search

megvii-research / FSCE

Apache License 2.0
280 stars 47 forks source link

Unable to reproduce the results on MS COCO datasets #24

Closed chenf99 closed 2 years ago

chenf99 commented 3 years ago

Dear author, I tried to reproduce the results on MS COCO using the default config and base model from http://dl.yf.io/fs-det/models/coco/base_model/model_final.pth, but got average AP 10.460 on 10 seeds, could you please provide the model_final.pth you trained? Thank you very much!

bsun0802 commented 3 years ago

How many GPUs did you use?

chenf99 commented 3 years ago

How many GPUs did you use?

4 with batchsize unchanged.

bsun0802 commented 3 years ago

The specification we provide is for 8-GPUs.

kike-0304 commented 2 years ago

我们提供的规格适用于 8-GPU。

Why do you need 8gpu? Can't I just keep the batch size consistent?

bsun0802 commented 2 years ago

Batch size is a hyper parameter, just like learning rate. Few-shot learning is sensitive to training setting due to low data regime.

You can experiments with 4 cards, but we didn't have recommendations for hyper-params for it.

A quick theoretical background on batch size and learning rate: https://stackoverflow.com/questions/53033556/how-should-the-learning-rate-change-as-the-batch-size-change