-
Notifications
You must be signed in to change notification settings - Fork 68
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[MobRecon] ResNet50 backbone reproduce #65
Comments
Hi, ResNet50的改造看上去没有问题。可能有两个地方可以做修改(非必需):
image size变化后,不需要对verts,joint_img做额外修改。为确保数据没问题,可使用以下代码做可视化 HandMesh/mobrecon/datasets/freihand.py Line 345 in dbdca0d
|
感謝!! 我嘗試後再將結果分享上來 非常感謝 |
期待你的结果!在学习如何复现resnet50 backbone... |
Hi Xingyu,
模型權重部分,與 ResNet50 同樣的部分( 兩個 Stack 中縮小的部分 - 雖然絕對不會像 ReaNetStack 修改:ResNetStack 放大部分的 layer4 ~ layer1,我是用與 ReaNet50 的 layer 完全相同的 BottleNeck 大小 註1,在最後再接上 註1: ResNet 中,各 layer 的第一層因為調整 resolution 與 channel 數,與後續幾層有不同的設計 Code 部分:重建 ResNet_Stack commit 架構圖呈現如下:與原版 ResNet50, DenseStack 不同的部分用紅色與黃色特別標明 感謝!! |
你好,my_resnet.py能提供一下吗?或者是resnetstack.pth |
啊! 其實 Bottleneck 跟 所以可以直接改從 |
好的,成功了!感谢你! |
非常感謝!!! 我再嘗試看看 |
@clashroyaleisgood @SeanChenxy |
你好,我已收到你的邮件,我会及时处理^_^
|
@brian208579 關於下面給作者的問題,有個應該可以幫忙回答,有誤再麻煩作者更正了
我自己在做實驗的時候,曾經將 Backbone 輸出的 feature 經過 detach( 將 gradient 截斷 ) 後再往後傳,最後的訓練成果蠻糟糕的,而將 detach 取消掉後就恢復正常了 由此推測 Backbone 所需的 Loss 除了原先的 2D joint loss 以外,也需要整個模型的 3D Loss |
非常感謝您的回覆!你提供的資訊對我有非常大的幫助,我會再加以嘗試著墨,在這邊預祝您論文與口試都順利! |
@clashroyaleisgood sorry to bother you, could you tell me how many epoches did you use when you train on the resnet50? I also tried on resnet50, but the outcome is worse. |
Hi @Shirley-olivia, I don’t remember clearly, but it should be the same as monrecon. Maybe you can ask author @SeanChenxy for correct answer. |
Thanks a lot! I did not pretrain ResNet50 backbone, and that may be the reason for my bad outcome.And I notice that the author pretrain the model based on the synthesized data, which include 2D pose position and heatmap based on L1 loss. Did you pretrain like this or could you share the way you pretrain resnet50? |
Hi Xingyu,
我想請教關於重建 ResNet50 Backbone 的問題,以下是我照著 code 畫出來原版 Backbone 架構圖,以及我猜測的 ResNet50 Backbone 架構
請問這樣的架構正確嗎? 還是其實有某些地方是有問題的?
做的更改有
HandMesh/mobrecon/configs/mobrecon_ds.yml
檔中新增或是直接修改HandMesh/mobrecon/configs/default.py
latent
uv_reg
另外想請問: 在 image 大小的部分,是調整好 cfg.DATA.SIZE 就 OK 了嗎? 還是在 Dataset 裡面的 verts, joint_img 等也需要做更改嗎?
感謝!!!!
The text was updated successfully, but these errors were encountered: