search

NetEase-GameAI / ResNeSt-caffe

A Caffe version of official PyTorch ResNeSt
27 stars 7 forks source link

train #3

Open mxy519 opened 4 years ago

mxy519 commented 4 years ago

Hello, I want to ask what is the meaning of caffe environment need permute layer from ssd-caffe ,and if i want to train the network with my own datas what should i do? Looking forward to your reply,thanks.

NetEase-GameAI commented 4 years ago

If you use bvlc-caffe (https://github.com/BVLC/caffe), you need manually add permute layer from ssd-caffe (https://github.com/weiliu89/caffe/tree/ssd), since no Permute layer in bvlc-caffe but ResNeSt need it. If you wanna train with your data, just do as other networks. For example, change the Input layer to Data layer to use LMDB/LevelDB, and add a SoftmaxWithLoss layer in the end.

mxy519 commented 4 years ago

Can I use the ssd-caffe directly without using bvlc-caffe?

NetEase-GameAI commented 4 years ago

Theoretically yes.

mxy519 commented 4 years ago

Thank you. Because my classification task is very simple, I only intercepted the part of resnest50 network, and did not use the pre training model during training, resulting in the network not converging and the loss has been very high. What should be done with this?

mxy519 commented 4 years ago

I cannot download the resnest50 caffemodel, can you share it with baidu cloud

NetEase-GameAI commented 4 years ago

Thank you. Because my classification task is very simple, I only intercepted the part of resnest50 network, and did not use the pre training model during training, resulting in the network not converging and the loss has been very high. What should be done with this?

We did not train resnest models, but only converted them from pytorch to caffe. For simple task, you could finetune pretrained resnest50.

NetEase-GameAI commented 4 years ago

I cannot download the resnest50 caffemodel, can you share it with baidu cloud

We have no Baidu account, but here is a personal link: https://pan.baidu.com/s/13rTjk-q1JJMNj30x3L8vHQ (code: ucpb)

mxy519 commented 4 years ago

Thanks. When I finetune the network, the convergence speed of the network is very slow. I want to ask you how to select the parameters when training the network, such as how to set the learning rate, momentum, weight decay coefficient.

NetEase-GameAI commented 4 years ago

We have no experience in training ResNeSt. Use the official repo for training help: https://github.com/zhanghang1989/ResNeSt

mxy519 commented 4 years ago

I want to convert the pretraining model in pytorch(.pth) into Caffe model(.caffemodel). What should I do