def test_resample_feature_adder_compile(self):
config = hparams_config.get_efficientdet_config("efficientdet-d0")
feat_sizes = utils.get_feat_sizes(config.image_size, config.max_level)
tf2.random.set_seed(SEED)
inputs = [
tf2.keras.Input(shape=[512, 512, 3]),
tf2.keras.Input(shape=[256, 256, 16]),
tf2.keras.Input(shape=[128, 128, 24]),
tf2.keras.Input(shape=[64, 64, 40]),
tf2.keras.Input(shape=[32, 32, 112]),
tf2.keras.Input(shape=[16, 16, 320])
]
outputs = efficientdet_arch_keras.ResampleFeatureAdder(config)(inputs)
model = tf2.keras.Model(inputs=inputs, outputs=outputs)
examples = [[
tf2.ones([1, 512, 512, 3]),
tf2.ones([1, 256, 256, 16]),
tf2.ones([1, 128, 128, 24]),
tf2.ones([1, 64, 64, 40]),
tf2.ones([1, 32, 32, 112]),
tf2.ones([1, 16, 16, 320])
]]
preds = model(examples)
try:
utils.verify_feats_size(preds,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
except ValueError as err:
self.assertFalse(True, msg=repr(err))
self.assertEqual(len(preds), 5, "P3-P7")
def build_feature_network(features, config):
"""Build FPN input features.
Args:
features: input tensor.
config: a dict-like config, including all parameters.
Returns:
A dict from levels to the feature maps processed after feature network.
"""
feat_sizes = utils.get_feat_sizes(config.image_size, config.max_level)
feats = []
if config.min_level not in features.keys():
raise ValueError(
'features.keys ({}) should include min_level ({})'.format(
features.keys(), config.min_level))
# Build additional input features that are not from backbone.
for level in range(config.min_level, config.max_level + 1):
if level in features.keys():
feats.append(features[level])
else:
h_id, w_id = (2,
3) if config.data_format == 'channels_first' else (1,
2)
# Adds a coarser level by downsampling the last feature map.
feats.append(
ResampleFeatureMap(
target_height=(feats[-1].shape[h_id] - 1) // 2 + 1,
target_width=(feats[-1].shape[w_id] - 1) // 2 + 1,
target_num_channels=config.fpn_num_filters,
apply_bn=config.apply_bn_for_resampling,
is_training=config.is_training_bn,
conv_after_downsample=config.conv_after_downsample,
use_native_resize_op=config.use_native_resize_op,
pooling_type=config.pooling_type,
strategy=config.strategy,
data_format=config.data_format,
name='resample_p{}'.format(level),
)(feats[-1]))
utils.verify_feats_size(feats,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
new_feats = FPNCells(feat_sizes, config)(feats)
return new_feats
def call(self, feats, training):
for cell in self.cells:
cell_feats = cell(feats, training)
min_level = self.config.min_level
max_level = self.config.max_level
feats = []
for level in range(min_level, max_level + 1):
for i, fnode in enumerate(reversed(self.fpn_config.nodes)):
if fnode['feat_level'] == level:
feats.append(cell_feats[-1 - i])
break
utils.verify_feats_size(feats, self.feat_sizes, min_level, max_level,
self.config.data_format)
return feats
def call(self, feats):
for cell in self.cells:
cell_feats = cell(feats)
min_level = self.config.min_level
max_level = self.config.max_level
new_feats = {}
for l in range(min_level, max_level + 1):
for i, fnode in enumerate(reversed(self.fpn_config.nodes)):
if fnode['feat_level'] == l:
new_feats[l] = cell_feats[-1 - i]
break
feats = [
new_feats[level] for level in range(min_level, max_level + 1)
]
utils.verify_feats_size(feats, self.feat_sizes, min_level,
max_level, self.config.data_format)
return new_feats
def build_feature_network(feats, config):
"""Build FPN input features.
Args:
feats: A list of input tensors starting from min_level.
config: a dict-like config, including all parameters.
Returns:
A dict from levels to the feature maps processed after feature network.
"""
feat_sizes = utils.get_feat_sizes(config.image_size, config.max_level)
if not feats:
raise ValueError('FPN input features cannot be empty.')
# Build additional input features that are not from backbone.
while len(feats) < config.max_level - config.min_level + 1:
level = len(feats) + config.min_level
h_id, w_id = (2, 3) if config.data_format == 'channels_first' else (1,
2)
# Adds a coarser level by downsampling the last feature map.
feats.append(
ResampleFeatureMap(
target_height=(feats[-1].shape[h_id] - 1) // 2 + 1,
target_width=(feats[-1].shape[w_id] - 1) // 2 + 1,
target_num_channels=config.fpn_num_filters,
apply_bn=config.apply_bn_for_resampling,
is_training=config.is_training_bn,
conv_after_downsample=config.conv_after_downsample,
strategy=config.strategy,
data_format=config.data_format,
name='resample_p{}'.format(level),
)(feats[-1]))
utils.verify_feats_size(feats,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
new_feats = FPNCells(feat_sizes, config)(feats)
return new_feats
def build_feature_network(features, config):
"""Build FPN input features.
Args:
features: input tensor.
config: a dict-like config, including all parameters.
Returns:
A dict from levels to the feature maps processed after feature network.
"""
feat_sizes = utils.get_feat_sizes(config.image_size, config.max_level)
feats = []
if config.min_level not in features.keys():
raise ValueError(
'features.keys ({}) should include min_level ({})'.format(
features.keys(), config.min_level))
# Build additional input features that are not from backbone.
for level in range(config.min_level, config.max_level + 1):
if level in features.keys():
feats.append(features[level])
else:
h_id, w_id = (2,
3) if config.data_format == 'channels_first' else (1,
2)
# Adds a coarser level by downsampling the last feature map.
feats.append(
resample_feature_map(
feats[-1],
name='p%d' % level,
target_height=(feats[-1].shape[h_id] - 1) // 2 + 1,
target_width=(feats[-1].shape[w_id] - 1) // 2 + 1,
target_num_channels=config.fpn_num_filters,
apply_bn=config.apply_bn_for_resampling,
is_training=config.is_training_bn,
conv_after_downsample=config.conv_after_downsample,
strategy=config.strategy,
data_format=config.data_format))
utils.verify_feats_size(feats,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
with tf.variable_scope('fpn_cells'):
for rep in range(config.fpn_cell_repeats):
with tf.variable_scope('cell_{}'.format(rep)):
logging.info('building cell %d', rep)
new_feats = build_bifpn_layer(feats, feat_sizes, config)
feats = [
new_feats[level]
for level in range(config.min_level, config.max_level + 1)
]
utils.verify_feats_size(feats,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
return new_feats
