EVA
EVA copied to clipboard
baaivision
EVA Series: Visual Representation Fantasies from BAAI
How do I add negative samples and train them in EVA02/det, I can't find where
Hello, I am fine-tuning EVA-02 LVIS checkpoints for custom few-shot object detection tasks, and it's performance is great. However, I hope it could be further increased if we use undertrained...
## Describe the bug Traceback (most recent call last): File "evaclip_infer.py", line 86, in main() File "evaclip_infer.py", line 61, in main image_features = model.encode_image(image) File "/mnt/EVA/EVA-CLIP/rei/eva_clip/model.py", line 297, in encode_image...
I met this problem，could you help me about it？
Have you tried how the S or Ti model works on COCO?
Hi, I found the hidden dimension of FFN in EVA-2 cannot be divided by 256 which would not be suitable for model parallelism. Would you mind releasing the model that...
I notice that you use separate linear projections for q, k, and v when `subln=True`: https://github.com/baaivision/EVA/blob/802e195104e67989777d4f696833368463046386/EVA-02/seg/configs/eva02/upernet/upernet_eva02_base_12_512_slide_60k.py#L45 https://github.com/baaivision/EVA/blob/802e195104e67989777d4f696833368463046386/EVA-02/seg/backbone/eva2.py#L233 https://github.com/baaivision/EVA/blob/802e195104e67989777d4f696833368463046386/EVA-02/seg/backbone/eva2.py#L290 I do not understand why. It should be equivalent to using single...
Could you provide the train config file of EVA-02-L on cocostuff164k which achieves 53.7 mIoU?
For example, Merged-38M MIM pretraining -> IN21K finetuning -> downstream task finetuning
Can you provide detailed Lion setting(learning rate and weight decay) in Table 18(EVA02 paper)?
Metadata
Owner
Metadata
EVA Series: Visual Representation Fantasies from BAAI