Difficulty calculating model gradients, shared variables with static values, and sparse matrices

[lucianopaz]

@Ian_Czekala

# flatten the RFFT output appropriately for the interpolation, taking the real and imag parts separately
  vis_real = rfft[0, :, :, 0].flatten() #
  real values vis_imag = rfft[0, :, :, 1].flatten() # imaginary values

  # interpolate the RFFT to the baselines by a sparse matrix multiply
  interp_real = theano.sparse.dot(C_real_sparse, vis_real)
  interp_imag = theano.sparse.dot(C_imag_sparse, vis_imag)

  # condition on the real and imaginary observations
  pm.Normal("obs_real", mu=interp_real, sd=noise, observed=real_data)
  pm.Normal("obs_imag", mu=interp_imag, sd=noise, observed=imag_data)

I think that your error comes from where you define the observations. interp_real and interp_imag are both arrays, from what I could follow, so your random variables should have a shape compatible with the supplied parameter mu. This means that you should add a shape=something to the pm.Normal("obs_...) RV creation. Regrettably, we can’t handle symbolic shapes, so shape=interp_real.shape wont work. You will have to write down the proper shape tuple that depends on the shapes of your shared arrays’ alpha and delta.