EVA

EVA copied to clipboard

Position Encoding of EVA-2

1

Hi, I found the EVA-2 adopts both learnable positional embedding and rotary embedding at the same time. Why do you choose such design? Thanks.

vateye

Hello！Here import accimage import display can not be resolved, how to solve this problem ，I hope to hear from you.Thanks. # TODO: specify the return type def accimage_loader(path: str) ->...

zhanglaoban-kk

ImageNet Classification Fine-tuning

4

Sorry to bother you. I met a mistake during fine-tuning on EVA-01. RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates...

LITTLEKKKK

the .sh script of Evaluate the fine-tuned EVA (336px, patch_size=14) on ImageNet-1K val with a single node (click to expand) can not execute.

1

(eva) root@nexus-nyz:~/EVA/EVA-01/eva# bash eva.sh /root/miniconda3/envs/eva/lib/python3.8/site-packages/torch/distributed/launch.py:178: FutureWarning: The module torch.distributed.launch is deprecated and will be removed in future. Use torchrun. Note that --use_env is set by default in torchrun. If your...

peter-ni-noob

Where can I find the script to test retrieval zero-shot performance?

2

Thanks for the impressive work. By the way, where can I find the script to test retrieval zero-shot performance?

1338199

Hello! What is the difference between EVA-01 and EVA-02? Best wishes, Egor

Egorundel

Hi, thanks for the good EVA-clip. I'm loading&training the EVAVisionTransformer in EVA02-CLIP-L-14 and find that setting the patch_dropout to 0.5 raise runtimeError The size of tensor a (128) must match...

chenmc1996

hello，thanks for your works, but when i run this work in deepspeed, the GPU memory can not reduce, thank very much if you can tell me why.

shuowoshishui

Every time I try to run _pip install -r requirements.txt_, where would be a mistake _AssertionError: Unable to pre-compile sparse_attn_, even when I download and install the triton compiled windows...

JeavanCode

Difference between eva-clip model and MIM pretrained model

2

I find that the eva-clip model has an extra inner_attn_ln layer compared to the original pretrained model. [EVA-CLIP](https://github.com/baaivision/EVA/blob/master/EVA-CLIP/rei/eva_clip/eva_vit_model.py#L164) [EVA](https://github.com/baaivision/EVA/blob/master/EVA-02/asuka/modeling_finetune.py#L188) Any reason to use another layer-norm in the clip model?

jihaonew