-|
| EasyOCR_recognizer|[code](https://github.com/JaidedAI/EasyOCR/blob/master/easyocr/recognition.py) |-|
x2paddle/core/program.py
浏览文件 @ f08b1f3f
...@@ -26,6 +26,7 @@ import six ...@@ -26,6 +26,7 @@ import six
import pickle import pickle
import numpy as np import numpy as np
from os import path as osp from os import path as osp
from x2paddle.core.util import *
class PaddleLayer(object): class PaddleLayer(object):
...@@ -210,6 +211,8 @@ class PaddleGraph(object): ...@@ -210,6 +211,8 @@ class PaddleGraph(object):
layer_id, 0) == 0 and layer.kernel != "prim.assert" \ layer_id, 0) == 0 and layer.kernel != "prim.assert" \
and layer.kernel != "prim.exception" \ and layer.kernel != "prim.exception" \
and layer.kernel != "prim.warnings": and layer.kernel != "prim.warnings":
if layer.kernel == "paddle.to_tensor":
self.inputs_info.pop(layer.outputs[0])
invalid_list.append(layer_id) invalid_list.append(layer_id)
for layer_id in invalid_list: for layer_id in invalid_list:
self.layers.pop(layer_id) self.layers.pop(layer_id)
...@@ -272,7 +275,7 @@ class PaddleGraph(object): ...@@ -272,7 +275,7 @@ class PaddleGraph(object):
def gen_dygraph_model(self, save_dir, jit_type=None): def gen_dygraph_model(self, save_dir, jit_type=None):
if jit_type == "trace": if jit_type == "trace":
from x2paddle.optimizer.code_optimizer import HierarchicalTree from x2paddle.optimizer.pytorch_code_optimizer import HierarchicalTree
hierarchical_tree = HierarchicalTree(self) hierarchical_tree = HierarchicalTree(self)
for layer_id, layer in self.layers.items(): for layer_id, layer in self.layers.items():
hierarchical_tree.insert(layer) hierarchical_tree.insert(layer)
...@@ -280,7 +283,7 @@ class PaddleGraph(object): ...@@ -280,7 +283,7 @@ class PaddleGraph(object):
self.dump_dygraph_parameter(save_dir) self.dump_dygraph_parameter(save_dir)
else: else:
if self.source_type == "pytorch": if self.source_type == "pytorch":
from x2paddle.optimizer.code_optimizer import ModuleGraph from x2paddle.optimizer.pytorch_code_optimizer import ModuleGraph
module_graph = ModuleGraph(self) module_graph = ModuleGraph(self)
module_graph.save_source_files(save_dir) module_graph.save_source_files(save_dir)
self.dump_dygraph_parameter(save_dir) self.dump_dygraph_parameter(save_dir)
...@@ -324,12 +327,10 @@ class PaddleGraph(object): ...@@ -324,12 +327,10 @@ class PaddleGraph(object):
write_code( write_code(
f, [ f, [
"from paddle.fluid.initializer import Constant",
"from paddle.fluid.param_attr import ParamAttr",
"import paddle.fluid as fluid",
custom_import, custom_import,
"import paddle", "import math", "", "import paddle",
"import math",
"",
], ],
indent=0) indent=0)
if self.custom_code is not None: if self.custom_code is not None:
...@@ -346,6 +347,8 @@ class PaddleGraph(object): ...@@ -346,6 +347,8 @@ class PaddleGraph(object):
], ],
indent=1) indent=1)
for layer_id, layer in self.layers.items(): for layer_id, layer in self.layers.items():
if layer.kernel.startswith("paddle"):
remove_default_attrs(layer.kernel, layer.attrs)
edges_in = self.edges_in.get(layer_id, []) edges_in = self.edges_in.get(layer_id, [])
edges_out = self.edges_out.get(layer_id, []) edges_out = self.edges_out.get(layer_id, [])
if len(edges_in) == 0 and len(edges_out) == 0: if len(edges_in) == 0 and len(edges_out) == 0:
...@@ -425,8 +428,7 @@ class PaddleGraph(object): ...@@ -425,8 +428,7 @@ class PaddleGraph(object):
continue continue
if layer.kernel == "paddle.to_tensor": if layer.kernel == "paddle.to_tensor":
data = layer.attrs["data"] data = layer.attrs["data"]
if not data.startswith("params["): self.inputs.append(data)
self.inputs.append(data)
if len(layer.blocks) > 0: if len(layer.blocks) > 0:
for block in layer.blocks: for block in layer.blocks:
block.get_dygraph_inputs() block.get_dygraph_inputs()
...@@ -473,10 +475,7 @@ class PaddleGraph(object): ...@@ -473,10 +475,7 @@ class PaddleGraph(object):
custom_import = "" custom_import = ""
self.head = gen_codes( self.head = gen_codes(
[ [
"from paddle.fluid.initializer import Constant",
"from paddle.fluid.param_attr import ParamAttr",
"import paddle", "import paddle",
"import paddle.fluid as fluid",
"import math", "import math",
custom_import, custom_import,
"", "",
...@@ -548,6 +547,8 @@ class PaddleGraph(object): ...@@ -548,6 +547,8 @@ class PaddleGraph(object):
gen_head() gen_head()
for layer_id, layer in self.layers.items(): for layer_id, layer in self.layers.items():
if layer.kernel.startswith("paddle"):
remove_default_attrs(layer.kernel, layer.attrs)
if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel
) or layer.kernel == "paddle.to_tensor" or \ ) or layer.kernel == "paddle.to_tensor" or \
layer.kernel.startswith("custom_layer") or \ layer.kernel.startswith("custom_layer") or \
...@@ -578,7 +579,10 @@ class PaddleGraph(object): ...@@ -578,7 +579,10 @@ class PaddleGraph(object):
elif len(layer.outputs) == 2: elif len(layer.outputs) == 2:
line = layer.outputs[1] line = layer.outputs[1]
else: else:
line = ','.join(layer.outputs[1:]) if layer.kernel == "paddle.nn.LSTM":
line = "{}, ({})".format(layer.outputs[1], ', '.join(layer.outputs[-2:]))
else:
line = ','.join(layer.outputs[1:])
if layer.kernel == "paddle.to_tensor" and layer.attrs[ if layer.kernel == "paddle.to_tensor" and layer.attrs[
"data"].startswith("params["): "data"].startswith("params["):
line += " = self.{}".format(layer.outputs[0]) line += " = self.{}".format(layer.outputs[0])
......
x2paddle/core/util.py
浏览文件 @ f08b1f3f
# -*- coding:UTF-8 -*-
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. # Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# #
# Licensed under the Apache License, Version 2.0 (the "License" # Licensed under the Apache License, Version 2.0 (the "License"
...@@ -14,15 +15,61 @@ ...@@ -14,15 +15,61 @@
import numpy import numpy
import math import math
import os import os
import inspect
def string(param): def string(param):
""" 生成字符串。
"""
return "\'{}\'".format(param) return "\'{}\'".format(param)
def name_generator(nn_name, nn_name2id): def name_generator(nn_name, nn_name2id):
""" 生成paddle.nn类op的名字。
Args:
nn_name (str): 名字。
nn_name2id (dict): key为名字，value为名字出现的次数-1。
"""
if nn_name in nn_name2id: if nn_name in nn_name2id:
nn_name2id[nn_name] += 1 nn_name2id[nn_name] += 1
else: else:
nn_name2id[nn_name] = 0 nn_name2id[nn_name] = 0
real_nn_name = nn_name + str(nn_name2id[nn_name]) real_nn_name = nn_name + str(nn_name2id[nn_name])
return real_nn_name return real_nn_name
\ No newline at end of file
def remove_default_attrs(kernel, attrs):
""" 删除每个OP的默认参数。
Args:
kernel (str): OP的类型名字。
attrs (dict): 目前该OP所包含的参数， key为参数名，value为参数值。
"""
def get_default_args(func):
signature = inspect.signature(func)
return {
k: v.default
for k, v in signature.parameters.items()
if v.default is not inspect.Parameter.empty
}
is_func = True
if "paddle.nn" in kernel and "functional"not in kernel:
is_func = False
import paddle
obj = paddle
for i, part in enumerate(kernel.split(".")):
if i == 0:
continue
obj = getattr(obj, part)
if is_func:
func = obj
else:
func = obj.__init__
default_attrs = get_default_args(func)
for default_k, default_v in default_attrs.items():
if default_k in attrs:
if (isinstance(attrs[default_k], list) or isinstance(attrs[default_k], tuple)) \
and not is_func:
if len(set(attrs[default_k])) == 1:
attrs[default_k] = attrs[default_k][0]
if default_v == attrs[default_k]:
attrs.pop(default_k)
\ No newline at end of file
x2paddle/decoder/onnx_decoder.py
浏览文件 @ f08b1f3f
...@@ -571,4 +571,4 @@ class ONNXDecoder(object): ...@@ -571,4 +571,4 @@ class ONNXDecoder(object):
node.input[i] = self.make_variable_name(node.input[i]) node.input[i] = self.make_variable_name(node.input[i])
for i in range(len(node.output)): for i in range(len(node.output)):
node.output[i] = self.make_variable_name(node.output[i]) node.output[i] = self.make_variable_name(node.output[i])
return model return model
\ No newline at end of file
x2paddle/op_mapper/dygraph/caffe2paddle/caffe_op_mapper.py
浏览文件 @ f08b1f3f
...@@ -367,57 +367,46 @@ class CaffeOpMapper(OpMapper): ...@@ -367,57 +367,46 @@ class CaffeOpMapper(OpMapper):
output_size=kernel) output_size=kernel)
else: else:
layer_attrs = { layer_attrs = {
'pool_size': kernel, 'kernel_size': kernel,
'pool_stride': stride, 'stride': stride,
'pool_padding': pad, 'padding': pad,
'ceil_mode': ceil_mode, 'ceil_mode': ceil_mode,
'pool_type': string(pool_type),
'exclusive': False,
'global_pooling': global_pool,
} }
self.paddle_graph.add_layer( if params.pool == 0:
"paddle.fluid.dygraph.Pool2D", self.paddle_graph.add_layer(
inputs={"input": input.name}, "paddle.nn.MaxPool2D",
outputs=layer_outputs, inputs={"input": input.name},
**layer_attrs) outputs=layer_outputs,
# layer_attrs = { **layer_attrs)
# 'kernel_size': kernel, else:
# 'stride': stride, self.paddle_graph.add_layer(
# 'padding': pad, "paddle.nn.AvgPool2D",
# 'ceil_mode': ceil_mode, inputs={"input": input.name},
# } outputs=layer_outputs,
# if params.pool == 0: **layer_attrs)
# self.paddle_graph.add_layer(
# "paddle.nn.MaxPool2D",
# inputs={"input": input.name},
# outputs=layer_outputs,
# **layer_attrs)
# else:
# layer_attrs["count_include_pad"] = True
# self.paddle_graph.add_layer(
# "paddle.nn.AvgPool2D",
# inputs={"input": input.name},
# outputs=layer_outputs,
# **layer_attrs)
def LRN(self, node): def LRN(self, node):
lrn_name = name_generator("lrn", self.nn_name2id)
output_name = node.layer_name
layer_outputs = [lrn_name, output_name]
assert len(node.inputs) == 1, "The count of LRN node\'s input is not 1." assert len(node.inputs) == 1, "The count of LRN node\'s input is not 1."
