zhaominyiz / rfda-pytorch Goto Github PK

您好，非常感谢您的工作。我在看您的论文的消融实验时我有些疑惑，为什么只加RF模块，指标甚至比不加的效果还要差，这是什么原因？只加dsta，指标好像已经很高了，即便又多加了RF和two-stage，psnr只有0.02的提升。如果用RF和DSTA，是不是two-stage就没那么重要了，我看表中并没有列出RF和DSTA的组合结果，期待您的解答，非常感谢🙏

In your paper, you train the RFDA using the two-stage training strategy, but we can not find the two stage training files in the release code. So we are very confused that how to achieve the two-stage training strategy. Could you provide the training strategy files in detail?

Thank you very much!

I am trying to run the code by following the instructions
error: '/usr/local/cuda/bin/nvcc' failed: No such file or directory: '/usr/local/cuda/bin/nvcc'

Please help me

hello，I'm interested in your great work，could you release the training code so that i can reproduce the result for my thesis. THX!

您好，我不太懂视频压缩。对您论文里的率失真曲线不太懂。我看您是取了测试集里的十八个视频中的一个，但是您的这个比特率是怎么算出来的？是根据不同的qp设置，得到的几个值吗？如果是这样，那您用的是1s多少帧，这个是不是涉及到和传输相关的码率控制的问题？即便是固定qp，比特率也会受别的影响吧，不同的图片编码出来的比特数也不一致吧？期待您的回复，万分感谢！🙏

您好，非常感谢您的工作！您的brf版本我看到又有改进，但是代码是不是还没开源？您的这个改进版本工作是不是准备发论文的？期待您的回复，非常感谢！

Hello,

Thank you for releasing the training code. Could you please provide the instructions for training?

I tried to use the instructions for training STDF - python -m torch.distributed.launch train_rf_ft.py --opt_path config/Final_QP37.yml and ran into an error with the STDF module's in_conv layer:
Expected 4-dimensional input for 4-dimensional weight [32, 7, 3, 3], but got 5-dimensional input of size [8, 15, 1, 96, 96] instead
Could you please tell me what radius should be specified in the config file for training? The currently released config file specifies a radius of 7 but from the error log, it seems the training code expects R=3.

Additionally, your paper mentions two-stage training where you train the STDF with radius 3 and then the full network with 15 frame clips. Does your training code support doing that?

Thank you!