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

Training Pefromance of ZSL code ,final mAP with tresnet-m is 15.6 #30

Closed aliman80 closed 2 years ago

aliman80 commented 2 years ago

Hi, I have tried to train the tresnet-m model on nus-wide dataset but i got mAP of 15.6 after 40 epochs. May i ask you suggestion what am i doing wrong here , i didn't change any code. If you can answer this will help a lot Regards

giladsharir commented 2 years ago

Hi @aliman80 , can you please share the training logs/outputs ? Which model_path did you use for tresnet-m ?

aliman80 commented 2 years ago

Thank you very much for your response, I tried to use tresnet-m but it was giving me error that my saved model and current model dimensions mismatch so i trained it from scratch. Can you guide what should i don next.

aliman80 commented 2 years ago

RuntimeError: Error(s) in loading state_dict for TResNet: size mismatch for head.decoder.duplicate_pooling: copying a param with shape torch.Size([80, 768, 1]) from checkpoint, the shape in current model is torch.Size([768, 1]). size mismatch for head.decoder.duplicate_pooling_bias: copying a param with shape torch.Size([80]) from checkpoint, the shape in current model is torch.Size([1]). This is the error which i obtain , Pleae guide where i am messing up

giladsharir commented 2 years ago

thanks for reporting this issue. We fixed it, and now you can train tresnet-m with : --model-path=https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/tresnet_m_miil_21k.pth

aliman80 commented 2 years ago

Thank you so very much, for your response and support Thanks again

aliman80 commented 2 years ago

Hi, Thanks for your continuous support, I have been able to run the ZSL file but I am receiving these results;

mAP score regular 1.14, mAP score EMA 1.14 current_mAP = 1.14, highest_mAP = 1.14

aliman80 commented 2 years ago

Thank you , i used wrong lr that's why it caused me error.

Thanks again for your support