def test_binarization(self, quantized_layer, layer, input_shape, kwargs,
keras_should_run_eagerly):
input_data = testing_utils.random_input(input_shape)
random_weight = np.random.random() - 0.5
with lq.metrics.scope(["flip_ratio"]):
quant_output = testing_utils.layer_test(
quantized_layer,
kwargs=dict(
**kwargs,
kernel_quantizer="ste_sign",
input_quantizer="ste_sign",
kernel_initializer=tf.keras.initializers.constant(
random_weight),
),
input_data=input_data,
should_run_eagerly=keras_should_run_eagerly,
)
fp_model = tf.keras.models.Sequential([
layer(
**kwargs,
kernel_initializer=tf.keras.initializers.constant(
np.sign(random_weight)),
input_shape=input_shape[1:],
)
])
np.testing.assert_allclose(quant_output,
fp_model.predict(np.sign(input_data)))
def test_depthwise_layers(self, keras_should_run_eagerly):
input_data = testing_utils.random_input((2, 3, 7, 6))
random_weight = np.random.random() - 0.5
with lq.metrics.scope(["flip_ratio"]):
quant_output = testing_utils.layer_test(
lq.layers.QuantDepthwiseConv2D,
kwargs=dict(
kernel_size=3,
depthwise_quantizer="ste_sign",
input_quantizer="ste_sign",
depthwise_initializer=tf.keras.initializers.constant(
random_weight),
),
input_data=input_data,
should_run_eagerly=keras_should_run_eagerly,
)
fp_model = tf.keras.models.Sequential([
tf.keras.layers.DepthwiseConv2D(
kernel_size=3,
depthwise_initializer=tf.keras.initializers.constant(
np.sign(random_weight)),
input_shape=input_data.shape[1:],
)
])
np.testing.assert_allclose(quant_output,
fp_model.predict(np.sign(input_data)))
def test_layer_as_quantizer(self, quantizer, keras_should_run_eagerly):
"""Test whether a keras.layers.Layer can be used as quantizer."""
input_data = testing_utils.random_input((1, 10))
model = tf.keras.Sequential(
[lq.layers.QuantDense(1, **{quantizer: DummyTrainableQuantizer()})]
)
model.compile(optimizer="sgd", loss="mse", run_eagerly=keras_should_run_eagerly)
model.fit(input_data, np.ones((1,)), epochs=1)
assert any(["dummy_weight" in var.name for var in model.trainable_variables])
def test_separable_layers(
self, quantized_layer, layer, input_shape, keras_should_run_eagerly
):
input_data = testing_utils.random_input(input_shape)
random_d_kernel = np.random.random() - 0.5
random_p_kernel = np.random.random() - 0.5
with lq.context.metrics_scope(["flip_ratio"]):
quant_output = testing_utils.layer_test(
quantized_layer,
kwargs=dict(
filters=3,
kernel_size=3,
depthwise_quantizer="ste_sign",
pointwise_quantizer="ste_sign",
input_quantizer="ste_sign",
depthwise_initializer=tf.keras.initializers.constant(
random_d_kernel
),
pointwise_initializer=tf.keras.initializers.constant(
random_p_kernel
),
),
input_data=input_data,
should_run_eagerly=keras_should_run_eagerly,
)
fp_model = tf.keras.models.Sequential(
[
layer(
filters=3,
kernel_size=3,
depthwise_initializer=tf.keras.initializers.constant(
np.sign(random_d_kernel)
),
pointwise_initializer=tf.keras.initializers.constant(
np.sign(random_p_kernel)
),
input_shape=input_shape[1:],
)
]
)
np.testing.assert_allclose(quant_output, fp_model.predict(np.sign(input_data)))
