sense-x / mixmim Goto Github PK

请问，我在win系统下，from petrel_client.client import Client提示无法解析导入“petrel_client.client”，这个问题怎么解决？
安装也找不到这个库

It seems that your attention masking part is wrong, as the image patches from different images should have a different mask, while you use a single mask definition for all patches regardless of where they are from.

BTW, how do you cope with mixing more images? I see only mixing 2 images case in the code.

非常感谢您的研究，正在深入您的研究内容，请问是否能提供预训练权重，感谢！

Training with the default Settings, gradients always appear inf， is it normal？

Thank you for this great model!

Can you publish the self-sup pre-trained model on Imagenet?

我在代码里看见您在计算损失函数时，只计算重建的两张图像被mask区域的损失，对于这一点我没有异议。但是在target的设置上，我看您是直接是用源单张图像作为目标图像标签，而非源混合图像作为目标图像标签。不知道我描述得是否正确，还请您解惑

请问替换为其它backbone还需要加上absolute_pos_embed位置编码吗，因为我看SimMIM也采用swin，但是没有加absolute_pos_embed， 而且最新的一些transformer设计都已经用conv卷积来替换绝对位置编码了

Is there a visualization demo using pre-trained MixMIM models? Thanks.

Love the work. I would like to run MixMAE with PVT-V2, is the any way you could release the code for training other encoders (e.g. PVT-L in your paper)?

请问在下游任务上收敛速度比imagenet权重要慢很多是正常的吗，， ？
还有复现来看，感觉和simmim并没有明显的差距呢

hi，我注意到模型的attention部分存在 masked self-attention计算，如果用到原始swin中，它里面shifted window也会存在一个mask计算， 是不是会重叠不好部署呢？

在运行代码的过程中发现了两个小错误
add_weight_decay ——> param_groups_weight_decay

module ——> _modules

In your paper, you mentioned that you 'add two mix embeddings to the visual tokens to implicitly differentiate the two mixing groups' and 'use different mix embeddings for the 4 stages of the encoder', which is in section 2.3. However, it appears to me that there isn't such mix embedding in the function forward_encoder of models_mixmim.py. So, if it's my silly mistake that I didn't find it, would you kindly point it out to me? Thanks!

Hello,I can‘t find the implement of the mixed images input .Can you tell me where is ?

你好，请问您用的16张GPU是什么型号内存 呢，我只有2x48G的两张显卡可以复现结果么

你好，请问可以分享一下MixMIM以Swin-B作为backbone在预训练过程中得到的权重吗？包括Swin-B编码器的权重和Transformer解码器的权重，非常感谢！

What's the hardware & training time? Specifically, I'm interested in the statistics below:

• Arch & Epoch: [e.g. ViT-B, 300 ep]
• Hardware: [e.g. single 8 V100 node]
• Batch time: [e.g. 0.8s]
• Epoch time: [e.g. 12 mins 30s ]
• Training time: [e.g. 28 h]

Hello, great work!
Can you release the pretrained weights of your MixMIM-B?