def test_resnext_block_can_be_called_channels_first(self):
inputs = tf.random.normal([2, 32, 16, 16], dtype=tf.float32)
block = blocks.ResNextBlock(filters=64,
strides=2,
use_projection=True,
data_format='channels_first')
outputs = block(inputs, training=True)
grads = tf.gradients(outputs, inputs)
self.assertTrue(tf.compat.v1.trainable_variables())
self.assertTrue(grads)
self.assertTrue(tf.get_collection(tf.GraphKeys.UPDATE_OPS))
self.assertListEqual([2, 256, 8, 8], outputs.shape.as_list())
def test_resnext_block_can_be_called_float16(self):
inputs = tf.random.normal([2, 16, 16, 32], dtype=tf.float16)
with tf.variable_scope('float16', custom_getter=custom_float16_getter):
block = blocks.ResNextBlock(filters=64,
strides=2,
use_projection=True,
data_format='channels_last')
outputs = block(inputs, training=True)
grads = tf.gradients(outputs, inputs)
self.assertTrue(tf.compat.v1.trainable_variables())
self.assertTrue(grads)
self.assertTrue(tf.get_collection(tf.GraphKeys.UPDATE_OPS))
self.assertListEqual([2, 8, 8, 256], outputs.shape.as_list())