In my python script, I've created a class which, among others, contains keras models like so:
from keras.layers import Input, Activation, Dense
from keras.models import Model
class Klass:
def __init__(self, input_dims, output_dims, hidden_dims, optimizer, a, b):
self.input_dims = input_dims
self.output_dims = output_dims
self.hidden_dims = hidden_dims
self.optimizer = optimizer
self.a = a
self.b = b
self.__build_nn()
def __build_nn(self):
inputs = Input(shape=(self.input_dims,))
net = inputs
for h_dim in self.hidden_dims:
net = Dense(h_dim, kernel_initializer='he_uniform')(net)
net = Activation("relu")(net)
outputs = Dense(self.output_dims)(net)
outputs = Activation("linear")(outputs)
self.nn1 = Model(inputs=inputs, outputs=outputs)
self.nn2 = Model(inputs=inputs, outputs=outputs)
self.nn1.compile(optimizer=self.optimizer, loss='mean_squared_error')
self.nn2.compile(optimizer=self.optimizer, loss='mean_squared_error')
After creating a Klass instance, I would like to make a deep copy of it:
import copy
obj = Klass(10, 10, (20, 20), Adam(), 1, 2)
obj_dc = copy.deepcopy(obj)
However, this throws a TypeError: can't pickle _thread.RLock objects. I am pretty sure that the error is related to the keras models in the class object since I was able to get a deep copy of a similar class without the keras models.
Unfortunately, I was unable to find a solution to this on the internet, since most questions concerning deep copying a keras model were trying to clone a keras model like here.
So, how can I get a deep copy of a class containing the keras models?
EDIT
These three questions (1, 2, 3) mention a similar error under different circumstances. Yet, the solutions offered there do not apply in my case.
EDIT 2
As suggested in the comments, I've added a copy method in the class. Would that be a viable solution?
class Klass:
def __init__(self, input_dims, output_dims, hidden_dims, optimizer, a, b):
self.input_dims = input_dims
self.output_dims = output_dims
self.hidden_dims = hidden_dims
self.optimizer = optimizer
self.a = a
self.b = b
self.__build_nn()
# [...]
def copy(self):
new = Klass(self.input_dims, self.output_dims, self.hidden_dims,
self.optimizer, self.a, self.b)
new.nn1.set_weights(self.nn1.get_weights())
new.nn2.set_weights(self.nn2.get_weights())
return new
