search

iamhankai / ghostnet.pytorch

[CVPR2020] GhostNet: More Features from Cheap Operations
https://arxiv.org/abs/1911.11907
523 stars 118 forks source link

Release of PyTorch pretrained model and ablation #17

Closed d-li14 closed 4 years ago

d-li14 commented 4 years ago

Thanks for your awesome work. I have provided the PyTorch pre-trained GhostNet 1.0x with comparable accuracy in https://github.com/d-li14/ghostnet.pytorch. Furthermore, I conducted ablation experiments about different training settings, which may echo some questions in #3, #8.

iamhankai commented 4 years ago

Great job! Thanks for your works. In addition, 360 epochs, warmup and lower dropout ratio may help to further improve the performance.

d-li14 commented 4 years ago

Thanks for your kind suggestion! I have tried training longer until 360 epochs, but it brings about marginal benefit. I have also applied warmup in the original experiments. Seen from the training curve, I conjecture that keeping the dropout ratio as 0.2 will be a better choice compared to 0.1, 0.15. Correct me if I am wrong.

iamhankai commented 4 years ago

Official pytorch pretrained model is released: https://github.com/huawei-noah/ghostnet/tree/master/pytorch