ECO-pytorch
ECO-pytorch copied to clipboard
Initialization models and accuracy@1 in pytorch
Thanks for your work! I want to reproduce this paper,But I have some questions about Initialization models and accuracy@1 Q1: does the models use Initialization models in paper if you just set args.pretrained_parts == "both", In other words, weight_url_2d='https://yjxiong.blob.core.windows.net/ssn-models/bninception_rgb_kinetics_init-d4ee618d3399.pth' and models/C3DResNet18_rgb_16F_kinetics_v1.pth.tar is your Initialization models in your paper ECO-pytorch Q2: is it normal that baseline is 53.4% when i set args.pretrained_parts == "scratch" and num_segments=16 , could you provide some information about your model accuracy@1 when you set pretrained_parts='scratch', '2D', '3D', 'both','finetune' and segments=16 i am new learner in this, If you can help me, I will be very grateful. Looking forward to your reply @zhang-can
@chen-ming2019
Hi czl, Do you run this code successfully?i want to ask you some question.
i have reproduced this code, can i talk about this code with you?QQ726421577
@Fovever1 你说的哪个数据集?
UCF101,我不明白为什么我的是top1=76%,top5=92%,loss=1.0,到第十个epoch的时候,test loss就降不下去了,我是完全按照您的要求来的,为什么结果不一样,我是clone原作者的pytorch代码
将视频转成帧也是用的TSN的Denseflow提取的,预训练模型用的不是eco_lite_rgb_16F_kinetics_v3.pth,用的是ECO_Lite_rgb_model_Kinetics.pth
@chen-ming2019
@Fovever1 你这个损失还挺大的,应该是还没有完全收敛,
@chen-ming2019 您觉得是什么原因造成的,能说下您的具体细节吗
@Fovever1 3好像是新的,应该会好些,我之前用的是2
@Fovever1你最后面全连接层的维度有改吗,他代码上默认的好像是400
@chen-ming2019 我赶紧看一下
@chen-ming2019 可是num_class不就是从这传下去的吗?是101,不是400呀,您说的是指的哪个py文件里的,model?吗
@Fovever1对的,是我记混了,那就可能是你参数设置的问题了,这个不太好说
参数设置我是完全按照opt.py里面来的?您当时还调了其他参数吗?
@Fovever1你把你参数发来,我看看
因为我要在这个实验上继续优化,发论文才能毕业,现在测试情况不太好,所以希望能得到您的一些帮助,定重谢!您的微信和QQ是多少?
@chen-ming2019 发我邮箱一下[email protected]
785432809 QQ
@Fovever1 @chen-ming2019 大佬们,我想请教下,我也是在ufc101上训练,最后top1是75%,请问还要调整那个参数吗,thanks in advance~,我直接按照readme跑的ECO_Lite_rgb_model_Kinetics