input = self.graph.get_input_node(node, idx=0, copy=True) input = self.graph.get_input_node(node, idx=0, copy=True)
params = node.layer.lrn_param params = node.layer.lrn_param
assert params.local_size % 2 == 1 assert params.local_size % 2 == 1
alpha = params.alpha / float(params.local_size) alpha = params.alpha / float(params.local_size)
layer_attrs = { layer_attrs = {
"n": params.local_size, "size": params.local_size,
"k": params.k, "k": params.k,
"alpha": alpha, "alpha": alpha,
"beta": params.beta, "beta": params.beta
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
"fluid.layers.lrn", "paddle.nn.LocalResponseNorm",
inputs={"input": input.name}, inputs={"input": input.name},
outputs=[node.layer_name], outputs=layer_outputs,
**layer_attrs) **layer_attrs)
def InnerProduct(self, node): def InnerProduct(self, node):
linear_name = name_generator("linear", self.nn_name2id) linear_name = name_generator("linear", self.nn_name2id)
output_name = node.layer_name output_name = node.layer_name
...@@ -1131,7 +1120,7 @@ class CaffeOpMapper(OpMapper): ...@@ -1131,7 +1120,7 @@ class CaffeOpMapper(OpMapper):
input = self.graph.get_input_node(node, idx=0, copy=True) input = self.graph.get_input_node(node, idx=0, copy=True)
params = node.layer.shuffle_channel_param params = node.layer.shuffle_channel_param
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
"fluid.layers.shuffle_channel", "paddle.fluid.layers.shuffle_channel",
inputs={"x": input.name}, inputs={"x": input.name},
outputs=[node.layer_name], outputs=[node.layer_name],
group=params.group) group=params.group)
......
x2paddle/op_mapper/dygraph/onnx2paddle/opset9/opset.py
浏览文件 @ f08b1f3f
...@@ -14,8 +14,6 @@ ...@@ -14,8 +14,6 @@
from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode
from x2paddle.core.graph import GraphNode from x2paddle.core.graph import GraphNode
from x2paddle.core.fluid_code import Layer
from x2paddle.core.fluid_code import FluidCode
from x2paddle.core.util import * from x2paddle.core.util import *
from functools import reduce from functools import reduce
import numpy as np import numpy as np
...@@ -86,7 +84,7 @@ class OpSet9(): ...@@ -86,7 +84,7 @@ class OpSet9():
elementwise_ops = { elementwise_ops = {
'Add': 'paddle.add', 'Add': 'paddle.add',
'Div': 'paddle.divide', 'Div': 'paddle.divide',
'Sub': 'fluid.layers.elementwise_sub', 'Sub': 'paddle.subtract',
'Mul': 'paddle.multiply', 'Mul': 'paddle.multiply',
'Pow': 'paddle.pow', 'Pow': 'paddle.pow',
} }
...@@ -281,16 +279,11 @@ class OpSet9(): ...@@ -281,16 +279,11 @@ class OpSet9():
inputs={"x": var_hw}, inputs={"x": var_hw},
outputs=[var_hw], outputs=[var_hw],
dtype=string('int32')) dtype=string('int32'))
# inputs['size'] = var_hw inputs['size'] = var_hw
attrs = {"align_corners": False,
# TODO(syf): all use "mode": string(node.get_attr('mode', 'nearest'))}
inputs['out_shape'] = var_hw
ipt = inputs.pop("x")
inputs["input"] = ipt
mode = node.get_attr('mode', 'nearest')
attrs.update({"align_corners": False})
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.resize_nearest", kernel="paddle.nn.functional.interpolate",
inputs=inputs, inputs=inputs,
outputs=[node.name], outputs=[node.name],
**attrs) **attrs)
...@@ -356,7 +349,7 @@ class OpSet9(): ...@@ -356,7 +349,7 @@ class OpSet9():
'sampling_ratio': sampling_ratio, 'sampling_ratio': sampling_ratio,
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
'fluid.layers.roi_align', 'paddle.fluid.layers.roi_align',
inputs={'input': val_x.name, inputs={'input': val_x.name,
'rois': val_rois.name}, 'rois': val_rois.name},
outputs=[node.name], outputs=[node.name],
...@@ -376,7 +369,7 @@ class OpSet9(): ...@@ -376,7 +369,7 @@ class OpSet9():
'spatial_scale': spatial_scale, 'spatial_scale': spatial_scale,
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
'fluid.layers.roi_pool', 'paddle.fluid.layers.roi_pool',
inputs={'input': val_x.name, inputs={'input': val_x.name,
'rois': val_rois.name}, 'rois': val_rois.name},
outputs=[node.name], outputs=[node.name],
...@@ -405,7 +398,7 @@ class OpSet9(): ...@@ -405,7 +398,7 @@ class OpSet9():
layer_attrs['data_format'] = string('NCHW') layer_attrs['data_format'] = string('NCHW')
layer_attrs['value'] = value layer_attrs['value'] = value
else: else:
paddle_op = 'fluid.layers.pad' paddle_op = 'paddle.fluid.layers.pad'
layer_attrs["pad_value"] = value layer_attrs["pad_value"] = value
if len(pads) == 4: if len(pads) == 4:
paddings = np.array(pads).reshape( paddings = np.array(pads).reshape(
...@@ -1062,40 +1055,23 @@ class OpSet9(): ...@@ -1062,40 +1055,23 @@ class OpSet9():
strides[1]) strides[1])
paddings = pad_h + pad_w paddings = pad_h + pad_w
paddle_op = 'fluid.layers.pool{}d'.format(poolnd) op_name = name_generator("pool", self.nn_name2id)
assert 2 <= poolnd <= 3, 'only pool2d and pool3d are supported' 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 = { layer_attrs = {
"pool_size": kernel_shape, "kernel_size": kernel_shape,
"pool_type": string('avg'), "stride": strides,
"pool_stride": strides, "padding": paddings,
"pool_padding": paddings,
"ceil_mode": ceil_mode, "ceil_mode": ceil_mode,
"exclusive": 'True', "exclusive": 'True',
"name": string(node.name)
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
paddle_op, paddle_op,
inputs={'input': val_x if isinstance(val_x, str) else val_x.name}, inputs={'x': val_x.name},
outputs=[node.name], outputs=layer_outputs,
**layer_attrs) **layer_attrs)
# TODO(syf): op has diff
# op_name = name_generator("pool", self.nn_name2id)
# output_name = node.name
# layer_outputs = [op_name, output_name]
# paddle_op = 'paddle.nn.Pool{}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 @print_mapping_info
def Concat(self, node): def Concat(self, node):
...@@ -1657,4 +1633,4 @@ class OpSet9(): ...@@ -1657,4 +1633,4 @@ class OpSet9():
'paddle.argmax', 'paddle.argmax',
inputs={"x": val_x.name}, inputs={"x": val_x.name},
outputs=[node.name], outputs=[node.name],
**layer_attrs) **layer_attrs)
\ No newline at end of file
x2paddle/op_mapper/dygraph/pytorch2paddle/aten.py
浏览文件 @ f08b1f3f
...@@ -426,11 +426,11 @@ def aten_avg_pool2d(mapper, graph, node): ...@@ -426,11 +426,11 @@ def aten_avg_pool2d(mapper, graph, node):
# 获取当前节点输入的list # 获取当前节点输入的list
current_inputs = list(layer_inputs.values()) current_inputs = list(layer_inputs.values())
# 处理输入1，即%538 # 处理输入1，即%538
layer_attrs["pool_size"] = mapper.attrs[inputs_name[1]] layer_attrs["kernel_size"] = mapper.attrs[inputs_name[1]]
# 处理输入2，即%539 # 处理输入2，即%539
layer_attrs["pool_stride"] = mapper.attrs[inputs_name[2]] layer_attrs["stride"] = mapper.attrs[inputs_name[2]]
# 处理输入3，即%540 # 处理输入3，即%540
layer_attrs["pool_padding"] = mapper.attrs[inputs_name[3]] layer_attrs["padding"] = mapper.attrs[inputs_name[3]]
# 处理输入4，即%273 # 处理输入4，即%273
layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[4]] layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[4]]
# 处理输入5，即%272 # 处理输入5，即%272
...@@ -445,22 +445,13 @@ def aten_avg_pool2d(mapper, graph, node): ...@@ -445,22 +445,13 @@ def aten_avg_pool2d(mapper, graph, node):
key=mapper.attrs[inputs_name[6]], key=mapper.attrs[inputs_name[6]],
value=None) value=None)
# TODO(syf): The op has diff.
# self.paddle_graph.add_layer(
# kernel="paddle.nn.AvgPool2D",
# inputs={"input": input_name},
# outputs=layer_outputs,
# kernel_size=k_size[2:4],
# stride=strides[2:4],
# padding=string(pad_mode))
layer_attrs["pool_type"] = string("avg")
graph.add_layer( graph.add_layer(
"fluid.layers.pool2d", kernel="paddle.nn.AvgPool2D",
inputs=layer_inputs, inputs=layer_inputs,
outputs=layer_outputs[1:], outputs=layer_outputs,
scope_name=scope_name, scope_name=scope_name,
**layer_attrs) **layer_attrs)
return current_inputs, current_outputs return current_inputs, current_outputs
def aten_avg_pool3d(mapper, graph, node): def aten_avg_pool3d(mapper, graph, node):
...@@ -493,11 +484,11 @@ def aten_avg_pool3d(mapper, graph, node): ...@@ -493,11 +484,11 @@ def aten_avg_pool3d(mapper, graph, node):
# 获取当前节点输入的list # 获取当前节点输入的list
current_inputs = list(layer_inputs.values()) current_inputs = list(layer_inputs.values())
# 处理输入1，即%538 # 处理输入1，即%538
layer_attrs["pool_size"] = mapper.attrs[inputs_name[1]] layer_attrs["kernel_size"] = mapper.attrs[inputs_name[1]]
# 处理输入2，即%539 # 处理输入2，即%539
layer_attrs["pool_stride"] = mapper.attrs[inputs_name[2]] layer_attrs["stride"] = mapper.attrs[inputs_name[2]]
# 处理输入3，即%540 # 处理输入3，即%540
layer_attrs["pool_padding"] = mapper.attrs[inputs_name[3]] layer_attrs["padding"] = mapper.attrs[inputs_name[3]]
# 处理输入4，即%273 # 处理输入4，即%273
layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[4]] layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[4]]
# 处理输入5，即%272 # 处理输入5，即%272
...@@ -512,20 +503,10 @@ def aten_avg_pool3d(mapper, graph, node): ...@@ -512,20 +503,10 @@ def aten_avg_pool3d(mapper, graph, node):
key=mapper.attrs[inputs_name[6]], key=mapper.attrs[inputs_name[6]],
value=None) value=None)
