def __init__(self, encoder=None, decoder=None):
super(AutoEncoder, self).__init__()
if encoder is None:
encoder = {
"name": "VGG16",
"module": "deeplodocus.app.models.vgg",
"kwargs": {
"batch_norm": True,
"pretrained": True,
"num_channels": 3,
"include_top": False
}
}
else:
if "include_top" not in encoder["kwargs"]:
Notification(
DEEP_NOTIF_WARNING,
"include_top not specified for the encoder, should be set to False"
)
if decoder is None:
decoder = {
"name": "VGG16Decode",
"module": "deeplodocus.app.models.vgg",
"kwargs": {
"batch_norm": True,
"num_classes": 10
}
}
self.encoder = get_specific_module(
name=encoder["name"],
module=encoder["module"])(**encoder["kwargs"])
self.decoder = get_specific_module(
name=decoder["name"],
module=decoder["module"],
)(**decoder["kwargs"])
def __init__(self,
backbone=None,
num_classes=10,
skip_layers=None,
anchors=(((10, 13), (16, 30), (22, 23)), ((30, 61), (62, 45),
(59, 119)),
((116, 90), (156, 198), (373, 326)))):
super(YOLO, self).__init__()
self.num_classes = num_classes
self.anchors = anchors
# If no backbone specified, set default backbone arguments
if backbone is None:
backbone = {
"name": "Darknet53",
"module": "deeplodocus.app.models.darknet",
"kwargs": {
"num_channels": 3,
"include_top": False
}
}
skip_layers = (36, 61)
# If skip layers not specified, select the appropriate indices in accordance with skip_dict
if skip_layers is None:
try:
self.skip_layers = skip_dict[backbone["name"]]
Notification(
DEEP_NOTIF_INFO,
MSG_AUTO_SKIP % (backbone["name"], str(self.skip_layers)))
except KeyError:
Notification(DEEP_NOTIF_FATAL, MSG_NO_SKIP % backbone["name"])
else:
self.skip_layers = skip_layers
# Get and initialise the backbone module
backbone_module = get_specific_module(name=backbone["name"],
module=backbone["module"],
fatal=True)
self.backbone = backbone_module(**backbone["kwargs"])
# CONVOLUTIONAL LAYERS/BLOCKS (ConvBlocks contain 4 conv layers)
self.conv_1_1 = ConvLayer(in_channels=1024,
out_channels=512,
kernel_size=1,
negative_slope=0.1)
self.conv_1_2 = ConvBlock(1024)
self.conv_1_3 = ConvLayer(in_channels=512,
out_channels=1024,
kernel_size=3,
padding=1,
negative_slope=0.1)
self.conv_1_4 = nn.Conv2d(in_channels=1024,
out_channels=len(anchors) *
(num_classes + 5),
kernel_size=1)
self.conv_2_1 = ConvLayer(in_channels=512,
out_channels=256,
kernel_size=1,
negative_slope=0.1)
self.conv_2_2 = ConvLayer(in_channels=768,
out_channels=256,
kernel_size=1,
negative_slope=0.1)
self.conv_2_3 = ConvBlock(512)
self.conv_2_4 = ConvLayer(in_channels=256,
out_channels=512,
kernel_size=3,
padding=1,
negative_slope=0.1)
self.conv_2_5 = nn.Conv2d(in_channels=512,
out_channels=len(anchors) *
(num_classes + 5),
kernel_size=1)
self.conv_3_1 = ConvLayer(in_channels=256,
out_channels=128,
kernel_size=1,
negative_slope=0.1)
self.conv_3_2 = ConvLayer(in_channels=384,
out_channels=128,
kernel_size=1,
negative_slope=0.1)
self.conv_3_3 = ConvBlock(256)
self.conv_3_4 = ConvLayer(in_channels=128,
out_channels=256,
kernel_size=3,
padding=1,
negative_slope=0.1)
self.conv_3_5 = nn.Conv2d(in_channels=256,
out_channels=len(anchors) *
(num_classes + 5),
kernel_size=1)
# UPSAMPLE LAYER
self.upsample = nn.Upsample(scale_factor=2, mode="nearest")
# Classify after backbone
self.classifier = nn.Linear(in_features=1024, out_features=num_classes)
def __init__(
self,
backbone=None,
num_classes=80,
skip_layers=(36, 61),
input_shape=(256, 256),
anchors=(((116, 90), (156, 198), (373, 326)),
((30, 61), (62, 45), (59, 119)), ((10, 13), (16, 30), (22,
23))),
normalized_anchors=False,
):
super(YOLOv3, self).__init__()
# Default backbone arguments
if backbone is None:
backbone = {
"name": "Daknet53",
"module": "deeplodocus.app.models.darknet",
"kwargs": {
"num_channels": 3,
"include_top": False
}
}
# Scale anchors by image dimensions if the anchors are normalized
if normalized_anchors:
anchors = [[(a[0] * input_shape[1], a[1] * input_shape[0])
for a in anchor] for anchor in anchors]
# Get the number of anchor boxes
num_anchors = len(anchors)
# Get and initialise the backbone module
backbone_module = get_specific_module(name=backbone["name"],
module=backbone["module"],
fatal=True)
self.backbone = backbone_module(**backbone["kwargs"])
# The keys to extract from the list
self.skip_layers = skip_layers
# CONVOLUTIONAL LAYERS/BLOCKS (ConvBlocks consist of 4 conv layers)
self.conv_1_1 = ConvLayer(in_channels=1024,
out_channels=512,
kernel_size=1,
negative_slope=0.1)
self.conv_1_2 = ConvBlock(1024)
self.conv_1_3 = ConvLayer(in_channels=512,
out_channels=1024,
kernel_size=3,
padding=1,
negative_slope=0.1)
self.conv_1_4 = nn.Conv2d(in_channels=1024,
out_channels=num_anchors * (num_classes + 5),
kernel_size=1)
self.conv_2_1 = ConvLayer(in_channels=512,
out_channels=256,
kernel_size=1,
negative_slope=0.1)
self.conv_2_2 = ConvLayer(in_channels=768,
out_channels=256,
kernel_size=1,
negative_slope=0.1)
self.conv_2_3 = ConvBlock(512)
self.conv_2_4 = ConvLayer(in_channels=256,
out_channels=512,
kernel_size=3,
padding=1,
negative_slope=0.1)
self.conv_2_5 = nn.Conv2d(in_channels=512,
out_channels=num_anchors * (num_classes + 5),
kernel_size=1)
self.conv_3_1 = ConvLayer(in_channels=256,
out_channels=128,
kernel_size=1,
negative_slope=0.1)
self.conv_3_2 = ConvLayer(in_channels=384,
out_channels=128,
kernel_size=1,
negative_slope=0.1)
self.conv_3_3 = ConvBlock(256)
self.conv_3_4 = ConvLayer(in_channels=128,
out_channels=256,
kernel_size=3,
padding=1,
negative_slope=0.1)
self.conv_3_5 = nn.Conv2d(in_channels=256,
out_channels=num_anchors * (num_classes + 5),
kernel_size=1)
# UPSAMPLE LAYER
self.upsample = nn.Upsample(scale_factor=2, mode="nearest")
# YOLO LAYERS
self.yolo_layer_1 = YoloLayer(num_classes=num_classes,
image_shape=input_shape,
anchors=anchors[0])
self.yolo_layer_2 = YoloLayer(num_classes=num_classes,
image_shape=input_shape,
anchors=anchors[1])
self.yolo_layer_3 = YoloLayer(num_classes=num_classes,
image_shape=input_shape,
anchors=anchors[2])