胡钧耀

我也有这样的问题，我在使用常用的计算FLOPS的库（torchprofile、ptflops、thop）时，三种方法结果均有所不同，请问如何公平比较呢？ - torchprofile计算结果偏高，对GroupNorm、matmul等进行了FLOPS计算 - ptflops计算结果将乘积和加法视为两个操作 - thop对GroupNorm、matmul等没有进行计算 。。。感觉非常乱。

I trained on resolution 64 for KTH dataset and have the same problem. The possible reason is that training is difficult if the resolution is too high; If the resolution...

me too, have you solved it?

pt文件传输过程中，没有传输完整，重新下载pt文件试试，我尝试过没有问题，你也可以从这里下载，两个选一个下载即可 https://www.dropbox.com/s/ge9e5ujwgetktms/i3d_torchscript.pt https://onedrive.live.com/download?cid=78EEF3EB6AE7DBCB&resid=78EEF3EB6AE7DBCB%21199&authkey=AApKdFHPXzWLNyI

> I also encountered the same problem during the reproduction process. The weight of the I3D model is downloaded from the link you gave. And put the file under the...

Are you living in China? I have just tested that the link can be downloaded. If you live in China, you should use the ladder (VPN) to download the file....

> I can go in and it will show: File too large to preview. But clicking download doesn't reflect anything. > > > Are you living in China? I have...

Hello, the following is my understanding. The approach is really inconsistent, and that's really a potential problem. However, the calculation of metrics such as FVD may be aligned with the...

您好，我的看法是这样的。 作者使用的是基于朗之万动力学NCSN扩散模型。原论文在设置参数的时候，开始参数大于结束参数。betas按列表顺序由大到小。在降噪过程中，betas应该是由大到小，列表索引应该是0-999。X_T是真实图像。   https://github.com/voletiv/mcvd-pytorch/blob/451da2eb635bad50da6a7c03b443a34c6eb08b3a/configs/kth64_big.yml#L81-L82 https://github.com/voletiv/mcvd-pytorch/blob/226a3fd1601a6fde1c59d01fb22f82cb37b3b8c4/models/__init__.py#L24-L26 https://github.com/voletiv/mcvd-pytorch/blob/226a3fd1601a6fde1c59d01fb22f82cb37b3b8c4/models/better/ncsnpp_more.py#L736-L739 https://github.com/voletiv/mcvd-pytorch/blob/451da2eb635bad50da6a7c03b443a34c6eb08b3a/models/__init__.py#L267 作为对比： 在DDPM原论文中，开始参数小于结束参数。betas按列表顺序由小到大。降噪过程时间步从列表索引999到0。X_0是真实图像。   https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/scripts/run_celebahq.py#L132-L137 ``` def train( exp_name, tpu_name, bucket_name_prefix, model_name='unet2d16b2c112244', dataset='celebahq256', optimizer='adam', total_bs=64, grad_clip=1., lr=0.00002, warmup=5000, num_diffusion_timesteps=1000, beta_start=0.0001, beta_end=0.02, beta_schedule='linear',...

Hello! Which version of torch do you use? The version in my env is 1.13.1. Let me know your env then I will try to repeat your issue. > ubuntu@ubuntu:~/hjy/code/common_metrics_on_video_quality$...