# TODO(syf): The op has diff.
# self.paddle_graph.add_layer(
# kernel="paddle.nn.AvgPool2D",
# inputs={"input": input_name},
# outputs=layer_outputs,
# kernel_size=k_size[2:4],
# stride=strides[2:4],
# padding=string(pad_mode))
layer_attrs["pool_type"] = string("avg")
graph.add_layer( graph.add_layer(
"fluid.layers.pool3d", kernel="paddle.nn.AvgPool3D",
inputs=layer_inputs, inputs=layer_inputs,
outputs=layer_outputs[1:], outputs=layer_outputs,
scope_name=scope_name, scope_name=scope_name,
**layer_attrs) **layer_attrs)
return current_inputs, current_outputs return current_inputs, current_outputs
...@@ -561,11 +542,11 @@ def aten_avg_pool1d(mapper, graph, node): ...@@ -561,11 +542,11 @@ def aten_avg_pool1d(mapper, graph, node):
# 获取当前节点输入的list # 获取当前节点输入的list
current_inputs = list(layer_inputs.values()) current_inputs = list(layer_inputs.values())
# 处理输入1，即%538 # 处理输入1，即%538
layer_attrs["pool_size"] = mapper.attrs[inputs_name[1]] layer_attrs["kernel_size"] = mapper.attrs[inputs_name[1]]
# 处理输入2，即%539 # 处理输入2，即%539
layer_attrs["pool_stride"] = mapper.attrs[inputs_name[2]] layer_attrs["stride"] = mapper.attrs[inputs_name[2]]
# 处理输入3，即%540 # 处理输入3，即%540
layer_attrs["pool_padding"] = mapper.attrs[inputs_name[3]] layer_attrs["padding"] = mapper.attrs[inputs_name[3]]
# 处理输入4，即%273 # 处理输入4，即%273
layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[4]] layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[4]]
# 处理输入5，即%272 # 处理输入5，即%272
...@@ -580,20 +561,10 @@ def aten_avg_pool1d(mapper, graph, node): ...@@ -580,20 +561,10 @@ def aten_avg_pool1d(mapper, graph, node):
key=mapper.attrs[inputs_name[6]], key=mapper.attrs[inputs_name[6]],
value=None) value=None)
# TODO(syf): The op has diff.
# self.paddle_graph.add_layer(
# kernel="paddle.nn.AvgPool2D",
# inputs={"input": input_name},
# outputs=layer_outputs,
# kernel_size=k_size[2:4],
# stride=strides[2:4],
# padding=string(pad_mode))
layer_attrs["pool_type"] = string("avg")
graph.add_layer( graph.add_layer(
"fluid.layers.pool1d", kernel="paddle.nn.AvgPool1D",
inputs=layer_inputs, inputs=layer_inputs,
outputs=layer_outputs[1:], outputs=layer_outputs,
scope_name=scope_name, scope_name=scope_name,
**layer_attrs) **layer_attrs)
return current_inputs, current_outputs return current_inputs, current_outputs
...@@ -929,7 +900,7 @@ def aten_constant_pad_nd(mapper, graph, node): ...@@ -929,7 +900,7 @@ def aten_constant_pad_nd(mapper, graph, node):
outputs=[inputs_name[0] + "_list"], outputs=[inputs_name[0] + "_list"],
scope_name=scope_name) scope_name=scope_name)
block.add_layer( block.add_layer(
"paddle.tensor.unsqueeze", "paddle.unsqueeze",
inputs={"x": inputs_name[0], inputs={"x": inputs_name[0],
"axis": inputs_name[0] + "_list"}, "axis": inputs_name[0] + "_list"},
outputs=[inputs_name[0] + "_var"], outputs=[inputs_name[0] + "_var"],
...@@ -941,7 +912,7 @@ def aten_constant_pad_nd(mapper, graph, node): ...@@ -941,7 +912,7 @@ def aten_constant_pad_nd(mapper, graph, node):
scope_name=scope_name, scope_name=scope_name,
**layer_attrs) **layer_attrs)
block.add_layer( block.add_layer(
"paddle.tensor.squeeze", "paddle.squeeze",
inputs={"x": output_name, inputs={"x": output_name,
"axis": inputs_name[0] + "_list"}, "axis": inputs_name[0] + "_list"},
outputs=[output_name], outputs=[output_name],
...@@ -1703,7 +1674,7 @@ def aten_expand_as(mapper, graph, node): ...@@ -1703,7 +1674,7 @@ def aten_expand_as(mapper, graph, node):
outputs=[inputs_name[1] + "_type"], outputs=[inputs_name[1] + "_type"],
scope_name=scope_name) scope_name=scope_name)
block.add_layer( block.add_layer(
"fluid.layers.cast", "paddle.cast",
inputs={"x": inputs_name[0]}, inputs={"x": inputs_name[0]},
outputs=[inputs_name[0]], outputs=[inputs_name[0]],
scope_name=scope_name, scope_name=scope_name,
...@@ -1722,7 +1693,7 @@ def aten_expand_as(mapper, graph, node): ...@@ -1722,7 +1693,7 @@ def aten_expand_as(mapper, graph, node):
if_layer = graph.layers[list(graph.layers.keys())[-1]] if_layer = graph.layers[list(graph.layers.keys())[-1]]
block = PaddleGraph(source_type="pytorch", parent_layer=if_layer, graph_type="dygraph") block = PaddleGraph(source_type="pytorch", parent_layer=if_layer, graph_type="dygraph")
block.add_layer( block.add_layer(
"fluid.layers.cast", "paddle.cast",
inputs={"x": layer_outputs[0]}, inputs={"x": layer_outputs[0]},
outputs=copy.deepcopy(layer_outputs), outputs=copy.deepcopy(layer_outputs),
scope_name=scope_name, scope_name=scope_name,
...@@ -2515,6 +2486,89 @@ def aten_log(mapper, graph, node): ...@@ -2515,6 +2486,89 @@ def aten_log(mapper, graph, node):
return current_inputs, current_outputs return current_inputs, current_outputs
def aten_lstm(mapper, graph, node):
""" 构造长短期记忆网络（LSTM）的PaddleLayer。
TorchScript示例:
%input.96, %551, %552 = aten::lstm(%input.95, %734, %549, %526, %525, %524, %526, %526, %526)
参数含义:
%input.96 (Tensor): 输出，由前向和后向cell的输出拼接得到。
%551 (Tensor): cell state。
%552 (Tensor): hidden state。
%input.95 (Tensor): 网络输入。
%734 (Tensor): 网络的初始状态。
%549 (list): 所有权重组合成的list。
%526 (bool): 是否使用bias。
%525 (int): 网络层数。
%524 (float): dropout概率。
%526 (bool): 是否为训练阶段。
%526 (bool): 是否使用双向LSTM。
%526 (bool): 第一个维度是否为batch size。
"""
scope_name = mapper.normalize_scope_name(node)
op_name = name_generator("lstm", mapper.nn_name2id)
output_names = mapper._get_outputs_name(node)
layer_outputs = [op_name]
layer_outputs.extend(output_names)
layer_inputs = {}
layer_attrs = {}
inputs_name, inputs_node = mapper._get_inputs_name(node)
# 获取当前节点输出的list
current_outputs = output_names
# 处理输入0，即%input.95
mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs, scope_name)
layer_inputs["input0"] = inputs_name[0]
# 处理输入1，即%734
mapper._check_input(graph, inputs_node[1], inputs_name[1], current_outputs, scope_name)
layer_inputs["input1"] = inputs_name[1]
# 获取当前节点输入、输出的list
current_inputs = list(layer_inputs.values())
# 处理输入2，即%734
mapper._check_input(graph, inputs_node[2], inputs_name[2], current_outputs, scope_name)
graph.layers.pop(mapper.output2id[inputs_name[2]])
param_inputs_name, _ = mapper._get_inputs_name(inputs_node[2])
new_param_inputs_name = list()
for i, param_name in enumerate(param_inputs_name):
if i == 0:
layer_attrs["hidden_size"] = int(mapper.paddle_params[param_name].shape[0] / 4)
layer_attrs["input_size"] = int(mapper.paddle_params[param_name].shape[1])
if len(mapper.paddle_params[param_name].shape) > 1:
part_name = param_name.split("_weight_")[-1]
mapper.paddle_params["{}.weight_{}".format(op_name, part_name)] = mapper.paddle_params[param_name]
new_param_inputs_name.append("{}.weight_{}".format(op_name, part_name))
else:
part_name = param_name.split("_bias_")[-1]
mapper.paddle_params["{}.bias_{}".format(op_name, part_name)] = mapper.paddle_params[param_name]
mapper.paddle_params.pop(param_name)
# 处理输入3，即%526
is_bias = mapper.attrs[inputs_name[3]]
if not is_bias:
for param_name in new_param_inputs_name:
bias_name = param_name.replace("weight", "bias")
bias_shape= mapper.paddle_params[param_name].shape[:1]
mapper.paddle_params[bias_name] = np.zeros(bias_shape).astype("float32")
# 处理输入4，即%525
layer_attrs["num_layers"] = mapper.attrs[inputs_name[4]]
# 处理输入5，即%524
layer_attrs["dropout"] = mapper.attrs[inputs_name[5]]
# 处理输入7，即%526
is_bidirectional = mapper.attrs[inputs_name[7]]
if is_bidirectional:
layer_attrs["direction"] = string("bidirectional")
# 处理输入8，即%526
batch_first = mapper.attrs[inputs_name[8]]
if not batch_first:
layer_attrs["time_major"] = True
graph.add_layer(
"paddle.nn.LSTM",
inputs=layer_inputs,
outputs=layer_outputs,
scope_name=scope_name,
**layer_attrs)
return current_inputs, current_outputs
def aten_lt(mapper, graph, node): def aten_lt(mapper, graph, node):
""" 构造对比大小的PaddleLayer。 """ 构造对比大小的PaddleLayer。
...@@ -2847,22 +2901,13 @@ def aten_max_pool2d(mapper, graph, node): ...@@ -2847,22 +2901,13 @@ def aten_max_pool2d(mapper, graph, node):
# 处理输入5，即%19 # 处理输入5，即%19
layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[5]] layer_attrs["ceil_mode"] = mapper.attrs[inputs_name[5]]
layer_attrs_tmp["ceil_mode"] = mapper.attrs[inputs_name[5]] layer_attrs_tmp["ceil_mode"] = mapper.attrs[inputs_name[5]]
if mapper.attrs[inputs_name[5]] == True: graph.add_layer(
layer_attrs["pool_type"] = string("max") "paddle.nn.MaxPool2D",
graph.add_layer( inputs=layer_inputs,
"fluid.layers.pool2d", outputs=layer_outputs,
inputs=layer_inputs, scope_name=scope_name,
outputs=layer_outputs[1:], **layer_attrs)
scope_name=scope_name,
**layer_attrs_tmp)
else:
graph.add_layer(
"paddle.nn.MaxPool2D",
inputs=layer_inputs,
outputs=layer_outputs,
scope_name=scope_name,
**layer_attrs)
return current_inputs, current_outputs return current_inputs, current_outputs
...@@ -3991,7 +4036,7 @@ def aten_squeeze(mapper, graph, node): ...@@ -3991,7 +4036,7 @@ def aten_squeeze(mapper, graph, node):
layer_inputs["axis"] = inputs_name[1] layer_inputs["axis"] = inputs_name[1]
current_inputs.append(inputs_name[1]) current_inputs.append(inputs_name[1])
graph.add_layer( graph.add_layer(
"paddle.tensor.squeeze", "paddle.squeeze",
inputs=layer_inputs, inputs=layer_inputs,
outputs=layer_outputs, outputs=layer_outputs,
scope_name=scope_name, scope_name=scope_name,
......
