search

Alibaba-MIIL / ML_Decoder

Official PyTorch implementation of "ML-Decoder: Scalable and Versatile Classification Head" (2021)
MIT License
317 stars 53 forks source link

Reproducing the results on the VOC dataset #14

Closed kprokofi closed 2 years ago

kprokofi commented 2 years ago

Hi, thanks for your work. I tried to reproduce the results on the VOC dataset. I added the dataset to your code. This is my fork. I also added Timm wrapper to use timm models

I follow the article's recommendations and run the following command: python train.py --data /path/to/voc --model-name tresnet_l --num-classes 20 --batch-size 48 --lr 2e-4

I achieved only 96.1 % with tresnet_l (Q2L level)

Could you, please, help me to reproduce your results from the article ( 96.6 % ). Have you trained it with a larger batch size?

mrT23 commented 2 years ago

96.1% is not bad.

in any case, i trained with: --learning_rate=5e-5 --batch_size=128

so i think you need lower learning rates. play a bit with larger batches, and reduced learning rates.

please share your code and training params once you are able to reproduce the results