Settings for experiment with ViT-B/32 and ViT-L/14

[andylin12]

Thanks for the wonderful paper and repo.

I was able to reproduce MaskClip and MaskClip+ with ViT-B/16 + R101 on Pascal context dataset. The result mAp is 25.45 and 29.48 respecitively.

However, when I tried to change the model to ViT-B/32 and ViT-L/14 the result is not good, less than half of ViT-B/16 and the quanlitative result shows that the predicted dense label is generally a mess.

What I did was:

1. convert weight and backbone and extract text embeddings for ViT-B/32 and ViT-L/14
2. create a config accoding to ViT-B/16, with modifications:
   • change the patch size to 32 for ViT-B/32
   • change the pathc size to 14, embed_dims to 1024, num_layers to 24 for ViT-L/14

Is there anything I've done wrong or misunderstood? Do you have any suggestions on why the result is bad?

Thanks in advance.

[hewenbin]

Same observation here. Any thoughts?

[111chengxuyuan]

Hello,I want to ask you a question,Which version of mmsegmentation should I install to run this code properly? I installed 0.20.0 but couldn't run it

[ngfuong]

Can you please share with me the configuration file of your reproduced MaskCLIP+ ViT16?