x2paddle/op_mapper/dygraph/pytorch2paddle/prim.py
浏览文件 @ f08b1f3f
...@@ -33,11 +33,33 @@ def prim_Constant(mapper, graph, node): ...@@ -33,11 +33,33 @@ def prim_Constant(mapper, graph, node):
output_type = output.type() output_type = output.type()
if isinstance(value, str): if isinstance(value, str):
value = string(value) value = string(value)
if str(output_type) == "Tensor": if "Tensor" in str(output_type):
tensor_value = value tensor_value = value
value = "{}".format(value) value = "{}".format(value)
if "tensor" in value: if "tensor" in value:
mapper.pytorch_params[output_name] = tensor_value.cpu().detach().numpy() if isinstance(tensor_value, list) or isinstance(tensor_value, tuple):
name_dict = dict()
for i, tv in enumerate(tensor_value):
output_name_i = "{}_p{}".format(output_name,i)
key_i = "input{}".format(i)
mapper.paddle_params[output_name_i] = tv.cpu().detach().numpy()
graph.add_layer(
"self.create_parameter",
inputs={},
outputs=[output_name_i],
scope_name=scope_name,
dtype=string(str(mapper.paddle_params[output_name_i].dtype)),
shape = mapper.paddle_params[output_name_i].shape,
default_initializer="paddle.nn.initializer.Constant(value=0.0)")
name_dict[key_i] = output_name_i
graph.add_layer(
"prim.list",
inputs=name_dict,
outputs=[output_name],
scope_name=scope_name)
return [], [output_name]
else:
mapper.pytorch_params[output_name] = tensor_value.cpu().detach().numpy()
if "inf" in str(value): if "inf" in str(value):
t = str(type(value)).split("'")[1] t = str(type(value)).split("'")[1]
...@@ -218,11 +240,13 @@ def prim_ListConstruct(mapper, graph, node): ...@@ -218,11 +240,13 @@ def prim_ListConstruct(mapper, graph, node):
current_outputs = [output_name] current_outputs = [output_name]
# 处理每个输入 # 处理每个输入
for i, input_name in enumerate(inputs_name): for i, input_name in enumerate(inputs_name):
mapper._check_input(graph, inputs_node[i], input_name, current_outputs, scope_name)
layer_inputs["input{}".format(i)] = input_name layer_inputs["input{}".format(i)] = input_name
# 获取当前节点输入的list # 获取当前节点输入的list
current_inputs = list(layer_inputs.values()) current_inputs = list(layer_inputs.values())
graph.add_layer("prim.list", inputs=layer_inputs, outputs=layer_outputs, scope_name=scope_name) layer_id = graph.add_layer("prim.list", inputs=layer_inputs, outputs=layer_outputs, scope_name=scope_name)
mapper.output2id[output_name] = layer_id
return current_inputs, current_outputs return current_inputs, current_outputs
......
x2paddle/op_mapper/dygraph/pytorch2paddle/pytorch_custom_layer/gather.py
浏览文件 @ f08b1f3f
...@@ -13,7 +13,6 @@ ...@@ -13,7 +13,6 @@
# limitations under the License. # limitations under the License.
import paddle import paddle
import paddle.fluid as fluid
from itertools import product from itertools import product
import numpy as np import numpy as np
......
x2paddle/op_mapper/dygraph/pytorch2paddle/pytorch_op_mapper.py
浏览文件 @ f08b1f3f
...@@ -37,6 +37,7 @@ class PyTorchOpMapper(OpMapper): ...@@ -37,6 +37,7 @@ class PyTorchOpMapper(OpMapper):
self.scope_name_list = list() self.scope_name_list = list()
self.scope_name2id = dict() self.scope_name2id = dict()
self.inputs_info = dict() self.inputs_info = dict()
self.output2id = dict() # output名字和layer_id的关系，用于lstm去除前面的node
# 转换 # 转换
if not self.op_checker(decoder.graph): if not self.op_checker(decoder.graph):
raise Exception("Model is not supported yet.") raise Exception("Model is not supported yet.")
...@@ -175,7 +176,7 @@ class PyTorchOpMapper(OpMapper): ...@@ -175,7 +176,7 @@ class PyTorchOpMapper(OpMapper):
if add_dim: if add_dim:
param = param[np.newaxis, :] param = param[np.newaxis, :]
self.paddle_params[output_name] = param self.paddle_params[output_name] = param
graph.add_layer( layer_id = graph.add_layer(
"self.create_parameter", "self.create_parameter",
inputs={}, inputs={},
outputs=[output_name], outputs=[output_name],
...@@ -183,6 +184,7 @@ class PyTorchOpMapper(OpMapper): ...@@ -183,6 +184,7 @@ class PyTorchOpMapper(OpMapper):
dtype=string(str(param.dtype)), dtype=string(str(param.dtype)),
shape = param.shape, shape = param.shape,
default_initializer="paddle.nn.initializer.Constant(value=0.0)") default_initializer="paddle.nn.initializer.Constant(value=0.0)")
self.output2id[output_name] = layer_id
else: else:
if isinstance(param, dict) and "Tensor" in param and \ if isinstance(param, dict) and "Tensor" in param and \
"parent_layer_id" in param: "parent_layer_id" in param:
...@@ -202,7 +204,7 @@ class PyTorchOpMapper(OpMapper): ...@@ -202,7 +204,7 @@ class PyTorchOpMapper(OpMapper):
if add_dim: if add_dim:
param = param[np.newaxis, :] param = param[np.newaxis, :]
self.paddle_params[output_name] = param self.paddle_params[output_name] = param
graph.add_layer( layer_id = graph.add_layer(
"self.create_parameter", "self.create_parameter",
inputs={}, inputs={},
outputs=[output_name], outputs=[output_name],
...@@ -211,6 +213,7 @@ class PyTorchOpMapper(OpMapper): ...@@ -211,6 +213,7 @@ class PyTorchOpMapper(OpMapper):
shape = param.shape, shape = param.shape,
default_initializer="paddle.nn.initializer.Constant(value=0.0)") default_initializer="paddle.nn.initializer.Constant(value=0.0)")
node_outputs.append(output_name) node_outputs.append(output_name)
self.output2id[output_name] = layer_id
return return
# 若if-else外，则可直接引用if-else中的赋值结果 # 若if-else外，则可直接引用if-else中的赋值结果
graph.add_layer( graph.add_layer(
...@@ -231,14 +234,15 @@ class PyTorchOpMapper(OpMapper): ...@@ -231,14 +234,15 @@ class PyTorchOpMapper(OpMapper):
elif node.kind() == "prim::Constant" and output_name in self.pytorch_params: elif node.kind() == "prim::Constant" and output_name in self.pytorch_params:
param = self.pytorch_params[output_name] param = self.pytorch_params[output_name]
self.paddle_params[output_name] = param self.paddle_params[output_name] = param
graph.add_layer( layer_id = graph.add_layer(
"self.create_parameter", "self.create_parameter",
inputs={}, inputs={},
outputs=[output_name], outputs=[output_name],
scope_name=scope_name, scope_name=scope_name,
dtype=string(str(param.dtype)), dtype=string(str(param.dtype)),
shape = param.shape, shape = param.shape,
default_initializer="paddle.nn.initializer.Constant(value=0.0)") default_initializer="paddle.nn.initializer.Constant(value=0.0)")
self.output2id[output_name] = layer_id
def _get_inputs_name(self, node): def _get_inputs_name(self, node):
......
x2paddle/op_mapper/dygraph/tf2paddle/tf_op_mapper.py
浏览文件 @ f08b1f3f
...@@ -70,7 +70,7 @@ class TFOpMapper(OpMapper): ...@@ -70,7 +70,7 @@ class TFOpMapper(OpMapper):
'AddV2': 'paddle.add', 'AddV2': 'paddle.add',
'RealDiv': 'paddle.divide', 'RealDiv': 'paddle.divide',
'DivNoNan': 'paddle.divide', 'DivNoNan': 'paddle.divide',
'Sub': 'fluid.layers.elementwise_sub', 'Sub': 'paddle.subtract',
'Maximum': 'paddle.maximum', 'Maximum': 'paddle.maximum',
'Minimum': 'paddle.minimum', 'Minimum': 'paddle.minimum',
'Mul': 'paddle.multiply', 'Mul': 'paddle.multiply',
...@@ -346,7 +346,7 @@ class TFOpMapper(OpMapper): ...@@ -346,7 +346,7 @@ class TFOpMapper(OpMapper):
shape=[0, c, h, w]) shape=[0, c, h, w])
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.pixel_shuffle", kernel="paddle.nn.functional.pixel_shuffle",
inputs={"x": reshape_name}, inputs={"x": reshape_name},
outputs=[node.name], outputs=[node.name],
upscale_factor=block_size) upscale_factor=block_size)
...@@ -858,22 +858,22 @@ class TFOpMapper(OpMapper): ...@@ -858,22 +858,22 @@ class TFOpMapper(OpMapper):
layer_outputs = [op_name, output_name] layer_outputs = [op_name, output_name]
# TODO(syf): The op has diff. # TODO(syf): The op has diff.
# self.paddle_graph.add_layer(
# kernel="paddle.nn.AvgPool2D",
# inputs={"input": input_name},
# outputs=layer_outputs,
# kernel_size=k_size[2:4],
# stride=strides[2:4],
# padding=string(pad_mode))
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.pool2d", kernel="paddle.nn.AvgPool2D",
inputs={"input": input_name}, inputs={"input": input_name},
outputs=[node.name], outputs=layer_outputs,
pool_size=k_size[2:4], kernel_size=k_size[2:4],
pool_type=string("avg"), stride=strides[2:4],
pool_stride=strides[2:4], padding=string(pad_mode))
pool_padding=string(pad_mode))
# self.paddle_graph.add_layer(
# kernel="fluid.layers.pool2d",
# inputs={"input": input_name},
# outputs=[node.name],
# pool_size=k_size[2:4],
# pool_type=string("avg"),
# pool_stride=strides[2:4],
# pool_padding=string(pad_mode))
if data_format == "NHWC": if data_format == "NHWC":
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
...@@ -1118,14 +1118,6 @@ class TFOpMapper(OpMapper): ...@@ -1118,14 +1118,6 @@ class TFOpMapper(OpMapper):
begin = begin.value.tolist() begin = begin.value.tolist()
attrs['offsets'] = begin attrs['offsets'] = begin
else: else:
# shape = begin.out_shapes[0]
# reshape_name = gen_name("slice", "reshape")
# self.paddle_graph.add_layer(
# kernel="fluid.layers.reshape",
# inputs={"x": begin.name},
# outputs=[reshape_name],
# shape=shape)
# inputs['offsets'] = reshape_name
begin = self.decoder.infer_tensor(begin, use_diff_inputs=False).tolist() begin = self.decoder.infer_tensor(begin, use_diff_inputs=False).tolist()
attrs['offsets'] = begin attrs['offsets'] = begin
if size.layer_type == "Const": if size.layer_type == "Const":
...@@ -1433,7 +1425,7 @@ class TFOpMapper(OpMapper): ...@@ -1433,7 +1425,7 @@ class TFOpMapper(OpMapper):
y_shape = y.out_shapes[0] y_shape = y.out_shapes[0]
# TODO(syf) # TODO(syf)
layer_id = self.paddle_graph.add_layer( layer_id = self.paddle_graph.add_layer(
"fluid.layers.elementwise_sub", inputs=inputs, outputs=[node.name]) "paddle.subtract", inputs=inputs, outputs=[node.name])
self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape} self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape}
inputs = {"x": node.name, "y": node.name} inputs = {"x": node.name, "y": node.name}
......
x2paddle/op_mapper/static/caffe2paddle/caffe_op_mapper.py
浏览文件 @ f08b1f3f
...@@ -401,18 +401,14 @@ class CaffeOpMapper(OpMapper): ...@@ -401,18 +401,14 @@ class CaffeOpMapper(OpMapper):
padding=pad, padding=pad,
ceil_mode=ceil_mode) ceil_mode=ceil_mode)
else: else:
# TODO(syf): The op has diff.
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.pool2d", kernel="paddle.nn.functional.avg_pool2d",
inputs={"input": input.name}, inputs={"x": input.name},
outputs=[node.name], outputs=[node.name],
pool_size=kernel, kernel_size=kernel,
pool_type=string("avg"), stride=stride,
pool_stride=stride, padding=pad,
pool_padding=pad, ceil_mode=ceil_mode)
ceil_mode=ceil_mode,
exclusive=False,
global_pooling=False)
def LRN(self, node): def LRN(self, node):
assert len(node.inputs) == 1, 'The count of LRN node\'s input is not 1.' assert len(node.inputs) == 1, 'The count of LRN node\'s input is not 1.'
