Silentssss

Results 6 issues of Silentssss

I cant download KsponSpeech dataset on AIHhub net. Can you teach me how to download, thanks!

![1695629181138](https://github.com/Sierkinhane/CRNN_Chinese_Characters_Rec/assets/108161275/d6ee8618-98c2-4593-bba6-e6a2d041de23) 在修改完8P后出现以上报错,于是我在网上寻找解决办法,发现反向传播时没有设置loss.requires_grad_(True),于是在以下地方添加了改代码: ![image](https://github.com/Sierkinhane/CRNN_Chinese_Characters_Rec/assets/108161275/b317b391-dfec-43f9-9887-53aaebddd320) 修改完后虽然可以跑通,但是loss却一直不收敛,如下图,请问有什么解决办法可以跑8P并使其收敛,不然训练步数太多,单P跑的太久了: ![image](https://github.com/Sierkinhane/CRNN_Chinese_Characters_Rec/assets/108161275/0c630e83-00b1-409d-934c-2e34cfcf7bab)

我使用ucf101数据集先训练了1个epoch,得到的loss曲线是这样的,非常不平滑,请问这是正常的吗,下面是我的结果和训练参数: `torchrun --standalone --nproc_per_node=8 train.py \ --model VDiT-XL/1x2x2 \ --use_video \ --data_path preprocess/ucf101_train.csv \ --batch_size 1 \ --num_frames 16 \ --image_size 256 \ --frame_interval 3 \ --epochs 1` ![image](https://github.com/NUS-HPC-AI-Lab/OpenDiT/assets/108161275/b88ff588-c89a-46d4-b42d-5b30d0c8ef8f)

你好,有个flash-atten的问题想请教下,当我想使能flash-attn时,我发现以下图1的逻辑根本走不进去,为此我打印了self.using_flash、attn_bias、qkv.dtype,最后发现attn_bias一直不是None(图2) 图1: ![image](https://github.com/FoundationVision/VAR/assets/108161275/2e65feea-0f95-4035-8bae-887ee2038931) 图2: ![image](https://github.com/FoundationVision/VAR/assets/108161275/e56d39da-8f12-4471-80cc-6431b94f85f8) 于是我将代码修改成以下逻辑: `using_flash = self.using_flash and attn_bias is None and qkv.dtype != torch.float32` 修改为 `using_flash = self.using_flash and qkv.dtype != torch.float32` `assert attn_bias is None and...

![image](https://user-images.githubusercontent.com/108161275/177292985-b6954fd2-d75b-4fa2-bc77-a57995e47387.png) ![image](https://user-images.githubusercontent.com/108161275/177293645-e093b1a7-6850-4a2a-95b8-2f2ccf24b239.png) I have installed pynd==0.3.6,why report this error?please

![image](https://github.com/CircleRadon/Osprey/assets/108161275/1cc5471e-30c7-4093-b1f2-5f18b6bf0541) Hello,vg images link is invalid. Could you tell me how to download it?