Why is Python eval returning same object for Keras regularizer?

Question

I am trying to convert strings (which I read from a JSON) to arguments that can be used by Keras layers. However when I find that all the regularizer objects created by the eval function are the same.

a = eval('l1(0.1)')
b = eval('l2(0.1)')
c = eval('l1_l2(0.1)')
print(a,b,c)

gives:

<tensorflow.python.keras.regularizers.L1L2 object at 0x0000013F003C2F60>
<tensorflow.python.keras.regularizers.L1L2 object at 0x0000013F003C2D68> 
<tensorflow.python.keras.regularizers.L1L2 object at 0x0000013F0032F160>

Shouldn't eval('l1(0.1)') give

<function tensorflow.python.keras.regularizers.l1(l=0.01)>

Any thoughts on why this is happening would be much appreciated.

Answer 1

L1L2 stores both l1 and l2; on the regularizer, run, for example:

print(model.layers[1].kernel_regularizer.__dict__)
# {'l1': array(0., dtype=float32), 'l2': array(1., dtype=float32)}

To access one or the other:

print(model.layers[1].kernel_regularizer.l1) # 0.0
print(model.layers[1].kernel_regularizer.l2) # 1.0

In your code, a sets l1, b sets l2, and c sets both.

---

Full sample script:

from keras.layers import Input, Dense
from keras.regularizers import l2
from keras.models import Model

ipt   = Input(shape=(100,4))
x     = Dense(10, activation='relu', kernel_regularizer=l2(1))(ipt)
out   = Dense(1, activation='sigmoid')(x)
model = Model(ipt, out)
model.compile(optimizer='adam', loss='binary_crossentropy')


print(model.layers[1].kernel_regularizer.__dict__)
print(model.layers[1].kernel_regularizer.l1)
print(model.layers[1].kernel_regularizer.l2)