...@@ -433,7 +429,7 @@ class CaffeOpMapper(OpMapper): ...@@ -433,7 +429,7 @@ class CaffeOpMapper(OpMapper):
'name': string(node.name) 'name': string(node.name)
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.lrn", kernel="paddle.fluid.layers.lrn",
inputs={"input": input.name}, inputs={"input": input.name},
outputs=[node.name], outputs=[node.name],
**layer_attrs) **layer_attrs)
...@@ -1184,7 +1180,7 @@ class CaffeOpMapper(OpMapper): ...@@ -1184,7 +1180,7 @@ class CaffeOpMapper(OpMapper):
input = self.graph.get_input_node(node, idx=0, copy=True) input = self.graph.get_input_node(node, idx=0, copy=True)
params = node.layer.shuffle_channel_param params = node.layer.shuffle_channel_param
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
"fluid.layers.shuffle_channel", "paddle.fluid.layers.shuffle_channel",
inputs={"x": input.name}, inputs={"x": input.name},
outputs=[node.layer_name], outputs=[node.layer_name],
group=params.group) group=params.group)
......
x2paddle/op_mapper/static/onnx2paddle/opset9/opset.py
浏览文件 @ f08b1f3f
...@@ -14,8 +14,6 @@ ...@@ -14,8 +14,6 @@
from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode
from x2paddle.core.graph import GraphNode from x2paddle.core.graph import GraphNode
from x2paddle.core.fluid_code import Layer
from x2paddle.core.fluid_code import FluidCode
from x2paddle.core.util import string from x2paddle.core.util import string
from functools import reduce from functools import reduce
import numpy as np import numpy as np
...@@ -88,7 +86,7 @@ class OpSet9(): ...@@ -88,7 +86,7 @@ class OpSet9():
elementwise_ops = { elementwise_ops = {
'Add': 'paddle.add', 'Add': 'paddle.add',
'Div': 'paddle.divide', 'Div': 'paddle.divide',
'Sub': 'fluid.layers.elementwise_sub', 'Sub': 'paddle.subtract',
'Mul': 'paddle.multiply', 'Mul': 'paddle.multiply',
'Pow': 'paddle.pow', 'Pow': 'paddle.pow',
} }
...@@ -271,16 +269,11 @@ class OpSet9(): ...@@ -271,16 +269,11 @@ class OpSet9():
inputs={"x": var_hw}, inputs={"x": var_hw},
outputs=[var_hw], outputs=[var_hw],
dtype=string('int32')) dtype=string('int32'))
# inputs['size'] = var_hw inputs['size'] = var_hw
attrs = {"align_corners": False,
# TODO(syf): all use "mode": string(node.get_attr('mode', 'nearest'))}
inputs['out_shape'] = var_hw
ipt = inputs.pop("x")
inputs["input"] = ipt
mode = node.get_attr('mode', 'nearest')
attrs.update({"align_corners": False})
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.resize_nearest", kernel="paddle.nn.functional.interpolate",
inputs=inputs, inputs=inputs,
outputs=[node.name], outputs=[node.name],
**attrs) **attrs)
...@@ -346,7 +339,7 @@ class OpSet9(): ...@@ -346,7 +339,7 @@ class OpSet9():
'sampling_ratio': sampling_ratio, 'sampling_ratio': sampling_ratio,
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
'fluid.layers.roi_align', 'paddle.fluid.layers.roi_align',
inputs={'input': val_x.name, inputs={'input': val_x.name,
'rois': val_rois.name}, 'rois': val_rois.name},
outputs=[node.name], outputs=[node.name],
...@@ -365,7 +358,7 @@ class OpSet9(): ...@@ -365,7 +358,7 @@ class OpSet9():
'spatial_scale': spatial_scale, 'spatial_scale': spatial_scale,
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
'fluid.layers.roi_pool', 'paddle.fluid.layers.roi_pool',
inputs={'input': val_x.name, inputs={'input': val_x.name,
'rois': val_rois.name}, 'rois': val_rois.name},
outputs=[node.name], outputs=[node.name],
...@@ -394,7 +387,7 @@ class OpSet9(): ...@@ -394,7 +387,7 @@ class OpSet9():
layer_attrs['data_format'] = string('NCHW') layer_attrs['data_format'] = string('NCHW')
layer_attrs['value'] = value layer_attrs['value'] = value
else: else:
paddle_op = 'fluid.layers.pad' paddle_op = 'paddle.fluid.layers.pad'
layer_attrs["pad_value"] = value layer_attrs["pad_value"] = value
if len(pads) == 4: if len(pads) == 4:
paddings = np.array(pads).reshape( paddings = np.array(pads).reshape(
...@@ -1046,23 +1039,21 @@ class OpSet9(): ...@@ -1046,23 +1039,21 @@ class OpSet9():
strides[1]) strides[1])
paddings = pad_h + pad_w paddings = pad_h + pad_w
paddle_op = 'fluid.layers.pool{}d'.format(poolnd) paddle_op = 'paddle.nn.functional.avg_pool{}d'.format(poolnd)
assert 2 <= poolnd <= 3, 'only pool2d and pool3d are supported' assert 1 <= poolnd <= 3, 'only avg_pool1d, avg_pool2d and avg_pool3d are supported'
layer_attrs = { layer_attrs = {
"pool_size": kernel_shape, "kernel_size": kernel_shape,
"pool_type": string('avg'), "stride": strides,
"pool_stride": strides, "padding": paddings,
"pool_padding": paddings,
"ceil_mode": ceil_mode, "ceil_mode": ceil_mode,
"exclusive": 'True', "exclusive": True,
"name": string(node.name) "name": string(node.name)
} }
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
paddle_op, paddle_op,
inputs={'input': val_x if isinstance(val_x, str) else val_x.name}, inputs={'x': val_x if isinstance(val_x, str) else val_x.name},
outputs=[node.name], outputs=[node.name],
**layer_attrs) **layer_attrs)
# TODO(syf): op has diff
@print_mapping_info @print_mapping_info
def Concat(self, node): def Concat(self, node):
......
x2paddle/op_mapper/static/tf2paddle/tf_op_mapper.py
浏览文件 @ f08b1f3f
...@@ -72,7 +72,7 @@ class TFOpMapper(OpMapper): ...@@ -72,7 +72,7 @@ class TFOpMapper(OpMapper):
'RealDiv': 'paddle.divide', 'RealDiv': 'paddle.divide',
'DivNoNan': 'paddle.divide', 'DivNoNan': 'paddle.divide',
# TODO (syf): replace # TODO (syf): replace
'Sub': 'fluid.layers.elementwise_sub', 'Sub': 'paddle.subtract',
'Maximum': 'paddle.maximum', 'Maximum': 'paddle.maximum',
'Minimum': 'paddle.minimum', 'Minimum': 'paddle.minimum',
'Mul': 'paddle.multiply', 'Mul': 'paddle.multiply',
...@@ -315,7 +315,7 @@ class TFOpMapper(OpMapper): ...@@ -315,7 +315,7 @@ class TFOpMapper(OpMapper):
shape=[0, c, h, w]) shape=[0, c, h, w])
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.pixel_shuffle", kernel="paddle.nn.functional.pixel_shuffle",
inputs={"x": reshape_name}, inputs={"x": reshape_name},
outputs=[node.name], outputs=[node.name],
upscale_factor=block_size) upscale_factor=block_size)
...@@ -437,8 +437,6 @@ class TFOpMapper(OpMapper): ...@@ -437,8 +437,6 @@ class TFOpMapper(OpMapper):
if c == -1: if c == -1:
attr = {"shape": [0, k_size[2], 0, 0]} attr = {"shape": [0, k_size[2], 0, 0]}
node.fluid_code.add_layer(
"reshape", inputs=input, output=input, param_attr=attr)
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="paddle.reshape", kernel="paddle.reshape",
inputs={"x": input_name}, inputs={"x": input_name},
...@@ -842,13 +840,12 @@ class TFOpMapper(OpMapper): ...@@ -842,13 +840,12 @@ class TFOpMapper(OpMapper):
# TODO(syf): The op has diff. # TODO(syf): The op has diff.
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
kernel="fluid.layers.pool2d", kernel="paddle.nn.functional.avg_pool2d",
inputs={"input": input_name}, inputs={"x": input_name},
outputs=[node.name], outputs=[node.name],
pool_size=k_size[2:4], kernel_size=k_size[2:4],
pool_type=string("avg"), stride=strides[2:4],
pool_stride=strides[2:4], padding=string(pad_mode))
pool_padding=string(pad_mode))
if data_format == "NHWC": if data_format == "NHWC":
self.paddle_graph.add_layer( self.paddle_graph.add_layer(
...@@ -1406,7 +1403,7 @@ class TFOpMapper(OpMapper): ...@@ -1406,7 +1403,7 @@ class TFOpMapper(OpMapper):
y_shape = y.out_shapes[0] y_shape = y.out_shapes[0]
# TODO(syf) # TODO(syf)
layer_id = self.paddle_graph.add_layer( layer_id = self.paddle_graph.add_layer(
"fluid.layers.elementwise_sub", inputs=inputs, outputs=[node.name]) "paddle.subtract", inputs=inputs, outputs=[node.name])
self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape} self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape}
inputs = {"x": node.name, "y": node.name} inputs = {"x": node.name, "y": node.name}
......
