def darknet_base(inputs, include_yolo_head=True):
"""
Builds Darknet53 by reading the YOLO configuration file
:param inputs: Input tensor
:param include_yolo_head: Includes the YOLO head
:return: A list of output layers and the network config
"""
path = os.path.join(ROOT_DIR, 'cfg', '{}.cfg'.format(YOLO_VERSION))
blocks = Parser.parse_cfg(path)
x, layers, yolo_layers = inputs, [], []
ptr = 0
config = {}
for block in blocks:
block_type = block['type']
if block_type == 'net':
config = _read_net_config(block)
elif block_type == 'convolutional':
x, layers, yolo_layers, ptr = _build_conv_layer(
x, block, layers, yolo_layers, ptr, config)
elif block_type == 'shortcut':
x, layers, yolo_layers, ptr = _build_shortcut_layer(
x, block, layers, yolo_layers, ptr)
elif block_type == 'yolo':
x, layers, yolo_layers, ptr = _build_yolo_layer(
x, block, layers, yolo_layers, ptr, config)
elif block_type == 'route':
x, layers, yolo_layers, ptr = _build_route_layer(
x, block, layers, yolo_layers, ptr)
elif block_type == 'upsample':
x, layers, yolo_layers, ptr = _build_upsample_layer(
x, block, layers, yolo_layers, ptr)
elif block_type == 'maxpool':
x, layers, yolo_layers, ptr = _build_maxpool_layer(
x, block, layers, yolo_layers, ptr)
else:
raise ValueError(
'{} not recognized as block type'.format(block_type))
_verify_weights_completed_consumed(ptr)
if include_yolo_head:
output_layers = yolo_layers
return tf.keras.layers.Concatenate(axis=1)(output_layers), config
else:
output_layers = [
layers[i - 1] for i in range(len(layers)) if layers[i] is None
]
# NOTE: Apparently TFLite doesn't like Concatenate.
# return tf.keras.layers.Concatenate(axis=1)(output_layers), config
return output_layers, config
def darknet_base(inputs,
yolo_version='yolov3',
data_dir='./weights',
include_yolo_head=True):
"""
Builds Darknet53 by reading the YOLO configuration file
:param inputs: Input tensor
:param include_yolo_head: Includes the YOLO head
:return: A list of output layers and the network config
"""
weights_file = open(
os.path.join(data_dir, '{}.weights'.format(yolo_version)), 'rb')
major, minor, revision = np.ndarray(shape=(3, ),
dtype='int32',
buffer=weights_file.read(12))
if (major * 10 + minor) >= 2 and major < 1000 and minor < 1000:
seen = np.ndarray(shape=(1, ),
dtype='int64',
buffer=weights_file.read(8))
else:
seen = np.ndarray(shape=(1, ),
dtype='int32',
buffer=weights_file.read(4))
print('Weights Header: ', major, minor, revision, seen)
path = os.path.join(data_dir, '{}.cfg'.format(yolo_version))
blocks = Parser.parse_cfg(path)
x, layers, yolo_layers = inputs, [], []
ptr = 0
config = {}
for block in blocks:
block_type = block['type']
if block_type == 'net':
config = _read_net_config(block, data_dir, yolo_version)
elif block_type == 'convolutional':
x, layers, yolo_layers, ptr = _build_conv_layer(
x, block, layers, yolo_layers, ptr, config, weights_file)
elif block_type == 'shortcut':
x, layers, yolo_layers, ptr = _build_shortcut_layer(
x, block, layers, yolo_layers, ptr)
elif block_type == 'yolo':
x, layers, yolo_layers, ptr = _build_yolo_layer(
x, block, layers, yolo_layers, ptr, config)
elif block_type == 'route':
x, layers, yolo_layers, ptr = _build_route_layer(
x, block, layers, yolo_layers, ptr)
elif block_type == 'upsample':
x, layers, yolo_layers, ptr = _build_upsample_layer(
x, block, layers, yolo_layers, ptr)
elif block_type == 'maxpool':
x, layers, yolo_layers, ptr = _build_maxpool_layer(
x, block, layers, yolo_layers, ptr)
else:
raise ValueError(
'{} not recognized as block type'.format(block_type))
_verify_weights_completed_consumed(ptr, weights_file)
if include_yolo_head:
output_layers = yolo_layers
return tf.keras.layers.Concatenate(axis=1)(output_layers), config
else:
output_layers = [
layers[i - 1] for i in range(len(layers)) if layers[i] is None
]
# NOTE: Apparently TFLite doesn't like Concatenate.
# return tf.keras.layers.Concatenate(axis=1)(output_layers), config
return output_layers, config