You’ve got your answer right there:
coeffs_d_2[1][1].size:4489
coeffs_u_4.size:4624
This happens when you end up with non-integer dimensions when downsampling, and pywt.dwt2 rounds it up.
for example, you have an image and data.shape = (250, 400).
cA1, (cH1, cV1, cD1) = pywt.dwt2(data, 'haar') << cA1.shape = (125, 200)
cA2, (cH2, cV2, cD2) = pywt.dwt2(CA1, 'haar') << cA2.shape = (63, 100)
ca1 = pywt.dwt2((cA2, (cH2, cV2, cD2)), 'haar') >> ca1.shape = (126, 200)
data2= pywt.dwt2((ca1, (cH1, cV1, cD1)), 'haar') >> ca1.shape = (126, 200), ‘cA1.shape = cH1.shape = cv1.shape = cD1.shape = (125, 200)’
pywt.idwt2 does not like different dimensions of coefficients. I am not familiar with ‘db3’ but I tried your code with an image of shape (760, 680) and printed .shape of the coefficients:
Link to my notebook: https://github.com/Alaqian/PyWavelets-idtw2-error/blob/main/pywt-idwt2-error.ipynb
Code:
import pywt.data
import cv2
import numpy as np
import matplotlib.pyplot as plt
original = cv2.imread('gigachad.jpg',cv2.IMREAD_GRAYSCALE)
print(original.shape)
coeffs_d_1 = pywt.dwt2(original,'db3')
print(f"coeffs_d_1[1][1].shape = {coeffs_d_1[1][1].shape}")
coeffs_d_2 = pywt.dwt2(coeffs_d_1[0],'db3')
print(f"coeffs_d_2[1][1].shape = {coeffs_d_2[1][1].shape}")
coeffs_d_3 = pywt.dwt2(coeffs_d_2[0],'db3')
print(f"coeffs_d_3[1][1].shape = {coeffs_d_3[1][1].shape}")
coeffs_d_4 = pywt.dwt2(coeffs_d_3[0],'db3')
print(f"coeffs_d_4[1][1].shape = {coeffs_d_4[1][1].shape}")
coeffs_d_5 = pywt.dwt2(coeffs_d_4[0],'db3')
print(f"coeffs_d_5[1][1].shape = {coeffs_d_5[1][1].shape}")
coeffs_d_6 = pywt.dwt2(coeffs_d_5[0],'db3')
print(f"coeffs_d_6[1][1].shape = {coeffs_d_6[1][1].shape}")
coeffs_u_1 = pywt.idwt2((np.empty(coeffs_d_6[0].shape),coeffs_d_6[1]),'db3')
print(f"coeffs_u_1.shape = {coeffs_u_1.shape}")
coeffs_u_2 = pywt.idwt2((coeffs_u_1,coeffs_d_5[1]),'db3')
print(f"coeffs_u_2.shape = {coeffs_u_2.shape}")
coeffs_u_3 = pywt.idwt2((coeffs_u_2,coeffs_d_4[1]),'db3')
print(f"coeffs_u_3.shape = {coeffs_u_3.shape}")
Output:
original.shape = (760, 680)
coeffs_d_1[1][1].shape = (382, 342)
coeffs_d_2[1][1].shape = (193, 173)
coeffs_d_3[1][1].shape = (99, 89)
coeffs_d_4[1][1].shape = (52, 47)
coeffs_d_5[1][1].shape = (28, 26)
coeffs_d_6[1][1].shape = (16, 15)
coeffs_u_1.shape = (28, 26)
coeffs_u_2.shape = (52, 48)
ValueError Traceback (most recent call last)
Cell In[6], line 29
27 coeffs_u_2 = pywt.idwt2((coeffs_u_1,coeffs_d_5[1]),‘db3’)
28 print(f"coeffs_u_2.shape = {coeffs_u_2.shape}“)
—> 29 coeffs_u_3 = pywt.idwt2((coeffs_u_2,coeffs_d_4[1]),‘db3’)
30 print(f"coeffs_u_3.shape = {coeffs_u_3.shape}”)
31 coeffs_u_4 = pywt.idwt2((coeffs_u_3,coeffs_d_3[1]),‘db3’)
File ~\miniconda3\envs\ImV\lib\site-packages\pywt_multidim.py:118, in idwt2(coeffs, wavelet, mode, axes)
115 raise ValueError(“Expected 2 axes”)
117 coeffs = {‘aa’: LL, ‘da’: HL, ‘ad’: LH, ‘dd’: HH}
→ 118 return idwtn(coeffs, wavelet, mode, axes)
File ~\miniconda3\envs\ImV\lib\site-packages\pywt_multidim.py:280, in idwtn(coeffs, wavelet, mode, axes)
277 raise ValueError("coeffs must contain at least one non-null wavelet "
278 “band”)
279 if any(s != coeff_shape for s in coeff_shapes):
→ 280 raise ValueError(“coeffs must all be of equal size (or None)”)
282 if axes is None:
283 axes = range(ndim_transform)
ValueError: coeffs must all be of equal size (or None)
So coeffs_d_4[1][1].shape = (52, 47) and coeffs_u_2.shape = (52, 48) which idwt2 does not know how to deal with.
One solution is to check the dimensions of the coefficients and truncate the first row using coeffs_u_2[1:, :] but I am not a fan of this solution and the reconstruction results are poorer. The solution that worked for me using ‘haar’ was cropping the original image so that it is a multiple of 2 ** d, where d is the level of decomposition. In your case, d = 6, so your dimensions should be a multiple of 2**6 = 64 if you are using haar, but db3 works differently.
Hope this helps!
More information on this thread: https://github.com/PyWavelets/pywt/issues/54