x2paddle/optimizer/fusion/dygraph/bn_scale_fuser.py
浏览文件 @ f08b1f3f
...@@ -21,47 +21,94 @@ from x2paddle.core.util import * ...@@ -21,47 +21,94 @@ from x2paddle.core.util import *
class DygraphBNScaleFuser(FuseBase): class DygraphBNScaleFuser(FuseBase):
def __init__(self): def __init__(self):
super(DygraphBNScaleFuser, self).__init__(graph_type="dygraph") super(DygraphBNScaleFuser, self).__init__(graph_type="dygraph")
patterns = list()
def build_pattern(self): def build_pattern(self):
""" 描述需要替换的batchnorm2d图结构。 """ 描述需要替换的batchnorm2d图结构。
batchnorm2d层模式python实现代码示例: batchnorm2d层模式python实现代码示例:
模式一：
bn_conv1 = self.batchnorm0(conv1) bn_conv1 = self.batchnorm0(conv1)
scale_conv1_cparam1 = self.scale_conv1_cparam1 scale_conv1_cparam1 = self.scale_conv1_cparam1
scale_conv1_mul = paddle.multiply(x=bn_conv1, y=scale_conv1_cparam1, axis=1) scale_conv1_mul = paddle.multiply(x=bn_conv1, y=scale_conv1_cparam1, axis=1)
scale_conv1_cparam2 = self.scale_conv1_cparam2 scale_conv1_cparam2 = self.scale_conv1_cparam2
scale_conv1 = fluid.layers.elementwise_add(x=scale_conv1_mul, y=scale_conv1_cparam2, axis=1) scale_conv1 = paddle.add(x=scale_conv1_mul, y=scale_conv1_cparam2, axis=1)
模式二：
bn_conv1 = self.batchnorm0(conv1)
scale_conv1_cparam1 = self.scale_conv1_cparam1
scale_conv1_mul = paddle.multiply(x=bn_conv1, y=scale_conv1_cparam1, axis=1)
scale_conv1_cparam2 = self.scale_conv1_cparam2
scale_conv1_cparam2 = paddle.reshape(x=scale_conv1_cparam2, shape=[32, 1, 1])
scale_conv1 = paddle.add(x=scale_conv1_mul, y=scale_conv1_cparam2, axis=1)
""" """
def gen_name(id): def gen_name(id):
return "x" + str(id) return "x" + str(id)
self.pattern.add_layer( pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.nn.BatchNorm2D",
inputs={"input": "bn-input-0"},
outputs=[gen_name(0)])
pattern.add_layer(
"self.create_parameter",
inputs={},
outputs=[gen_name(1)])
inputs_dict = {}
inputs_dict['x'] = gen_name(0)
inputs_dict['y'] = gen_name(1)
pattern.add_layer(
"paddle.multiply",
inputs=inputs_dict,
outputs=[gen_name(2)])
pattern.add_layer(
"self.create_parameter",
inputs={},
outputs=[gen_name(3)])
inputs_dict = {}
inputs_dict['x'] = gen_name(2)
inputs_dict['y'] = gen_name(3)
pattern.add_layer(
"paddle.add",
inputs=inputs_dict,
outputs=[gen_name(4)])
pattern.build(inputs={"input-0": "bn-input-0"})
self.patterns.append(pattern)
pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.nn.BatchNorm2D", "paddle.nn.BatchNorm2D",
inputs={"input": "bn-input-0"}, inputs={"input": "bn-input-0"},
outputs=[gen_name(0)]) outputs=[gen_name(0)])
self.pattern.add_layer( pattern.add_layer(
"self.create_parameter", "self.create_parameter",
inputs={}, inputs={},
outputs=[gen_name(1)]) outputs=[gen_name(1)])
inputs_dict = {} inputs_dict = {}
inputs_dict['x'] = gen_name(0) inputs_dict['x'] = gen_name(0)
inputs_dict['y'] = gen_name(1) inputs_dict['y'] = gen_name(1)
self.pattern.add_layer( pattern.add_layer(
"paddle.multiply", "paddle.multiply",
inputs=inputs_dict, inputs=inputs_dict,
outputs=[gen_name(2)]) outputs=[gen_name(2)])
self.pattern.add_layer( pattern.add_layer(
"self.create_parameter", "self.create_parameter",
inputs={}, inputs={},
outputs=[gen_name(3)]) outputs=[gen_name(3)])
pattern.add_layer(
"paddle.reshape",
inputs={"x": gen_name(3)},
outputs=[gen_name(3)])
inputs_dict = {} inputs_dict = {}
inputs_dict['x'] = gen_name(2) inputs_dict['x'] = gen_name(2)
inputs_dict['y'] = gen_name(3) inputs_dict['y'] = gen_name(3)
self.pattern.add_layer( pattern.add_layer(
"fluid.layers.elementwise_add", "paddle.add",
inputs=inputs_dict, inputs=inputs_dict,
outputs=[gen_name(4)]) outputs=[gen_name(4)])
self.pattern.build(inputs={"input-0": "bn-input-0"}) pattern.build(inputs={"input-0": "bn-input-0"})
self.patterns.append(pattern)
def insert_new_layer(self, graph, parameters, matches): def insert_new_layer(self, graph, parameters, matches):
new_layer = self.gen_new_layer(parameters, matches) new_layer = self.gen_new_layer(parameters, matches)
...@@ -78,7 +125,7 @@ class DygraphBNScaleFuser(FuseBase): ...@@ -78,7 +125,7 @@ class DygraphBNScaleFuser(FuseBase):
layer_attrs = layer.attrs layer_attrs = layer.attrs
layer_attrs.pop("weight_attr") layer_attrs.pop("weight_attr")
layer_attrs.pop("bias_attr") layer_attrs.pop("bias_attr")
layer = matches[layers_id[4]] layer = matches[layers_id[-1]]
layer_outputs = [bn_name] + layer.outputs layer_outputs = [bn_name] + layer.outputs
layer = matches[layers_id[1]] layer = matches[layers_id[1]]
data0_name = layer.outputs[0] data0_name = layer.outputs[0]
......
x2paddle/optimizer/fusion/dygraph/reshape_fuser.py
浏览文件 @ f08b1f3f
...@@ -27,7 +27,7 @@ class DygraphReshapeFuser(FuseBase): ...@@ -27,7 +27,7 @@ class DygraphReshapeFuser(FuseBase):
reshape层模式python实现代码示例: reshape层模式python实现代码示例:
x165 = int(x164) x165 = int(x164)
x166 = [x158, x159, x165] x166 = [x158, x159, x165]
x167 = fluid.layers.reshape(x=x157, shape=x166) x167 = paddle.reshape(x=x157, shape=x166)
""" """
def gen_name(id): def gen_name(id):
...@@ -46,7 +46,7 @@ class DygraphReshapeFuser(FuseBase): ...@@ -46,7 +46,7 @@ class DygraphReshapeFuser(FuseBase):
}, },
outputs=[gen_name(1)]) outputs=[gen_name(1)])
self.pattern.add_layer( self.pattern.add_layer(
"fluid.layers.reshape", "paddle.reshape",
inputs={"x": "reshape-input-3", inputs={"x": "reshape-input-3",
"shape": gen_name(1)}, "shape": gen_name(1)},
outputs=[gen_name(2)]) outputs=[gen_name(2)])
......
x2paddle/optimizer/fusion/dygraph/trace_fc_fuser.py
浏览文件 @ f08b1f3f
...@@ -49,7 +49,7 @@ class TraceFcFuser(FuseBase): ...@@ -49,7 +49,7 @@ class TraceFcFuser(FuseBase):
inputs={}, inputs={},
outputs=[gen_name(0)]) outputs=[gen_name(0)])
pattern.add_layer( pattern.add_layer(
"fluid.layers.transpose", "paddle.transpose",
inputs={"x": gen_name(0)}, inputs={"x": gen_name(0)},
outputs=[gen_name(1)], outputs=[gen_name(1)],
perm=[1, 0]) perm=[1, 0])
......
x2paddle/optimizer/fusion/static/bn_scale_fuser.py
浏览文件 @ f08b1f3f
...@@ -21,12 +21,14 @@ from x2paddle.core.util import * ...@@ -21,12 +21,14 @@ from x2paddle.core.util import *
class Static_BNScaleFuser(FuseBase): class Static_BNScaleFuser(FuseBase):
def __init__(self): def __init__(self):
super(Static_BNScaleFuser, self).__init__(graph_type="static") super(Static_BNScaleFuser, self).__init__(graph_type="static")
patterns = list() self.patterns = list()
def build_pattern(self): def build_pattern(self):
""" 描述需要替换的batchnorm2d图结构。 """ 描述需要替换的batchnorm2d图结构。
batchnorm2d层模式python实现代码示例: batchnorm2d层模式python实现代码示例:
模式一： 模式一：
conv1_bn_mean = paddle.static.create_parameter(shape=(128,), dtype='float32', name='conv1_bn_mean')
conv1_bn_variance = paddle.static.create_parameter(shape=(128,), dtype='float32', name='conv1_bn_variance')
conv1_bn = paddle.nn.functional.batch_norm(x=conv1, weight=conv1_bn_weight, bias=conv1_bn_bias, running_mean=conv1_bn_mean, running_var=conv1_bn_variance, epsilon=9.999999747378752e-06, momentum=0.9990000128746033) conv1_bn = paddle.nn.functional.batch_norm(x=conv1, weight=conv1_bn_weight, bias=conv1_bn_bias, running_mean=conv1_bn_mean, running_var=conv1_bn_variance, epsilon=9.999999747378752e-06, momentum=0.9990000128746033)
conv1_scale_cparam1 = paddle.static.create_parameter(shape=(32,), dtype='float32', name='conv1_scale_cparam1') conv1_scale_cparam1 = paddle.static.create_parameter(shape=(32,), dtype='float32', name='conv1_scale_cparam1')
conv1_scale_mul = paddle.multiply(x=conv1_bn, y=conv1_scale_cparam1, axis=1) conv1_scale_mul = paddle.multiply(x=conv1_bn, y=conv1_scale_cparam1, axis=1)
...@@ -34,6 +36,8 @@ class Static_BNScaleFuser(FuseBase): ...@@ -34,6 +36,8 @@ class Static_BNScaleFuser(FuseBase):
conv1_scale_cparam2 = paddle.reshape(x=conv1_scale_cparam2, shape=[32, 1, 1]) conv1_scale_cparam2 = paddle.reshape(x=conv1_scale_cparam2, shape=[32, 1, 1])
conv1_scale = paddle.add(x=conv1_scale_mul, y=conv1_scale_cparam2) conv1_scale = paddle.add(x=conv1_scale_mul, y=conv1_scale_cparam2)
模式二： 模式二：
conv1_bn_mean = paddle.static.create_parameter(shape=(128,), dtype='float32', name='conv1_bn_mean')
conv1_bn_variance = paddle.static.create_parameter(shape=(128,), dtype='float32', name='conv1_bn_variance')
conv1_bn = paddle.nn.functional.batch_norm(x=conv1, weight=conv1_bn_weight, bias=conv1_bn_bias, running_mean=conv1_bn_mean, running_var=conv1_bn_variance, epsilon=9.999999747378752e-06, momentum=0.9990000128746033) conv1_bn = paddle.nn.functional.batch_norm(x=conv1, weight=conv1_bn_weight, bias=conv1_bn_bias, running_mean=conv1_bn_mean, running_var=conv1_bn_variance, epsilon=9.999999747378752e-06, momentum=0.9990000128746033)
conv1_scale_cparam1 = paddle.static.create_parameter(shape=(32,), dtype='float32', name='conv1_scale_cparam1') conv1_scale_cparam1 = paddle.static.create_parameter(shape=(32,), dtype='float32', name='conv1_scale_cparam1')
conv1_scale_mul = paddle.multiply(x=conv1_bn, y=conv1_scale_cparam1, axis=1) conv1_scale_mul = paddle.multiply(x=conv1_bn, y=conv1_scale_cparam1, axis=1)
...@@ -45,13 +49,21 @@ class Static_BNScaleFuser(FuseBase): ...@@ -45,13 +49,21 @@ class Static_BNScaleFuser(FuseBase):
return "x" + str(id) return "x" + str(id)
pattern = PaddleGraph(graph_type="dygraph") pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(10)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(11)])
pattern.add_layer( pattern.add_layer(
"paddle.nn.functional.batch_norm", "paddle.nn.functional.batch_norm",
inputs={"input": "bn-input-0", inputs={"input": "bn-input-0",
"weight": "bn-input-1", "weight": "bn-input-1",
"bias": "bn-input-2", "bias": "bn-input-2",
"running_mean": "bn-input-3", "running_mean": gen_name(10),
"running_var": "bn-input-4",}, "running_var": gen_name(11)},
outputs=[gen_name(0)]) outputs=[gen_name(0)])
pattern.add_layer( pattern.add_layer(
"paddle.static.create_parameter", "paddle.static.create_parameter",
...@@ -81,19 +93,25 @@ class Static_BNScaleFuser(FuseBase): ...@@ -81,19 +93,25 @@ class Static_BNScaleFuser(FuseBase):
outputs=[gen_name(5)]) outputs=[gen_name(5)])
pattern.build(inputs={"input-0": "bn-input-0", pattern.build(inputs={"input-0": "bn-input-0",
"input-1": "bn-input-1", "input-1": "bn-input-1",
"input-2": "bn-input-2", "input-2": "bn-input-2"})
"input-3": "bn-input-3",
"input-4": "bn-input-4"})
self.patterns.append(pattern) self.patterns.append(pattern)
pattern = PaddleGraph(graph_type="dygraph") pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(10)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(11)])
pattern.add_layer( pattern.add_layer(
"paddle.nn.functional.batch_norm", "paddle.nn.functional.batch_norm",
inputs={"input": "bn-input-0", inputs={"input": "bn-input-0",
"weight": "bn-input-1", "weight": "bn-input-1",
"bias": "bn-input-2", "bias": "bn-input-2",
"running_mean": "bn-input-3", "running_mean": gen_name(10),
"running_var": "bn-input-4",}, "running_var": gen_name(11),},
outputs=[gen_name(0)]) outputs=[gen_name(0)])
pattern.add_layer( pattern.add_layer(
"paddle.static.create_parameter", "paddle.static.create_parameter",
...@@ -119,25 +137,25 @@ class Static_BNScaleFuser(FuseBase): ...@@ -119,25 +137,25 @@ class Static_BNScaleFuser(FuseBase):
outputs=[gen_name(4)]) outputs=[gen_name(4)])
pattern.build(inputs={"input-0": "bn-input-0", pattern.build(inputs={"input-0": "bn-input-0",
"input-1": "bn-input-1", "input-1": "bn-input-1",
"input-2": "bn-input-2", "input-2": "bn-input-2"})
"input-3": "bn-input-3",
"input-4": "bn-input-4"})
self.patterns.append(pattern) self.patterns.append(pattern)
def insert_new_layer(self, graph, parameters, matches): def insert_new_layer(self, graph, parameters, matches):
new_layer = self.gen_new_layer(parameters, matches) new_layer = self.gen_new_layer(parameters, matches)
new_layer_id = list(matches.keys())[-1] new_layer_id = list(matches.keys())[-1]
graph.layers[new_layer_id] = new_layer graph.layers[new_layer_id] = new_layer
matches.pop(list(matches.keys())[0])
matches.pop(list(matches.keys())[0])
matches.pop(list(matches.keys())[1]) matches.pop(list(matches.keys())[1])
matches.pop(list(matches.keys())[2]) matches.pop(list(matches.keys())[2])
matches.pop(new_layer_id) matches.pop(new_layer_id)
def gen_new_layer(self, parameters, matches): def gen_new_layer(self, parameters, matches):
layers_id = list(matches.keys()) layers_id = list(matches.keys())
bn_layer = matches[layers_id[0]] bn_layer = matches[layers_id[2]]
layer = matches[layers_id[1]]
bn_layer.inputs["weight"] = layer.outputs[0]
layer = matches[layers_id[3]] layer = matches[layers_id[3]]
bn_layer.inputs["weight"] = layer.outputs[0]
layer = matches[layers_id[5]]
bn_layer.inputs["bias"] = layer.outputs[0] bn_layer.inputs["bias"] = layer.outputs[0]
bn_layer.id = layers_id[-1] bn_layer.id = layers_id[-1]
layer = matches[layers_id[-1]] layer = matches[layers_id[-1]]
......
