I made a small example, which basically does the same as your code, and I show that the expected results are obtained:
Imports
In [1]: from keras.models import Sequential
In [2]: from keras.layers import Dense
Build a model
Note the name
arguments:
In [3]: model = Sequential()
In [4]: model.add(Dense(50, input_shape=(20, 20), name='dense1'))
In [5]: model.add(Dense(30, name='dense2'))
In [6]: model.add(Dense(1, name='dense3'))
In [7]: model.compile('rmsprop', 'mse')
In [8]: model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense1 (Dense) (None, 20, 50) 1050
_________________________________________________________________
dense2 (Dense) (None, 20, 30) 1530
_________________________________________________________________
dense3 (Dense) (None, 20, 1) 31
=================================================================
Rename the layers
EDIT: it seems newer versions of Keras and the tf.keras
API now do not allow renaming layers via layer.name = "new_name"
. Instead you must assign your new name to the private attribute, layer._name
.
In [9]: for i, layer in enumerate(model.layers):
...: # layer.name = 'layer_' + str(i) <-- old way
...: layer._name = 'layer_' + str(i)
In [10]: model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
layer_0 (Dense) (None, 20, 50) 1050
_________________________________________________________________
layer_1 (Dense) (None, 20, 30) 1530
_________________________________________________________________
layer_2 (Dense) (None, 20, 1) 31
=================================================================
So we see the layers now have the new names!
Further inspection
I can see that the internal identifiers of the layers themselves are not changed. I show the first layer's weights, as we see the original label is still present:
In [11]: a = model.layers[0]
In [12]: a.weights
Out[12]:
[<tf.Variable 'dense1/kernel:0' shape=(20, 50) dtype=float32_ref>,
<tf.Variable 'dense1/bias:0' shape=(50,) dtype=float32_ref>]
We see the dense1
name is still visible in the layer container, along with its unchanged shape and datatype.
I also initialised the model weights by running model.get_weights()
, then renamed the layers as before to new names, compiled the model and saw that the layers' names were altered as before, and the weights themselves were not changed i.e. reinitialised. The exact same values were contained. This means that renaming your layers cannot be behind the drop in accuracy.
name
attribute of a layer should not affect the accuracy of a model. They are simply descriptors. $\endgroup$