Number of stride in EfficientNet 5th MBconv block layer

toandaominh1997 / EfficientDet.Pytorch

Implementation EfficientDet: Scalable and Efficient Object Detection in PyTorch

MIT License

1.44k stars 306 forks source link

Number of stride in EfficientNet 5th MBconv block layer #92

Open SungjoonPark opened 4 years ago

SungjoonPark commented 4 years ago

Hi, I've found that # of stride in 5th MBconv block layer of the backbone (efficient net) is set to 2 ('r3_k5_s22_e6_i80_o112_se0.25'), which are different from the official Tensorflow implementation. ('r3_k5_s11_e6_i80_o112_se0.25') This might affect halving the resolution of the output of that layer, but seems should be kept to the same (paper, Table 1). Is it intended behavior?

SungjoonPark commented 4 years ago

along with the last (7th) block. Should it be "r1_k3_s11_e6_i192_o320_se0.25"?