x2paddle/optimizer/pattern_matcher.py
浏览文件 @ f08b1f3f
...@@ -99,7 +99,7 @@ class PatternMatcher(object): ...@@ -99,7 +99,7 @@ class PatternMatcher(object):
return False return False
else: else:
subgraph_id2layers.pop(layer_id) subgraph_id2layers.pop(layer_id)
continue continue
else: else:
if len(graph.edges_out[layer_id]) != len( if len(graph.edges_out[layer_id]) != len(
pattern.edges_out[pattern_layer_id]): pattern.edges_out[pattern_layer_id]):
...@@ -116,7 +116,20 @@ class PatternMatcher(object): ...@@ -116,7 +116,20 @@ class PatternMatcher(object):
else: else:
subgraph_id2layers.pop(layer_id) subgraph_id2layers.pop(layer_id)
continue continue
else:
layer_out = graph.edges_out[layer_id]
pattern_layer_out = pattern.edges_out[pattern_layer_id]
is_pop = False
for i in range(len(layer_out)):
layer_id_out = layer_out[i]
pattern_layer_id_out = pattern_layer_out[i]
if layer_id_out != -1:
if graph_layers[layer_id_out].kernel != pattern.layers[pattern_layer_id_out].kernel:
is_pop = True
break
if is_pop:
subgraph_id2layers.pop(layer_id)
continue
# 当为控制流时的处理 # 当为控制流时的处理
if layer.kernel == "prim.if" or layer.kernel == "prim.loop": if layer.kernel == "prim.if" or layer.kernel == "prim.loop":
if len(pattern_layer.blocks) != len(layer.blocks): if len(pattern_layer.blocks) != len(layer.blocks):
...@@ -161,7 +174,7 @@ class PatternMatcher(object): ...@@ -161,7 +174,7 @@ class PatternMatcher(object):
for i, (layer_id, layer) in enumerate(graph.layers.items()): for i, (layer_id, layer) in enumerate(graph.layers.items()):
match_info = get_subgraph(self.pattern, graph, i) match_info = get_subgraph(self.pattern, graph, i)
if match_info: if match_info and match_info not in self.matches:
self.matches.append(match_info) self.matches.append(match_info)
for j, block in enumerate(layer.blocks): for j, block in enumerate(layer.blocks):
if len(block.layers) > 0: if len(block.layers) > 0:
...@@ -343,4 +356,5 @@ class FuseBase(object): ...@@ -343,4 +356,5 @@ class FuseBase(object):
if layer_id in subgraph.layers: if layer_id in subgraph.layers:
# layer_id可能是属于子图的，此时删除父layer，即删除整个子图 # layer_id可能是属于子图的，此时删除父layer，即删除整个子图
subgraph.layers.pop(layer_id) subgraph.layers.pop(layer_id)
\ No newline at end of file
x2paddle/optimizer/code_optimizer/__init__.py → x2paddle/optimizer/pytorch_code_optimizer/__init__.py
浏览文件 @ f08b1f3f
...@@ -13,5 +13,5 @@ ...@@ -13,5 +13,5 @@
# limitations under the License. # limitations under the License.
from x2paddle.optimizer.code_optimizer.hierachical_tree import HierarchicalTree from x2paddle.optimizer.pytorch_code_optimizer.hierachical_tree import HierarchicalTree
from x2paddle.optimizer.code_optimizer.module_graph import ModuleGraph from x2paddle.optimizer.pytorch_code_optimizer.module_graph import ModuleGraph
\ No newline at end of file \ No newline at end of file
x2paddle/optimizer/code_optimizer/hierachical_tree.py → x2paddle/optimizer/pytorch_code_optimizer/hierachical_tree.py
浏览文件 @ f08b1f3f
...@@ -18,10 +18,10 @@ import copy ...@@ -18,10 +18,10 @@ import copy
import os.path as osp import os.path as osp
from treelib import Tree from treelib import Tree
from queue import Queue from queue import Queue
from x2paddle.optimizer.code_optimizer.layer_code_generator import gen_layer_code, rename_layers, NN_KERNEL_WITH_PARAMS, NN_KERNEL_NAME from x2paddle.optimizer.pytorch_code_optimizer.layer_code_generator import gen_layer_code, rename_layers, NN_KERNEL_WITH_PARAMS, NN_KERNEL_NAME
from x2paddle.optimizer.code_optimizer.subgraphs_union import distinguish_sequential, get_inputs_outputs from x2paddle.optimizer.pytorch_code_optimizer.subgraphs_union import distinguish_sequential, get_inputs_outputs
from x2paddle.core.program import PaddleLayer from x2paddle.core.program import PaddleLayer
from x2paddle.optimizer.code_optimizer.parameter_tree import PamareterNode, PamareterTree from x2paddle.optimizer.pytorch_code_optimizer.parameter_tree import PamareterNode, PamareterTree
SEPARATOR_IN_SCOPE = "/" SEPARATOR_IN_SCOPE = "/"
...@@ -39,6 +39,7 @@ class HierarchicalTree(Tree): ...@@ -39,6 +39,7 @@ class HierarchicalTree(Tree):
self.identifier_idx = dict() self.identifier_idx = dict()
self.param_tree = PamareterTree() self.param_tree = PamareterTree()
self.module_name2count = dict() self.module_name2count = dict()
self.scope_name_list = list()
def insert(self, layer): def insert(self, layer):
""" 往层次树中插入节点。 """ 往层次树中插入节点。
...@@ -47,6 +48,7 @@ class HierarchicalTree(Tree): ...@@ -47,6 +48,7 @@ class HierarchicalTree(Tree):
layer (PaddleLayer): 需要插入的节点。 layer (PaddleLayer): 需要插入的节点。
""" """
scope_name = layer.scope_name scope_name = layer.scope_name
self.scope_name_list.append(scope_name)
if scope_name == "": if scope_name == "":
if layer.kernel == "prim.tuple" or layer.kernel == "prim.tuple_unpack": if layer.kernel == "prim.tuple" or layer.kernel == "prim.tuple_unpack":
layer_id = layer.id layer_id = layer.id
...@@ -55,12 +57,36 @@ class HierarchicalTree(Tree): ...@@ -55,12 +57,36 @@ class HierarchicalTree(Tree):
layer_id_list.append(int(input_layer_id)) layer_id_list.append(int(input_layer_id))
layer_id_list = list(set(layer_id_list)) layer_id_list = list(set(layer_id_list))
layer_id_list.sort(reverse=True) layer_id_list.sort(reverse=True)
for input_layer_id in layer_id_list:
input_layer_id_str = str(input_layer_id) if layer.kernel == "prim.tuple":
if self.pd_graph.layers[input_layer_id_str].scope_name != "": for i, input_layer_id in enumerate(layer_id_list):
input_layer_id_str = str(input_layer_id)
scope_name = self.pd_graph.layers[input_layer_id_str].scope_name scope_name = self.pd_graph.layers[input_layer_id_str].scope_name
break if i == 0:
layer.scope_name = scope_name min_scope_name = scope_name
else:
len1 = len(min_scope_name.split("/"))
len2 = len(scope_name.split("/"))
if scope_name not in self.scope_name_list:
min_scope_name = scope_name
continue
if len1 > len2:
min_scope_name = scope_name
if min_scope_name == "":
self.create_node(tag=layer.id,
identifier="no_scope_" + layer.id,
parent=self.pd_graph.name,
data=layer)
return
layer.scope_name = min_scope_name
scope_name = min_scope_name
else:
for input_layer_id in layer_id_list:
input_layer_id_str = str(input_layer_id)
if self.pd_graph.layers[input_layer_id_str].scope_name != "":
scope_name = self.pd_graph.layers[input_layer_id_str].scope_name
break
layer.scope_name = scope_name
else: else:
self.create_node(tag=layer.id, self.create_node(tag=layer.id,
identifier="no_scope_" + layer.id, identifier="no_scope_" + layer.id,
...@@ -369,9 +395,6 @@ class HierarchicalTree(Tree): ...@@ -369,9 +395,6 @@ class HierarchicalTree(Tree):
self.convert_subgraph_to_layer() self.convert_subgraph_to_layer()
self.update_parameters() self.update_parameters()
import_list = ["import paddle", import_list = ["import paddle",
"import paddle.fluid as fluid",
"from paddle.fluid.initializer import Constant",
"from paddle.fluid.param_attr import ParamAttr",
"import math", "import math",
"from x2paddle.op_mapper.dygraph.pytorch2paddle " + \ "from x2paddle.op_mapper.dygraph.pytorch2paddle " + \
"import pytorch_custom_layer as x2paddle_nn" "import pytorch_custom_layer as x2paddle_nn"
......
