Performance on Resnet101 network about simple-faster-rcnn-pytorch HOT 7 OPEN

I think maybe resnet101 is difficult to train.

This maybe helpful.

For Resnets, we fix the first block (total 4) when fine-tuning the network, and only use crop_and_resize to resize the RoIs (7x7) without max-pool (which Xinlei finds useless especially for COCO). The final feature maps are average-pooled for classification and regression. All batch normalization parameters are fixed. Learning rate for biases is not doubled.

from simple-faster-rcnn-pytorch.

@chenyuntc Thanks, I also fixed the weights for top layers, but the result didn't improve. As you mentioned, it might be the reason of BN and biases. I'll have another try.

from simple-faster-rcnn-pytorch.

Hi @chenyuntc, I've trained the model with:

1. Fix the first block.
2. Learning rate for biases is not doubled.
3. All batch normalization parameters are fixed.
4. Use 1e-4 as weight decay.

And I restrictedly followed the way of training as I did in caffe, while it seems the performance cannot be improved.
Have you tried to train the model on networks other than VGG16?

from simple-faster-rcnn-pytorch.

Actually, I only tried VGG16.

from simple-faster-rcnn-pytorch.

I recently also want to implement resnet structure based on this project. And I found your @twtygqyy codes are very helpful to me. But I have a question about batch normalization. Why we need to fix batch normalization parameters here?

from simple-faster-rcnn-pytorch.

@blateyang BN only work when batch_size>1 and only work well when batch_size>=16.You can see this paper https://arxiv.org/abs/2002.05712.However，@chenyuntc code only surport batch_size==1

from simple-faster-rcnn-pytorch.

from simple-faster-rcnn-pytorch.