TypeError: 'module' object is not callable

[marcogallotta]

Getting the following error using v0.2 on Ubuntu 20.10 using PyTorch. I was getting the same kind of error in a different line using the latest dev I built yesterday.

All I'm really doing right now is:

self.scattering = Scattering1D(J=8, shape=T, Q=12) 
out = self.scattering.forward(X)

where X is a 2D tensor.

  File "./torch_net.py", line 147, in <module>
    output = net(inputs)
  File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "./torch_net.py", line 26, in forward
    out = self.scattering.forward(X)
  File "/usr/local/lib/python3.8/dist-packages/kymatio/frontend/torch_frontend.py", line 23, in forward
    return self.scattering(x)
  File "/usr/local/lib/python3.8/dist-packages/kymatio/scattering1d/frontend/torch_frontend.py", line 114, in scattering
    S = scattering1d(x, self.backend.pad, self.backend.unpad, self.backend, self.J, self.psi1_f, self.psi2_f, self.phi_f,\
  File "/usr/local/lib/python3.8/dist-packages/kymatio/scattering1d/core/scattering1d.py", line 78, in scattering1d
    U_0_hat = fft(U_0, 'C2C')
  File "/usr/local/lib/python3.8/dist-packages/kymatio/backend/base_backend.py", line 58, in __call__
    return self.fft_forward(x, direction=direction, inverse=inverse)
  File "/usr/local/lib/python3.8/dist-packages/kymatio/backend/base_backend.py", line 53, in fft_forward
    output = self.fft(x)
  File "/usr/local/lib/python3.8/dist-packages/kymatio/scattering1d/backend/torch_backend.py", line 134, in <lambda>
    fft = FFT(lambda x: torch.fft(x, 1, normalized=False),
TypeError: 'module' object is not callable```

[dminskiy]

I have the same issue with torch 1.8.1, while it all works fine with torch 1.7.1
torch.fft was depreciated in this version of torch: pytorch/pytorch#42175
Annoyingly, when you try to use new versions of these functions like this:

fft = FFT(lambda x: torch.fft.fft(x, 2, norm='backward'),
          lambda x: torch.fft.ifft(x, 2, norm='backward'),
          lambda x: torch.fft.irfft(x, 2, norm='backward', onesided=False),
          type_checks)

Then you get this error:

  File ".../kymatio/scattering2d/core/scattering2d.py", line 22, in scattering2d
    U_1_c = cdgmm(U_0_c, phi[0])
  File ".../kymatio/backend/torch_backend.py", line 193, in cdgmm
    raise TypeError('Input and filter must be of the same dtype.')
TypeError: Input and filter must be of the same dtype.

This error is caused by the fact that after torch has moved to these updated functions they return torch.complex64 as opposed to previously used torch.float32

My question to authors would be: is there a quick fix to this issue?

[OverLordGoldDragon]

quick fix

Use master branch, comment out every "sanity check" that breaks code, it should work same. Just ensure input-output behavior of whatever you're changing remains the same (check types etc).

dev has lots of stiching together to do, next release won't happen anytime soon.

[dminskiy]

@OverLordGoldDragon Thank you for a quick response!
Looks like I was using an old version of the library. It seem to work fine when I updated to a newer release.

[OverLordGoldDragon]

Fixed in dev, set to merge with master.

[Vin-itall]

Hey, I have still facing this error. pip install doesn't work, are there any updates?

[OverLordGoldDragon]

@Vin-itall Please provide the following: 1) minimal code to reproduce the error; 2) output of below script:

import kymatio, os, inspect, importlib

def print_substring_in_dir(backend_dir, s="fft"):
    backend_files = [p.rstrip('.py') for p in os.listdir(backend_dir) 
                     if '__' not in p]
    
    backend_modules = {}
    for f in backend_files:
        try:
            mname = backend_dir.replace('\\', '.').split('kymatio')[-1]
            m = importlib.import_module('kymatio' + mname + '.' + f)
            backend_modules[f] = m
        except:
            pass
    
    print(backend_dir)
    for name, m in backend_modules.items():
        print("\n== {} {}".format(name, "="*80)[:80])
        txt = inspect.getsource(m).split('\n')
        for line in txt:
            if s in line:
                print(line)

print(kymatio.__version__)
_dir = os.path.dirname(os.path.abspath(kymatio.__file__))
print_substring_in_dir(os.path.join(_dir, 'scattering1d', 'backend'))
print_substring_in_dir(os.path.join(_dir, 'backend'))

[Vin-itall]

Thanks for the quick reply, turns out I was running version 0.2.0. Reinstalled kymatio, it works fine now. Thanks and sorry for any unnecessary inconvenience caused.