x2paddle/optimizer/code_optimizer/layer_code_generator.py → x2paddle/optimizer/pytorch_code_optimizer/layer_code_generator.py
浏览文件 @ f08b1f3f
...@@ -14,7 +14,11 @@ ...@@ -14,7 +14,11 @@
# limitations under the License. # limitations under the License.
import copy import copy
from x2paddle.optimizer.code_optimizer.parameter_tree import PamareterNode import os.path as osp
import x2paddle
from x2paddle.optimizer.pytorch_code_optimizer.parameter_tree import PamareterNode
from x2paddle.core.util import *
NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn", NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn",
"paddle.nn.LayerNorm": "layernorm", "paddle.nn.LayerNorm": "layernorm",
...@@ -22,6 +26,7 @@ NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn", ...@@ -22,6 +26,7 @@ NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn",
"paddle.nn.Embedding": "embedding", "paddle.nn.Embedding": "embedding",
"paddle.nn.Linear": "linear", "paddle.nn.Linear": "linear",
"paddle.nn.Conv2DTranspose": "conv", "paddle.nn.Conv2DTranspose": "conv",
"paddle.nn.LSTM": "lstm",
"paddle.nn.ReLU": "relu", "paddle.nn.ReLU": "relu",
"paddle.nn.ReLU6": "relu", "paddle.nn.ReLU6": "relu",
"paddle.nn.Softmax": "softmax", "paddle.nn.Softmax": "softmax",
...@@ -36,7 +41,7 @@ NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn", ...@@ -36,7 +41,7 @@ NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn",
"paddle.nn.GELU": "gelu", "paddle.nn.GELU": "gelu",
"paddle.nn.Hardtanh": "tanh", "paddle.nn.Hardtanh": "tanh",
"paddle.nn.LeakyReLU": "leakly_relu"} "paddle.nn.LeakyReLU": "leakly_relu"}
NN_KERNEL_WITH_PARAMS = list(NN_KERNEL_NAME.keys())[:6] NN_KERNEL_WITH_PARAMS = list(NN_KERNEL_NAME.keys())[:7]
def rename_layers(layers, param_tree=None, is_rename_module=False): def rename_layers(layers, param_tree=None, is_rename_module=False):
""" 对子模块的输入输出等进行重命名。 """ 对子模块的输入输出等进行重命名。
...@@ -125,14 +130,30 @@ def rename_layers(layers, param_tree=None, is_rename_module=False): ...@@ -125,14 +130,30 @@ def rename_layers(layers, param_tree=None, is_rename_module=False):
return layers_cp, nn_param_nodes, new_names return layers_cp, nn_param_nodes, new_names
def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()): def _update_attrs(layer, different_attrs):
if "module" in layer.kernel or "prim" in layer.kernel:
return
common_attrs = copy.deepcopy(layer.attrs)
special_attrs = dict()
for k, v in layer.attrs.items():
if len(layer.outputs) < 1:
break
key_name = "{}_{}".format(layer.outputs[0], k)
if key_name in different_attrs:
common_attrs.pop(k)
special_attrs[k] = v
remove_default_attrs(layer.kernel, common_attrs)
common_attrs.update(special_attrs)
layer.attrs = common_attrs
def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=dict()):
""" 根据sub_layers生成对应的Module代码。 """ 根据sub_layers生成对应的Module代码。
Args: Args:
graph (x2paddle.core.program.PaddleGraph): 整个Paddle图。 graph (x2paddle.core.program.PaddleGraph): 整个Paddle图。
sub_layers (dict): 子图的id和其对应layer组成的字典。 sub_layers (dict): 子图的id和其对应layer组成的字典。
sub_layers_name (str): 子图的名字。 sub_layers_name (str): 子图的名字。
different_attrs (list): 属性列表，这些属性表明在被调用时赋予不同值。 different_attrs (dict/list): 属性字典/列表，这些属性表明在被调用时赋予不同值。
""" """
def gen_codes(code_list, indent=0): def gen_codes(code_list, indent=0):
""" 根据code_list生成代码段。 """ 根据code_list生成代码段。
...@@ -157,7 +178,13 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()): ...@@ -157,7 +178,13 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
# 生成Layer的头部代码 # 生成Layer的头部代码
head = gen_codes(["class {}(paddle.nn.Layer):".format(sub_layers_name)], indent=0) head = gen_codes(["class {}(paddle.nn.Layer):".format(sub_layers_name)], indent=0)
# 生成init函数的头部代码 # 生成init函数的头部代码
attrs_str = ", ".join(different_attrs) diff_str_list = list()
if isinstance(different_attrs, dict):
for k, v in different_attrs.items():
diff_str_list.append("{}={}".format(k, v))
attrs_str = ", ".join(diff_str_list)
else:
attrs_str = ", ".join(different_attrs)
init_func_head = \ init_func_head = \
gen_codes(["def __init__(self, {}):".format(attrs_str)], indent=1) + \ gen_codes(["def __init__(self, {}):".format(attrs_str)], indent=1) + \
gen_codes(["super({}, self).__init__()".format(sub_layers_name)], indent=2) gen_codes(["super({}, self).__init__()".format(sub_layers_name)], indent=2)
...@@ -213,6 +240,7 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()): ...@@ -213,6 +240,7 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
outputs.append(layer.outputs[0]) outputs.append(layer.outputs[0])
no_output_count = 0 no_output_count = 0
for i, (layer_id, layer) in enumerate(sub_layers.items()): for i, (layer_id, layer) in enumerate(sub_layers.items()):
_update_attrs(layer, different_attrs)
if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel) or \ if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel) or \
layer.kernel.startswith("custom_layer"): layer.kernel.startswith("custom_layer"):
line = "self.{}".format(layer.outputs[0]) line = "self.{}".format(layer.outputs[0])
...@@ -235,7 +263,10 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()): ...@@ -235,7 +263,10 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
elif len(layer.outputs) == 2: elif len(layer.outputs) == 2:
line = layer.outputs[1] line = layer.outputs[1]
else: else:
line = ','.join(layer.outputs[1:]) if layer.kernel == "paddle.nn.LSTM":
line = "{}, ({})".format(layer.outputs[1], ', '.join(layer.outputs[-2:]))
else:
line = ','.join(layer.outputs[1:])
line += " = self.{}(".format(layer.outputs[0]) line += " = self.{}(".format(layer.outputs[0])
for k, v in layer.inputs.items(): for k, v in layer.inputs.items():
...@@ -263,7 +294,7 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()): ...@@ -263,7 +294,7 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
init_func=init_func, init_func=init_func,
forward_func=forward_func, forward_func=forward_func,
layer_id=layer_id, layer_id=layer_id,
different_attrs=different_attrs) different_attrs=list(different_attrs.keys()) if isinstance(different_attrs, dict) else different_attrs)
cur_outputs.extend(layer.outputs) cur_outputs.extend(layer.outputs)
else: else:
raise Exception( raise Exception(
......
x2paddle/optimizer/code_optimizer/module_graph.py → x2paddle/optimizer/pytorch_code_optimizer/module_graph.py
浏览文件 @ f08b1f3f
...@@ -17,9 +17,9 @@ import copy ...@@ -17,9 +17,9 @@ import copy
import os import os
import os.path as osp import os.path as osp
from x2paddle.core.program import PaddleLayer from x2paddle.core.program import PaddleLayer
from x2paddle.optimizer.code_optimizer.subgraphs_union import construct_attrs_table, get_inputs_outputs from x2paddle.optimizer.pytorch_code_optimizer.subgraphs_union import construct_attrs_table, get_inputs_outputs
from x2paddle.optimizer.code_optimizer.layer_code_generator import gen_layer_code, rename_layers from x2paddle.optimizer.pytorch_code_optimizer.layer_code_generator import gen_layer_code, rename_layers
from x2paddle.optimizer.code_optimizer.parameter_tree import PamareterNode, PamareterTree from x2paddle.optimizer.pytorch_code_optimizer.parameter_tree import PamareterNode, PamareterTree
NoModuleStart = ["paddle.nn.ReLU"] NoModuleStart = ["paddle.nn.ReLU"]
...@@ -179,16 +179,27 @@ class ModuleGraph(object): ...@@ -179,16 +179,27 @@ class ModuleGraph(object):
def analyze_attrs_table(self, attrs_table): def analyze_attrs_table(self, attrs_table):
""" 分析属性表格，哪些属性取值不一致。 """ 分析属性表格，哪些属性取值不一致。
""" """
diff_attrs_column = list() diff_attrs_column = dict()
for column in list(attrs_table.columns): for column in list(attrs_table.columns):
elements = list(attrs_table.get(column)) elements = list(attrs_table.get(column))
base = elements[0] elements_list = list()
for element in elements[1:]: count_list = list()
if isinstance(base, str) and "'" not in base: for element in elements:
break if isinstance(element, str) and "'" not in element:
if element != base:
diff_attrs_column.append(column)
break break
if element not in elements_list:
count_list.append(1)
elements_list.append(element)
else:
index = elements_list.index(element)
count_list[index] += 1
if len(elements_list) > 1:
max_ct = 0
for k, v in zip(elements_list, count_list):
if v > max_ct and str(k) != "nan" :
max_ele = k
max_ct = v
diff_attrs_column[column] = max_ele
return diff_attrs_column return diff_attrs_column
def analyze_graph(self, sub_layers_list): def analyze_graph(self, sub_layers_list):
...@@ -258,8 +269,10 @@ class ModuleGraph(object): ...@@ -258,8 +269,10 @@ class ModuleGraph(object):
outputs = ["{}_{}".format(mn, index)] + outputs outputs = ["{}_{}".format(mn, index)] + outputs
node_name = "{}_{}".format(module_name, index) node_name = "{}_{}".format(module_name, index)
diff_attrs = dict() diff_attrs = dict()
for column in diff_attrs_column: for column, element in diff_attrs_column.items():
diff_attrs[column] = attrs_table.get(column).loc[node_name] current_element = attrs_table.get(column).loc[node_name]
if current_element != element:
diff_attrs[column] = current_element
new_layer = PaddleLayer(id=list(sub_layers.keys())[-1], new_layer = PaddleLayer(id=list(sub_layers.keys())[-1],
kernel="module", kernel="module",
inputs=inputs_dict, inputs=inputs_dict,
...@@ -352,9 +365,6 @@ class ModuleGraph(object): ...@@ -352,9 +365,6 @@ class ModuleGraph(object):
self.convert_subgraph_to_layer(combination, combination_id) self.convert_subgraph_to_layer(combination, combination_id)
self.update_parameters() self.update_parameters()
import_list = ["import paddle", import_list = ["import paddle",
"import paddle.fluid as fluid",
"from paddle.fluid.initializer import Constant",
"from paddle.fluid.param_attr import ParamAttr",
"import math", "import math",
"from x2paddle.op_mapper.dygraph.pytorch2paddle " + \ "from x2paddle.op_mapper.dygraph.pytorch2paddle " + \
"import pytorch_custom_layer as x2paddle_nn" "import pytorch_custom_layer as x2paddle_nn"
......
x2paddle/optimizer/code_optimizer/subgraphs_union.py → x2paddle/optimizer/pytorch_code_optimizer/subgraphs_union.py
浏览文件 @ f08b1f3f
...@@ -16,7 +16,7 @@ ...@@ -16,7 +16,7 @@
import copy import copy
import pandas as pd import pandas as pd
from x2paddle.optimizer.code_optimizer.layer_code_generator import rename_layers from x2paddle.optimizer.pytorch_code_optimizer.layer_code_generator import rename_layers
def construct_attrs_table(sub_layers_list, node_name2sub_layers=None, module_name=None): def construct_attrs_table(sub_layers_list, node_name2sub_layers=None, module_name=None):
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或