def test_resample_feature_map(self):
feat = tf.random.uniform([1, 16, 16, 320])
for apply_bn in [True, False]:
for training in [True, False]:
for strategy in ['tpu', '']:
with self.subTest(apply_bn=apply_bn,
training=training,
strategy=strategy):
tf.random.set_random_seed(SEED)
expect_result = legacy_arch.resample_feature_map(
feat,
name='resample_p0',
target_height=8,
target_width=8,
target_num_channels=64,
apply_bn=apply_bn,
is_training=training,
strategy=strategy)
tf.random.set_random_seed(SEED)
resample_layer = efficientdet_keras.ResampleFeatureMap(
name='resample_p0',
feat_level=0,
target_num_channels=64,
apply_bn=apply_bn,
is_training_bn=training,
strategy=strategy)
all_feats = [tf.ones([1, 8, 8, 64])]
actual_result = resample_layer(feat, training,
all_feats)
self.assertAllCloseAccordingToType(
expect_result, actual_result)
def test_resample_feature_map(self):
feat = tf.random.uniform([1, 16, 16, 320])
for apply_fn in [True, False]:
for is_training in [True, False]:
for use_tpu in [True, False]:
with self.subTest(apply_fn=apply_fn,
is_training=is_training,
use_tpu=use_tpu):
tf.random.set_random_seed(111111)
expect_result = efficientdet_arch.resample_feature_map(
feat,
name='resample_p0',
target_height=8,
target_width=8,
target_num_channels=64,
apply_bn=apply_fn,
is_training=is_training,
use_tpu=use_tpu)
tf.random.set_random_seed(111111)
actual_result = efficientdet_arch_keras.ResampleFeatureMap(
name='resample_p0',
target_height=8,
target_width=8,
target_num_channels=64,
apply_bn=apply_fn,
is_training=is_training,
use_tpu=use_tpu)(feat)
self.assertAllCloseAccordingToType(
expect_result, actual_result)
def test_resample_var_names(self):
with tf.Graph().as_default():
feat = tf.random.uniform([1, 16, 16, 320])
resample_layer = efficientdet_keras.ResampleFeatureMap(
name='resample_p0', feat_level=0, target_num_channels=64)
all_feats = [tf.ones([1, 8, 8, 64])]
resample_layer(feat, True, all_feats)
vars1 = sorted([var.name for var in tf.trainable_variables()])
with tf.Graph().as_default():
feat = tf.random.uniform([1, 16, 16, 320])
legacy_arch.resample_feature_map(feat,
name='p0',
target_height=8,
target_width=8,
target_num_channels=64)
vars2 = sorted([var.name for var in tf.trainable_variables()])
self.assertEqual(vars1, vars2)
def test_var_names(self):
with tf.Graph().as_default():
feat = tf.random.uniform([1, 16, 16, 320])
resample_layer = efficientdet_arch_keras.ResampleFeatureMap(
name='resample_p0',
target_height=8,
target_width=8,
target_num_channels=64)
resample_layer(feat)
vars1 = [var.name for var in tf.trainable_variables()]
with tf.Graph().as_default():
feat = tf.random.uniform([1, 16, 16, 320])
legacy_arch.resample_feature_map(feat,
name='p0',
target_height=8,
target_width=8,
target_num_channels=64)
vars2 = [var.name for var in tf.trainable_variables()]
self.assertEqual(vars1, vars2)
