search

google / automl

Google Brain AutoML
Apache License 2.0
6.22k stars 1.45k forks source link

Why is the config in effnetv2_configs.py not consistent with that in paper? #1175

Closed HydrogenSulfate closed 1 year ago

HydrogenSulfate commented 1 year ago

for example: the max value of dropout_rate and ram is 0.2 and 10 in effnetv2_configs.py, but 0.3 and 15 in paper, and dropout_rate seems not progressively changed in 4 traning stages

'efficientnetv2-s':  # 83.9% @ 22M
        (v2_s_block, 1.0, 1.0, 300, 384, 0.2, 10, 0, 'randaug'),

image

Royyyy0 commented 1 year ago

I got the same question as yours, do you have any ideas now? Thx

HydrogenSulfate commented 1 year ago

I got the same question as yours, do you have any ideas now? Thx

We've tries several times for reproducing with paddlepaddle, but got many result up to Top1=83.58%, still has 0.27% diff to official Top1=83.85%.

We find this may be a slightly overfitting, so we add a Mixup to moderate, and finally achieve to 83.84%, please check the model and config below: efficientnet_v2.py EfficientNetV2_S.yaml

The reson might be framework differences between tf and paddlepaddle/pytorch. As we all know, many paper in tensorflow always hard to reproduce same metric by other DL frameworks and always got slightly lower metric.