如何将pytorch 3d cnn代码转换为keras

Question

我在pytorch有一个3D CNN网络，我曾试图将其转换为keras，但我不太确定转换是否正确。此外，当我运行keras代码时，出现以下错误：

ValueError: Negative dimension size caused by subtracting 3 from 2 for 'conv3d_13/convolution' (op: 'Conv3D) with input shapes [?,10,2,2,512],[3,3,3,512,512]

Pytorch代码：

class semi_C3D(nn.Module):


    def __init__(self, num_classes = 101):
        super(semi_C3D, self).__init__()

        self.features = nn.Sequential(
            nn.Conv3d(3, 64, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.Conv3d(64, 64, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=(1,2,2), stride=(1,2,2)),

            nn.Conv3d(64, 128, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.Conv3d(128, 128, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=(1,2,2), stride=(1,2,2)),

            nn.Conv3d(128, 256, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.Conv3d(256, 256, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.Conv3d(256, 256, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=(1,2,2), stride=(1,2,2)),

            nn.Conv3d(256, 512, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.Conv3d(512, 512, kernel_size=(1,3,3), padding=(0,1,1)),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=(1,2,2), stride=(1,2,2)),

        )


        self.conv3d = nn.Sequential(
            nn.Conv3d(512, 512, kernel_size=(3,1,1), padding=(1,0,0)),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=(3,1,1), stride=(2,1,1), padding=(1,0,0)),

            nn.Conv3d(512, 512, kernel_size=(3,3,3), padding=(1,1,1)),
            nn.ReLU(),
            nn.Conv3d(512, 512, kernel_size=(3,3,3), padding=(1,1,1)),
            nn.ReLU(),
            nn.MaxPool3d(kernel_size=(2,2,2), stride=(2,2,2)),

            nn.Conv3d(512, 512, kernel_size=(3,3,3), padding=(1,1,1)),
            nn.ReLU(),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)),

        )


        self.st_classifier = nn.Sequential(
            nn.Linear(9216, 4096), 
            nn.ReLU(),
            nn.Dropout(p=0.5),
            nn.Linear(4096, 4096), 
            nn.ReLU(),
            nn.Dropout(p=0.5),
            nn.Linear(4096, num_classes),
        )

    def forward(self, input):

        x = self.features(input)
        x = self.conv3d(x)

        x = x.view(-1, 9216)
        x = self.st_classifier(x)
        return x

Keras代码：

def semi_3d(self):

        model = Sequential()
    #First bloc of layers (self.features = nn.Sequential part for the pytorch code)
        model.add(Conv3D(64,kernel_size=(1, 3, 3),activation='relu',
                 input_shape=self.input_shape))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(Conv3D(64,kernel_size=(1, 3, 3),activation='relu',))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1,2,2)))

        model.add(Conv3D(128,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(Conv3D(128,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1,2,2)))

        model.add(Conv3D(256,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(Conv3D(256,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(Conv3D(256,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1,2,2)))

        model.add(Conv3D(512,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(Conv3D(512,kernel_size=(1, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(0, 1, 1)))
        model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1,2,2)))

    #Second bloc of layers (self.conv3d = nn.Sequential part for the pytorch code)
        model.add(Conv3D(512,kernel_size=(3, 1, 1),activation='relu'))
        model.add(ZeroPadding3D(padding=(1, 0, 0)))
        model.add(MaxPooling3D(pool_size=(3, 1, 1), strides=(2, 1, 1)))
        model.add(ZeroPadding3D(padding=(1, 0, 0)))

        model.add(Conv3D(512,kernel_size=(3, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(1, 1, 1)))
        model.add(Conv3D(512,kernel_size=(3, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(1, 1, 1)))
        model.add(MaxPooling3D(pool_size=(2, 2, 2), strides=(2, 2, 2)))

        model.add(Conv3D(512,kernel_size=(3, 3, 3),activation='relu'))
        model.add(ZeroPadding3D(padding=(1, 1, 1)))
        model.add(MaxPooling3D(pool_size=(2, 2, 2), strides=(2, 2, 2)))

    #FC Layers
        model.add(Dense(4096, activation='softmax', name='fc1'))
        model.add(Dropout(0.5))
        model.add(Dense(4096, activation='softmax', name='fc2'))
        model.add(Dropout(0.5))
        model.add(Dense(self.nb_classes, activation='softmax', name='fc3'))
        return model

Answer 1

您可以根据文档的说明使用onnx尝试此操作“ONNX是一种开放格式，用于表示机器学习模型。ONNX定义了一组通用的操作符（机器学习和深度学习模型的构建块）和一种通用的文件格式，使AI开发人员能够将模型与各种框架、工具、运行时和编译器一起使用。”

源代码在这里Open standard for machine learning interoperability 你会像下面这样做

import onnx
from keras.models import load_model

pytorch_model = '/path/to/pytorch/model'
keras_output = '/path/to/converted/keras/model.hdf5'
onnx.convert(pytorch_model, keras_output)
model = load_model(keras_output)
preds = model.predict(x)