edouardoyallon / scalingscattering Goto Github PK

hi ! thanks for your great job !
I debugede your code , but after this result:
"Total number of parameters: 12777726.000000"
the code stopped running,I waited for a few hours ,but still haven't any change, do you know the reason?

I run the CIFAR file try to train the CIFAR-10，but account this error，has anybody account？I didn't change any parameters

I tried using kymatio with this code. If you were interested in updating it, I switched the scattering method to kymatio but there are more things to be fixed, for instance, in STL, it seems that torch.legacy that you are using in utils has been deprecated and needs to be changed. In imagenet test, line torch.randn(8).cuda() again uses cuda, and when I comment .cuda() out, I fall into a rabbit hole of errors. the latest being

Traceback (most recent call last):
  File "/Users/Nshervt/Desktop/scalingscattering-master/imagenet/main_test.py", line 186, in <module>
    main()
  File "/Users/Nshervt/Desktop/scalingscattering-master/imagenet/main_test.py", line 65, in main
    model, params, stats = models.__dict__[opt.model](N=opt.N,J=opt.scat)
  File "/Users/Nshervt/Desktop/scalingscattering-master/imagenet/models/scatter_resnet.py", line 151, in scatresnet6_2
    model = ScatResNet(J,N)
  File "/Users/Nshervt/Desktop/scalingscattering-master/imagenet/models/scatter_resnet.py", line 61, in __init__
    self.bn0 = nn.BatchNorm2d(3*self.nfscat,eps=1e-5, momentum=0.9, affine=False)
  File "/Users/Nshervt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/batchnorm.py", line 34, in __init__
    self.register_buffer('running_mean', torch.zeros(num_features))
TypeError: zeros(): argument 'size' (position 1) must be tuple of ints, not float

Test [0,195/196] ; loss: 18.256 (17.631) ; err5: 98.75 (99.62) ; err1: 100.00 (99.91) ; data 0.013 ; time 1.792 [99.908, 99.622]

Training scripts for imagenet and SLE feature extractor will be added soon

I took a quick look at the models defined in each example and it seems that SLE is not implemented in anyone yet. I was wondering if it actually is not implemented or I can't find it.

First, congratulation for this very interesting work.
I would be interested in tested hybrid networks for 1D specific signals. When do you plan the new version you mention in pyscatwave description, including "1D-2D-3D fast, optimized, differentiable Scattering Transform" ?

Hello,

Thanks a lot for your generous sharing.

I want to use your proposed structure in some adversarial training framework.

My problem is that:

When batch size is 64, the previous data has the size of 64332*32

For J=1, after scattering, the matrix size is 64391616.
For J=2, after scattering, the matrix size is 6431788.
For J=3, after scattering, the matrix size is 6432544.
For J=4, after scattering, the matrix size is 6433322.

Looks like the tensor-shape for third dimesion is 1+J*8, the fourth and fifth dimension is 32/2^j.

I am not sure, is there a inverse operation API that I can restore the matrix after scattering to the original matrix, e.g., Tensor with size 64332*32.

It would be honor if you could answer the questions.
Thanks & regards!
Momo