super-gradients/src/super_gradients/recipes/arch_params/csp_resnet_arch_params.yaml

layers: [3, 6, 6, 3]                # model's structure
channels: [64, 128, 256, 512, 1024] # number of outputs channels for step and consecutive feature maps
activation: silu                    # model's width multiplier
return_idx: [1,2,3]                 # Indexes of feature maps to output
use_large_stem: True                # If True, uses 3 conv+bn+act instead of 2 in stem blocks
width_mult:                         # scaling factor to number of channels
depth_mult:                         # scaling factor to number of layers
use_alpha: False                    # If True, enables additional learnable weighting parameter for 1x1 branch in RepVGGBlock