[Over 17] Pytorch rewrites keras

任务

pytorchBuild your own modelkeras

想法

1. 有pytorchself-built model code
   • can be found line by linekerasreplace it with the corresponding layer（手工转写
   • Pytorch转keras的有效方法,以FlowNet为例_咆哮的阿杰的博客-CSDN博客_pytorch转keras
   • tensorflow和keras_手动 pytorch 模型转 tensorflow keras 模型_weixin_39637920的博客-CSDN博客
2. There are self-built models exportedonnx,作为中间件
   • 可以onnx转keras（自动转
   • 搞不出来

查到的博客

首先,We must have a clear understanding,网上以及githubSome of the so-called abovepytorch转换Keras或者Keras转换成PytorchThe tool code is almost inoperable or has limitations of use（For example, only certain models can be converted）,But we can see some clues in these conversion codes,For example, the size of the parameters of the two（shape）的形式、channel的排序（first or last）是否一样,Master the difference,You can write your own conversion code based on these differences,没错,Write the conversion code yourself,是最稳妥的办法.The whole process is divided into two parts.The author will use Nvidia开源的FlowNet为例,将开源的Pytorch代码转化为Keras模型.

按照Pytorch中模型的结构,编写对应的Keras代码,用keras的函数式API,It would be very convenient to build.
把Pytorch的模型参数,Assigned to in order according to the name of the layerKeras的模型
Although the above two steps seem simple,But in reality, I also took a lot of detours.A key point here,就是参数的shapeIs it unified in the two frameworks,That is of course not uniform.

查到的博客
pytorch 到 tensorflow 可以用onnxConvert as an intermediate tool,将pytorch转为onnx,再从onnx转为tensorflow,But there may be some messy problems in the middle.In fact, it is also very convenient to manually read the parameters and refill the corresponding model,This article summarizes manual model conversion.

过程

I would definitely try it first 自动转

pytorch模型转keras模型_AI算法-图哥的博客-CSDN博客_pytorch2keras

ONNX系列二 — 使用ONNX使KerasModels are portable_Igloo's blog-CSDN博客_keras onnx

安装一下

接触到了(load和load_state_dict)的问题

Pytorch保存和加载模型(load和load_state_dict)_木盏的博客-CSDN博客_pytorch 加载模型

插曲

安装whl

可以通过查看pythonThe version corresponds to the installationpythonversion of the offline installation package（附.whlInstall web link）_Az_z的博客-CSDN博客_Install the specified version offlinepython

真是服了,onnx1.8.1It was later removedoptimizer那个包,然后pytorch2kerasThis package uses the version is1.8.1及之前的,onnx1.8.1支持py3.8,然后我Py3.9.So had to be in another virtual environment(py3.6)Reinstalled the corresponding package inside（400M的tensorflow还有torch等等）.安好了,Thought it could be used now

然后报 modulelist的错

Then I export as onnx,再load,就报“TypeError: ‘ModelProto’ object is not callable”的错

我真是服了.

It seems that it can only be written by hand

python关于onnx模型的一些基本操作_A glass of salt water blog-CSDN博客_onnx 静态量化

学习keras,主要是 The parameters on both sides are different

It's better to change.（不问对错,就是搭积木）

Read some blogs

结果

input_data = keras.layers.Input(shape=(4,1), dtype='float64')
cnn_out = keras.layers.Conv1D(1, kernel_size=2, strides=1)(input_data)
cnn_out = keras.layers.MaxPool1D(2)(cnn_out)
lstm_out =keras.layers.LSTM(4)(cnn_out)
dense_out = keras.layers.Dense(3,activation='softmax')(lstm_out)
model = keras.Model(input_data, dense_out)

model.compile(optimizer='adam',loss='mean_absolute_error',metrics=['accuracy'])

model.summary()

The rewrite was successful,But what about the accuracy,very rubbish