def test_fail_due_to_call_in_invalid_place_subclass():
"""this test is failed. `set_seed` must be called at just before generating random variables"""
# not reproducible. if call `build` twice, not same variables are generated in each built models.
set_seed(200)
def build():
class SubClass(tf.keras.Model):
def __init__(self, *args, **kwargs):
super(SubClass, self).__init__(*args, **kwargs)
self.a = tf.keras.layers.Conv2D(4, 3)
self.b = tf.keras.layers.BatchNormalization()
self.c = tf.keras.layers.Activation("relu")
self.d = tf.keras.layers.Flatten()
self.e = tf.keras.layers.Dense(2)
def call(self, x):
x = self.a(x)
x = self.b(x)
x = self.c(x)
x = self.d(x)
return self.e(x)
layers = [SubClass()]
model = build_helper(layers, (16, 16, 3))
return model
assert not is_reproducible(build)
def build():
set_seed(200)
layers = [
tf.keras.layers.Conv2D(4, 3, input_shape=(16, 16, 3)),
tf.keras.layers.BatchNormalization(),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(2),
]
model = tf.keras.Sequential(layers)
return model
def build():
set_seed(200)
layers = [
tf.keras.layers.Conv2D(4, 3),
tf.keras.layers.BatchNormalization(),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(2),
]
model = build_helper(layers, (16, 16, 3))
return model
def test_fail_due_to_call_in_invalid_place_functional_api():
"""this test is failed. `set_seed` must be called at just before generating random variables"""
# not reproducible. if call `build` twice, not same variables are generated in each built models.
set_seed(200)
def build():
layers = [
tf.keras.layers.Conv2D(4, 3),
tf.keras.layers.BatchNormalization(),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(2),
]
model = build_helper(layers, (16, 16, 3))
return model
assert not is_reproducible(build)
def build():
set_seed(200)
class SubClass(tf.keras.Model):
def __init__(self, *args, **kwargs):
super(SubClass, self).__init__(*args, **kwargs)
self.a = tf.keras.layers.Conv2D(4, 3)
self.b = tf.keras.layers.BatchNormalization()
self.c = tf.keras.layers.Activation("relu")
self.d = tf.keras.layers.Flatten()
self.e = tf.keras.layers.Dense(2)
def call(self, x):
x = self.a(x)
x = self.b(x)
x = self.c(x)
x = self.d(x)
return self.e(x)
layers = [SubClass()]
model = build_helper(layers, (16, 16, 3))
return model