layer.trainable example not working

[ghost]

Please make sure that the boxes below are checked before you submit your issue. Thank you!

• Check that you are up-to-date with the master branch of Keras. You can update with:
  pip install git+git://github.com/fchollet/keras.git --upgrade --no-deps
• If running on Theano, check that you are up-to-date with the master branch of Theano. You can update with:
  pip install git+git://github.com/Theano/Theano.git --upgrade --no-deps
• Provide a link to a GitHub Gist of a Python script that can reproduce your issue (or just copy the script here if it is short).

The example in https://keras.io/getting-started/faq/#how-can-i-freeze-keras-layers does not work.
The example fails with both TF and theano.
Both models will do whatever you put in the second layer.trainable line.

Code to try:

from keras.layers import Input, Dense, Dropout, Activation, Flatten
from keras.layers import Convolution2D, MaxPooling2D, MaxPooling2D
from keras.layers import Reshape, Deconvolution2D
from keras.models import Model
import numpy as np

x = Input(shape=(32,))
layer = Dense(32)
layer.trainable = False
y = layer(x)
frozen_model = Model(x, y)
frozen_model.compile(optimizer='rmsprop', loss='mse')
#**** both models do whatever you put in the next line
layer.trainable = True
trainable_model = Model(x, y)
trainable_model.compile(optimizer='rmsprop', loss='mse')

data = labels = np.vstack((np.zeros((500,32)), np.ones((500,32))))
print layer.get_weights()[0][0]
frozen_model.fit(data, labels)
print layer.get_weights()[0][0]
trainable_model.fit(data, labels)
print layer.get_weights()[0][0]

[kuza55]

The docs do point out this behaviour:

# with this model the weights of the layer will be updated during training
# (which will also affect the above model since it uses the same layer instance)

Though we probably want to update the docs since the lines at the bottom imply the opposite:

frozen_model.fit(data, labels)  # this does NOT update the weights of `layer`
trainable_model.fit(data, labels)  # this updates the weights of `layer`

I don't see a trivial solution for doing what you want, so I would suggest trying to avoid using this pattern and just setting the flag to be what you need before training the model.

[ghost]

The docs are ok.

The lines you point out mean that, since they share a layer, training the non-frozen model will also change the weights of the frozen one. But training the frozen one should not change anything. This feature used to work properly in the past.

[fchollet]

That's a good point; the thing is that compile does not actually create the gradient updates in the current Keras version (it used to), it happens at the first call to fit. I have fixed this, so now the example in the docs does what it says it does. Update Keras.

[ghost]

Oh, I see. Great, thanks!

[ylmeng]

I load a pre-trained model as a layer in another model. Everything is fine is I set the pre-trained model trainable=False. However, if I set it True, I got a hundred lines of errors. The last chunk is this:
NotFoundError (see above for traceback): Resource __per_step_6/_tensor_arraysoutput_ta_965/N10tensorflow11TensorArrayE does not exist. [[Node: training/RMSprop/gradients/model_2/time_distributed_7/while_2/MatMul_grad/MatMul_1/StackPop/_1899 = _Send[T=DT_FLOAT, client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/gpu:0", send_device_incarnation=1, tensor_name="edge_11685_training/RMSprop/gradients/model_2/time_distributed_7/while_2/MatMul_grad/MatMul_1/StackPop", _device="/job:localhost/replica:0/task:0/gpu:0"](training/RMSprop/gradients/model_2/time_distributed_7/while_2/MatMul_grad/MatMul_1/StackPop/_1898)]]

It seems something is missing in order to make the combined system trainable?
My last layer is a TimeDistributed(Dense). Does that cause the issue? The error message seems to suggest so.

[fatih-ilhan]

I have the same problem while loading a pre-trained model checkpoint:

[[Node: training/Adam/gradients/time_distributed_2/while_1/TensorArrayStack/TensorArrayGatherV3_grad/TensorArrayGrad/TensorArrayGradV3 = TensorArrayGradV3[_class=["loc:@time_distributed_2/while_1/TensorArray_2"], source="training/Adam/gradients", _device="/job:localhost/replica:0/task:0/device:CPU:0"](training/Adam/gradients/time_distributed_2/while_1/TensorArrayStack/TensorArrayGatherV3_grad/TensorArrayGrad/TensorArrayGradV3/StackPopV2, training/Adam/gradients/time_distributed_2/while_1/TensorArrayStack/TensorArrayGatherV3_grad/TensorArrayGrad/TensorArrayGradV3/StackPopV2_1)]]
[[Node: training/Adam/gradients/StackPushV2_6/_868 = _Recvclient_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_2377_training/Adam/gradients/StackPushV2_6", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]]

Could you fix it?