ylsung / ecoflap Goto Github PK

Hello, I tried to run ecoflap_first.py, but the eval results are much less than the reported values.
I noticed that the sparsity_ratio_granularity config is not added in ecoflap_first.py, thus the model is pruned with fixed sparsity 0.5. Then I added this config, but the results became even less. I wonder if there is any other modification I need to do with the code to regenerate the experimental results.
As for the difference of experimental results, take task vqav2_zeroshot_flant5xl_eval for example:

1. with raw code ecoflap_first.py: 57.64
2. with sparsity_ratio_granularity added as block: 54.56
3. reported in paper: 59.4

Thanks a lot for your time.

Hello, your work has aroused great interest in me, which makes me really want to delve deeper into understanding and reproducing the pruning process through code. Therefore, I wonder if the code in this repository is complete. Unfortunately, I noticed that at least the lavis.compression.modify_model_with_weight_init module, mentioned in train.py, is not included in the repository.
Is this part necessary for pruning? Or, are you willing to release the complete code?

Hello, I found only blipt5_wanda_pruner in this repository, is it feasible to pruning BLIP(that is, with Bert as the backbone of language modal)? Or, is pruner for BLIP available?

Is the code for the llama part complete?
Is the code ECoFLaP used on LLM provided?

很感谢您的工作，但是在论文中，我对p的含义很混乱，在您文中提到的，p大概代表需要剪掉的比例，那也就是说零阶梯度越大，越应该被剪掉

只有这样那个s.t.才成立，但是如果p代表剪掉的比例，那么N_select就代表全局保留个数，那为什么s小的，我保留的就少呢，这不是和前面矛盾了嘛

Hello

Do you intend to release the checkpoints of the pruned BLIP-v2/ CLIP models via EcoFLap? Thanks!