def test_mobile_conv2d(self):
conv2d = movinet_layers.MobileConv2D(
filters=3,
kernel_size=(3, 3),
strides=(1, 1),
padding='same',
kernel_initializer='ones',
use_bias=False,
use_depthwise=False,
use_temporal=False,
use_buffered_input=True,
)
inputs = tf.ones([1, 2, 2, 2, 3])
predicted = conv2d(inputs)
expected = tf.constant(
[[[[[12., 12., 12.],
[12., 12., 12.]],
[[12., 12., 12.],
[12., 12., 12.]]],
[[[12., 12., 12.],
[12., 12., 12.]],
[[12., 12., 12.],
[12., 12., 12.]]]]])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
def test_mobile_conv2d_bn(self):
batch_norm_op = tf.keras.layers.BatchNormalization(
momentum=0.9,
epsilon=1.,
name='bn')
conv2d = movinet_layers.MobileConv2D(
filters=3,
kernel_size=(3, 3),
strides=(1, 1),
padding='same',
kernel_initializer='ones',
use_bias=False,
use_depthwise=False,
use_temporal=False,
use_buffered_input=True,
batch_norm_op=batch_norm_op,
)
inputs = tf.ones([1, 2, 2, 2, 3])
predicted = conv2d(inputs)
expected = tf.constant(
[[[[[8.48528, 8.48528, 8.48528],
[8.48528, 8.48528, 8.48528]],
[[8.48528, 8.48528, 8.48528],
[8.48528, 8.48528, 8.48528]]],
[[[8.48528, 8.48528, 8.48528],
[8.48528, 8.48528, 8.48528]],
[[8.48528, 8.48528, 8.48528],
[8.48528, 8.48528, 8.48528]]]]])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
def test_mobile_conv2d_temporal(self):
conv2d = movinet_layers.MobileConv2D(
filters=3,
kernel_size=(3, 1),
strides=(1, 1),
padding='causal',
kernel_initializer='ones',
use_bias=False,
use_depthwise=True,
use_temporal=True,
use_buffered_input=True,
)
inputs = tf.ones([1, 2, 2, 1, 3])
paddings = [[0, 0], [2, 0], [0, 0], [0, 0], [0, 0]]
padded_inputs = tf.pad(inputs, paddings)
predicted = conv2d(padded_inputs)
expected = tf.constant(
[[[[[1., 1., 1.]],
[[1., 1., 1.]]],
[[[2., 2., 2.]],
[[2., 2., 2.]]]]])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)