46 lines
1.5 KiB
Python
46 lines
1.5 KiB
Python
import numpy as np
|
|
import pytest
|
|
|
|
from keras_core import backend
|
|
from keras_core import layers
|
|
from keras_core import testing
|
|
|
|
|
|
class DropoutTest(testing.TestCase):
|
|
def test_dropout_basics(self):
|
|
self.run_layer_test(
|
|
layers.Dropout,
|
|
init_kwargs={
|
|
"rate": 0.2,
|
|
},
|
|
input_shape=(2, 3),
|
|
expected_output_shape=(2, 3),
|
|
expected_num_trainable_weights=0,
|
|
expected_num_non_trainable_weights=0,
|
|
expected_num_seed_generators=1,
|
|
expected_num_losses=0,
|
|
supports_masking=True,
|
|
)
|
|
|
|
def test_dropout_rescaling(self):
|
|
inputs = np.ones((20, 500))
|
|
layer = layers.Dropout(0.5, seed=1337)
|
|
outputs = layer(inputs, training=True)
|
|
self.assertAllClose(np.mean(outputs), 1.0, atol=0.02)
|
|
self.assertAllClose(np.max(outputs), 2.0)
|
|
|
|
@pytest.mark.skipif(
|
|
backend.backend() != "tensorflow", reason="Requires dynamic shapes"
|
|
)
|
|
def test_dropout_partial_noise_shape_dynamic(self):
|
|
inputs = np.ones((20, 5, 10))
|
|
layer = layers.Dropout(0.5, noise_shape=(None, 1, None))
|
|
outputs = layer(inputs, training=True)
|
|
self.assertAllClose(outputs[:, 0, :], outputs[:, 1, :])
|
|
|
|
def test_dropout_partial_noise_shape_static(self):
|
|
inputs = np.ones((20, 5, 10))
|
|
layer = layers.Dropout(0.5, noise_shape=(20, 1, 10))
|
|
outputs = layer(inputs, training=True)
|
|
self.assertAllClose(outputs[:, 0, :], outputs[:, 1, :])
|