model.trainable = 错误 - 权重冻结且无法训练？

Question

I am using EfficientNet model ( https://keras.io/api/applications/efficientnet/#efficientnetb0-function ) with weights from ImageNet and I want to use a customized top, so I stated top = False . I am now wondering if the weights of the EfficientNet are frozen and they are not getting retrained (that is what I want) when I use the following code:

efnB0_model = efn.EfficientNetB0(include_top=False, weights="imagenet", input_shape=(224, 224, 3))
efnB0_model.trainable = False

Or do I have to use another code?

Thanks a lot!

Answer 1

What you did works, but people generally do it layer by layer instead, because you might eventually decide to unfreeze certain layers:

for layer in model.layers:
    layer.trainable = False

model.layers returns a list, so you can also unfreeze just the last few layers:

for layer in model.layers[-10:]:
    layer.trainable = False

You can verify what can be trained with

model.trainable_variables

[]

In this case, nothing.