将CNN Pytorch中的预训练权重传递给Tensorflow中的CNN



我已经在 Pytorch 中训练了这个网络,用于 224x224 大小的图像和 4 个类。

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)))

我从第一个 Conv2D 中获得了重量,它的大小torch.Size([64, 3, 3, 3])

我已将其保存为:

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

这是我在Tensorflow中构建的模型。我想把我从 Pytorch 模型中保存的权重传递给这个 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)
)

我应该怎么做?Tensorflow需要什么形状的权重?谢谢!

您可以非常简单地检查所有keras图层的所有权重的形状:

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

这将为您提供所有权重(包括偏差)的形状,因此您可以相应地准备加载numpy权重。

要设置它们,请执行类似的操作:

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

其中torch_weights应该是一个列表,其中包含您必须加载的np.array列表。

通常,torch_weights的每个元素都会包含一个权重np.array和一个偏差。

请记住,从打印中收到的形状必须与您放入set_weights的形状完全相同

有关详细信息,请参阅文档。

顺便说一句。确切的形状取决于模型执行的图层和操作,有时可能需要转置某些数组以"适合它们"。

最新更新