Yuan xinbin comments

Results 8 comments of


                                            Yuan xinbin

The mAP of SSD-Resnet is lower than SSD-VGG16.

#我想问一下大家在训练的过程有碰到AP为0的情况吗，我是用的自己的数据集训练的。

About the learning rate setting of p_conv and m_conv

> You're right! I'll fix it. it seems this bug has not fixed yet

How to inference with batch_size（batched）

i have the same question too, has somebody solved this problem yet?

loss.backward()

> I got the same issue. After changed inplace=False, it returned `nan C: nan S: nan` while training i have got this problem too, have you solve it yet?

当我在训练MMrotate时，在config配置文件中加入optim_wrapper = dict( type='OptimWrapper', optimizer = dict( type='AdamW', lr=0.0002, #/8*gpu_number, betas=(0.9, 0.999), weight_decay=0.05), accumulative_counts=2)时，发现optim_wrapper并没有被解析，用的是schedule_1x.py里面的默认optimizer， optimizer = dict(type='SGD', lr=0.0025, momentum=0.9, weight_decay=0.0001)，相关库的版本： mmcv-full:1.7.0 mmengine：0.10.4 mmrotate:0.3.4 在MMrotate中怎样正确使用梯度累积呢

visual grounding poor performance after fine-tuning

我也是同样的问题，微调之后效果变得很差，我用的llama-factory 去微调Qwen2.5-VL-7B,微调之前模型的效果还凑合，微调之后完全不行，给出的grounding坐标全是错的

Yuan xinbin