Closed shiyanshilaoshu closed 4 months ago
Hi! Thank you very much for your feedback. Indeed, the true/false statements are a mistake; I should have removed them. They're leftover from my research! I wanted to add a small spatial attention block, like in the mobilenetv3 and EfficientDet paper, with squeeze-and-excite at specific parts of LeYOLO, using a true or false statement where you wish to use it. However, I never found any interest in using SE or channel-attention, and the parameter is not used anymore in the block code. However, I let it stay that way to keep my research going while publishing LeYOLO; I might change everything to false and document the .yaml file better! The code is in ultralytics/nn/modules/block.py.
Hello, your work is really impressive, and I really like your experiments on the number of channels in the inverted bottleneck structure. Additionally, I would like to ask about the meaning of the fourth parameter (True or False) in the args of MobileNetV3BLOCK in the source code you provided(./ultralytics/cfg/cfg/leyolo.yaml). Does it represent enabling the inductive bias of the convolutional layer? Could you please tell me where its detailed definition is in the code?