Scattering 2D doesn't work when using 2^J == image size (TensorFlow backend)

Similar issues have been reported before (#284, #363) and also fixed (#412) for Torch backend. However, still not working in the tensorflow backend:

test:

import numpy as np
from kymatio.tensorflow import Scattering2D

scattering = Scattering2D(J=5, shape=(32, 32))
test_im = np.ones((1,1,32,32))
test = scattering.scattering(test_im)

Gives error:

      7 scattering = Scattering2D(J=5, shape=(32, 32))
      8 test_im = np.ones((1,1,32,32))
----> 9 test = scattering.scattering(test_im)

~/.local/lib/python3.9/site-packages/kymatio/scattering2d/frontend/tensorflow_frontend.py in scattering(self, input)
     48             input = tf.reshape(input, tf.concat(((-1,), signal_shape), 0))
     49 
---> 50             S = scattering2d(input, self.pad, self.unpad, self.backend, self.J, self.L, self.phi, self.psi,
     51                              self.max_order, self.out_type)
     52 

~/.local/lib/python3.9/site-packages/kymatio/scattering2d/core/scattering2d.py in scattering2d(x, pad, unpad, backend, J, L, phi, psi, max_order, out_type)
     13     out_S_0, out_S_1, out_S_2 = [], [], []
     14 
---> 15     U_r = pad(x)
     16 
     17     U_0_c = fft(U_r, 'C2C')

~/.local/lib/python3.9/site-packages/kymatio/scattering2d/backend/tensorflow_backend.py in __call__(self, x)
     27         paddings = [[0, 0]] * len(x.shape[:-2])
     28         paddings += [[self.pad_size[0], self.pad_size[1]], [self.pad_size[2], self.pad_size[3]]]
---> 29         return tf.cast(tf.pad(x, paddings, mode="REFLECT"), tf.complex64)
     30 
     31 def unpad(in_):

~/.local/lib/python3.9/site-packages/tensorflow/python/util/traceback_utils.py in error_handler(*args, **kwargs)
    151     except Exception as e:
    152       filtered_tb = _process_traceback_frames(e.__traceback__)
--> 153       raise e.with_traceback(filtered_tb) from None
    154     finally:
    155       del filtered_tb

~/.local/lib/python3.9/site-packages/tensorflow/python/eager/execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name)
     56   try:
     57     ctx.ensure_initialized()
---> 58     tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
     59                                         inputs, attrs, num_outputs)
     60   except core._NotOkStatusException as e:

InvalidArgumentError: paddings must be less than the dimension size: 32, 32 not less than 32 [Op:MirrorPad]

(speculation) So possibly problems with the order of the padding being different in tensorflow from torch.

Should also probably include some tests for these types of problems like the tests for implemented for Torch in #346

[MatthijsMars]

Ah yes you are completely right, I’m only working with square images so that’s why it doesn’t matter too much but might cause some trouble when not doing that. Thanks for pointing it out. ( I’ll edit the comment above for if someone ends up copying something from here)

[janden]

So what’s the status on this?