super-gradients/recipes/arch_params/yolo_arch_params.yaml

defaults:
  - default_arch_params

anchors:
  _target_: super_gradients.training.utils.detection_utils.Anchors
  anchors_list: [[10, 13, 16, 30, 33, 23],[30, 61, 62, 45, 59, 119],[116, 90, 156, 198, 373, 326]]
  strides: [8, 16, 32]
num_classes: 80  # Number of classes to predict
depth_mult_factor: 1.0  # depth multiplier for the entire model, overridden for predefined YoloV5S, YoloV5M, YoloV5L
width_mult_factor: 1.0  # width multiplier for the entire model, overridden for predefined YoloV5S, YoloV5M, YoloV5L
channels_in: 3  # # of classes the model predicts
skip_connections_dict: {12: [6], 16: [4], 19: [14], 22: [10], 24: [17, 20]}
# A dictionary defining skip connections. format is target: [source1, source2, ...]. Each item defines a skip
# connection from all sources to the target according to the layers index (count starts from the backbone)
connection_layers_input_channel_size: [1024, 1024, 512]
# default number off channels for the connecting points between the backbone and the head
fuse_conv_and_bn: False  # Fuse sequential Conv + B.N layers into a single one
add_nms: False  # Add the NMS module to the computational graph
nms_conf: 0.25  # When add_nms is True during NMS predictions with confidence lower than this will be discarded
nms_iou: 0.45  # When add_nms is True IoU threshold for NMS algorithm
               # (with smaller value more boxed will be considered "the same" and removed)
yolo_version: 'v6.0'  # Release version of Ultralytics to built a model from: v.6.0 and v3.0 are supported
_convert_: all