How model.trainable = False works in keras (GAN model)

GAN-and-trainable.py

# coding: utf8

## based on this article: http://qiita.com/mokemokechicken/items/937a82cfdc31e9a6ca12

import numpy as np
from keras.models import Sequential
from keras.engine.topology import Input, Container
from keras.engine.training import Model
from keras.layers.core import Dense

def all_weights(m):
    return [list(w.reshape((-1))) for w in m.get_weights()]

def random_fit(m):
    x1 = np.random.random(10).reshape((5, 2))
    y1 = np.random.random(10).reshape((5, 2))
    m.fit(x1, y1, verbose=False)

np.random.seed(100)

# Discriminator model
x = in_x = Input((2, ))
x = Dense(1)(x)
x = Dense(2)(x)
model_D = Model(in_x, x)

# Compile D
model_D.compile(optimizer="sgd", loss="mse")

# Generator model
x = in_x = Input((2, ))
x = Dense(1)(x)
x = Dense(2)(x)
model_G = Model(in_x, x)

# Adversarial model
model_A = Sequential()
model_A.add(model_G)
model_A.add(model_D)

# Compile A 
model_D.trainable = False # set D in A "trainable=False"
model_A.compile(optimizer="sgd", loss="mse")

# Watch which weights are updated by model.fit
print("Initial Weights")
print("G: %s" % all_weights(model_G))
print("D: %s" % all_weights(model_D))
print("A : %s" % all_weights(model_A))

random_fit(model_D)

print 
print("after training D --- D and D in A changed")
print("G: %s" % all_weights(model_G))
print("D: %s" % all_weights(model_D))
print("A : %s" % all_weights(model_A))

random_fit(model_A)

print 
print("after training A --- D didn't changed!")
print("G: %s" % all_weights(model_G))
print("D: %s" % all_weights(model_D))
print("A : %s" % all_weights(model_A))


random_fit(model_D)
print 
print("after training D")
print("G: %s" % all_weights(model_G))
print("D: %s" % all_weights(model_D))
print("A : %s" % all_weights(model_A))

random_fit(model_A)

print
print("after training A")
print("G: %s" % all_weights(model_G))
print("D: %s" % all_weights(model_D))
print("A : %s" % all_weights(model_A))



# Initial Weights
# G: [[-0.27850878, -0.52411258], [0.0], [0.94569027, 0.83747566], [0.0, 0.0]]
# D: [[0.50677133, -0.43742394], [0.0], [1.2930039, -1.2365541], [0.0, 0.0]]
# A : [[-0.27850878, -0.52411258], [0.0], [0.94569027, 0.83747566], [0.0, 0.0], [0.50677133, -0.43742394], [0.0], [1.2930039, -1.2365541], [0.0, 0.0]]

# after training D --- D and D in A changed
# G: [[-0.27850878, -0.52411258], [0.0], [0.94569027, 0.83747566], [0.0, 0.0]]
# D: [[0.49537802, -0.4082337], [0.0034225769], [1.2876366, -1.2274913], [0.047490694, 0.046951186]]
# A : [[-0.27850878, -0.52411258], [0.0], [0.94569027, 0.83747566], [0.0, 0.0], [0.49537802, -0.4082337], [0.0034225769], [1.2876366, -1.2274913], [0.047490694, 0.046951186]]

# after training A --- D didn't changed!
# G: [[-0.27628738, -0.52191412], [0.0054477928], [0.93868071, 0.84325212], [0.021782838, -0.017950913]]
# D: [[0.49537802, -0.4082337], [0.0034225769], [1.2876366, -1.2274913], [0.047490694, 0.046951186]]
# A : [[-0.27628738, -0.52191412], [0.0054477928], [0.93868071, 0.84325212], [0.021782838, -0.017950913], [0.49537802, -0.4082337], [0.0034225769], [1.2876366, -1.2274913], [0.047490694, 0.046951186]]

# after training D
# G: [[-0.27628738, -0.52191412], [0.0054477928], [0.93868071, 0.84325212], [0.021782838, -0.017950913]]
# D: [[0.45315021, -0.42550534], [-0.069068611], [1.2836961, -1.222793], [0.054722041, 0.11372232]]
# A : [[-0.27628738, -0.52191412], [0.0054477928], [0.93868071, 0.84325212], [0.021782838, -0.017950913], [0.45315021, -0.42550534], [-0.069068611], [1.2836961, -1.222793], [0.054722041, 0.11372232]]

# after training A
# G: [[-0.27531064, -0.52016109], [0.0084079718], [0.93036431, 0.85106117], [0.042959597, -0.037835769]]
# D: [[0.45315021, -0.42550534], [-0.069068611], [1.2836961, -1.222793], [0.054722041, 0.11372232]]
# A : [[-0.27531064, -0.52016109], [0.0084079718], [0.93036431, 0.85106117], [0.042959597, -0.037835769], [0.45315021, -0.42550534], [-0.069068611], [1.2836961, -1.222793], [0.054722041, 0.11372232]]

Hi, I can't see why this is happening. How model_D is selectively trainable? Is this because in model_A, model_D is added as a layer? Can you please explain what's happening here. Thank you

@Arvinth-s It is because once you compiled the model, changing the trainable attribute does not affect the model. If you want to change this attribute during training, you need to recompile the model.
Basically, the trainable attribute will keep the value it had when the model was compiled.

Edit: I guess you found the answer since December 2019, but for some people that might be useful.

@Bornlex Thank you kind man. I spent a week looking for information about the "trainable" parameter and why it works this way exactly