search

guozix / TaI-DPT

MIT License
85 stars 7 forks source link

Some questions about reproducing "DualCoOp" #10

Closed iamxiaoyubei closed 10 months ago

iamxiaoyubei commented 10 months ago

Hi, zixian, I saw that the performace of reproducing DualCoOp(DualCoOp*) is better than that in the DualCoOp paper, as shown in the table below.

image

So want to use your DualCoOp for customized data. However, I saw that you obtained the output_g during training forward(https://github.com/guozix/TaI-DPT/blob/1333ecaa32bfffb4f2eb916f5532afb88ac457fe/trainers/Caption_dual.py#L398) but did not use it in the dualcoop_loss(https://github.com/guozix/TaI-DPT/blob/1333ecaa32bfffb4f2eb916f5532afb88ac457fe/trainers/utils.py#L151). So I want to ask whether output_g will be used in training? Is the result of the DualCoOp generated by the current code using the following command? bash main_dual.sh nuswide_partial rn101_nus end 16 True nuswide_partial_dualcoop_448_CSC_p0_1 0.1 2

guozix commented 10 months ago

Yes, only output is used to compute ASL loss, according to the DualCoOp paper. Since the author of DualCoOp has released their code, I recommend you use the official implementation :)