VideoMoCo

VideoMoCo copied to clipboard

Hi, Thanks a lot for your work. I am trying to train the model on a custom dataset and am curious if you could please share the logifles for the...

shashankvkt

The code does not work for ucf

pikeyang

Backpropagation through a masking operation does not seem easy, because you are essentially applying a non-differentiable operation onto a subset of frames. In this implementation, even though a list of...

basilevh

您好，论文中的公式（6），随时间衰减的因数t是乘在exp的外面的； 但是在代码中（moco/builder.py，第268行），这个因数weight是直接乘到了队列中的特征上； 按照代码中的实现方式，公式中分母的因数t是不是应该乘在k_i的前面？

weeknan

OOM error when using more than 24 GPUs

1

It's a very intresting work, but I was wondering how many GPUs when you're training videoMoco? we found if use more than 24 GPUs(v100 32G), the training processing would be...

rtygbwwwerr

train D for initialization

1

Hi authors, would you show more details about take 100 epochs to train D for initialization，does it use G but not update?

cobraheleah

Hi authors, If yes, can you show me some demo code for such a downstream task?

napohou

Hi, thanks for releasing the code and r2+1D model. Can you please share the kinetics pretrained 3D-ResNet18 weights too? I am comparing 3D-ResNet18 models from other self-supervised methods too in...

fmthoker

The paper say the architecture of G follows ConvLSTM, but in builder.py the Generator_Mask is different from ConvLSTM, why do you make this change？Could you explain the design of Generator_Mask...

tjuspAIC