search

wielandbrendel / bag-of-local-features-models

Pretrained bag-of-local-features neural networks
Other
312 stars 45 forks source link

How can I train from scratch for bagnet. #5

Closed Cold-Winter closed 5 years ago

Cold-Winter commented 5 years ago

Great work. It give me a lot insights. Can you please release the training code (Get the BagNet by myself)?

wielandbrendel commented 5 years ago

Training the BagNets from scratch is very simple: just use the standard Pytorch ImageNet training code (https://github.com/pytorch/examples/tree/master/imagenet) and load the BagNet as model. My code is a bit more adapted to the infrastructure that we use so I'd guess this is the most direct way.

Cold-Winter commented 5 years ago

Thanks for your quick reply. By the way, can you please tell me how do inference during test time. Given an image, how can we get the logits of each class.

wielandbrendel commented 5 years ago

The models are standard Pytorch and Keras models, so they return logits be default.

xuzhang5788 commented 5 years ago

Is it only good for 2D images? Is it possible to extend to 3D voxel? Thanks for your great work.

ajtejankar commented 5 years ago

@wielandbrendel, The inference code creates all possible patches. Should this also be the case when training BagNet? Also, I still need a patch generation and heatmap averaging layer in addition to the above PyTorch script for training BagNet, right? The paper is deceptively simple (really cool)! So, I just want to make sure that I haven't misunderstood anything.

wielandbrendel commented 5 years ago

@xuzhang5788 The same approach is applicable to 3D voxel data. @chigur There is no patch generation. BagNets are normal ResNet-50 but with many 3x3 convolutions replaced by 1x1 convolutions, which limits the receptive field size of the top-most convolutional layer.

ajtejankar commented 5 years ago

Aah, I see. Thanks for the prompt clarification. I was confused because the generate_heatmap_pytorch method creates patches.

wielandbrendel commented 5 years ago

The generate_heatmap_pytorch function is only for the heatmap visualisation because there we want a denser sampling (normal BagNets have a stride of 8).

xuzhang5788 commented 5 years ago

@wielandbrendel Thank you for your response. Would you mind telling us how to apply to 3D voxel data?

wielandbrendel commented 5 years ago

The principle behind BagNets is applicable to 3D voxel data, but of course you would have to modify an architecture designed for 3D voxel data. In any case, this discussion is unrelated to this issue and so I am closing this thread for now. Feel free to open a new issue.

jennyzu commented 5 years ago

@Cold-Winter how did you manage to train from scratch? The code here (https://github.com/pytorch/examples/tree/master/imagenet) has very specific models you can train (Resnet etc.), how is it possible to train a custom model? I am new to pytorch, so I would very much appreciate your help. Thanks!

ksouvik52 commented 5 years ago

Hi, How are you doing the training in patches. Are you taking q x q pixels and then training them through Resnet-50 model? As you mentioned in your paper fig. 1, I thought you are doing training patch wise. If not then I dont understand what the summation block is doing during inference, before the softmax. i assumed they are the summation of logits for each class from all the patches. Please help me understand.

xuzhang5788 commented 5 years ago

@wielandbrendel Is your BagNets only for imageNet dataset? I want to get my own weights of the BagNets with my own dataset, so I need to train from scratch. Althogh you said that you changes several 3X3 conv to 1X1 conv in ResNet, I don't know which 3X3 conv. Could you please show your original model which is modified version of ResNet? Mnay thanks.

wielandbrendel commented 5 years ago

@xuzhang5788 You find the full model description in https://github.com/wielandbrendel/bag-of-local-features-models/blob/master/bagnets/pytorchnet.py.