将 CNN Pytorch 中的预训练权重传递给 Tensorflow 中的 CNN

Question

I have trained this network in Pytorch for 224x224 size images and 4 classes.

class CustomConvNet(nn.Module):
    def __init__(self, num_classes):
        super(CustomConvNet, self).__init__()

        self.layer1 = self.conv_module(3, 64)
        self.layer2 = self.conv_module(64, 128)
        self.layer3 = self.conv_module(128, 256)
        self.layer4 = self.conv_module(256, 256)
        self.layer5 = self.conv_module(256, 512)
        self.gap = self.global_avg_pool(512, num_classes)
        #self.linear = nn.Linear(512, num_classes)
        #self.relu = nn.ReLU()
        #self.softmax = nn.Softmax()

    def forward(self, x):
        out = self.layer1(x)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = self.layer5(out)
        out = self.gap(out)
        out = out.view(-1, 4)
        #out = self.linear(out)

        return out

    def conv_module(self, in_num, out_num):
        return nn.Sequential(
            nn.Conv2d(in_num, out_num, kernel_size=3, stride=1, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=(2, 2), stride=None))

    def global_avg_pool(self, in_num, out_num):
        return nn.Sequential(
            nn.Conv2d(in_num, out_num, kernel_size=3, stride=1, padding=1),
            #nn.BatchNorm2d(out_num),
            #nn.LeakyReLU(),

            nn.ReLU(),
            nn.Softmax(),
            nn.AdaptiveAvgPool2d((1, 1)))

I got the weights from the first Conv2D and it's size torch.Size([64, 3, 3, 3])

I have saved it as:

weightsCNN = net.layer1[0].weight.data
np.save('CNNweights.npy', weightsCNN)

This is my model I built in Tensorflow. I would like to pass those weights I saved from the Pytorch model into this Tensorflow CNN.

    model = models.Sequential()
    model.add(layers.Conv2D(64, (3, 3), activation='relu', input_shape=(224, 224, 3)))
    model.add(layers.MaxPooling2D((2, 2)))

    model.add(layers.Conv2D(128, (3, 3), activation='relu'))
    model.add(layers.MaxPooling2D((2, 2)))

    model.add(layers.Conv2D(256, (3, 3), activation='relu'))
    model.add(layers.MaxPooling2D((2, 2)))

    model.add(layers.Conv2D(256, (3, 3), activation='relu'))
    model.add(layers.MaxPooling2D((2, 2)))

    model.add(layers.Conv2D(512, (3, 3), activation='relu'))
    model.add(layers.MaxPooling2D((2, 2)))

    model.add(layers.Conv2D(512, (3, 3), activation='relu'))

    model.add(layers.GlobalAveragePooling2D())
    model.add(layers.Dense(4, activation='softmax'))
    print(model.summary())


    adam = optimizers.Adam(learning_rate=0.0001, amsgrad=False)
    model.compile(loss='categorical_crossentropy',
                  optimizer=adam,
                  metrics=['accuracy'])


    nb_train_samples = 6596
    nb_validation_samples = 1290
    epochs = 10
    batch_size = 256


    history = model.fit_generator(
        train_generator,
        steps_per_epoch=np.ceil(nb_train_samples/batch_size),
        epochs=epochs,
        validation_data=validation_generator,
        validation_steps=np.ceil(nb_validation_samples / batch_size)
        )

How should I actually do that? What shape of weights does Tensorflow require? Thanks!

Answer 1

You can check shapes of all weights of all keras layers quite simply:

for layer in model.layers:
    print([tensor.shape for tensor in layer.get_weights()])

This would give you shapes of all weights (including biases), so you can prepare loaded numpy weights accordingly.

To set them, do something similar:

for torch_weight, layer in zip(model.layers, torch_weights):
    layer.set_weights(torch_weight)

where torch_weights should be a list containing lists of np.array which you would have to load.

Usually each element of torch_weights would contain one np.array for weights and one for bias.

Remember shapes received from print have to be exactly the same as the ones you put in set_weights .

See documentation for more info.

BTW. Exact shapes are dependent on layers and operations performed by model, you may have to transpose some arrays sometimes to "fit them in".