np.histogram2d exception too many values to unpack (expected 2)

The function is to calculate the distribution of color values in the ab color space from a set of images and the parameter images is array of rgb images.

When executing

np.histogram2d(pixel_ab[:, 0], pixel_ab[:, 1], bins=(ab_values[:-1], ab_values[:-1]), density=True)

> exception (<class 'ValueError'>, ValueError('too many values to unpack (expected 2)'))

is thrown in executing histogramdd function. It the shape it is getting is 3 but it is expecting 2.

How this can be altered to remove exception or alternate implementation?

import numpy as np
from skimage.color import rgb2lab
from scipy.ndimage import gaussian_filter

def compute_color_distribution(images, ab_bins=32, sigma=5, lam=0.5):
    ab_values = np.linspace(-1, 1, ab_bins)
    ab_space = np.stack(np.meshgrid(ab_values, ab_values), axis=-1).reshape(-1, 2)

    pixel_ab = []
    for img in images:
        lab_img = rgb2lab(img / 255.0)
        pixel_ab.append(lab_img[:, :, 1:] / 128)
    pixel_ab = np.vstack(pixel_ab)

    hist, x_edges, y_edges = np.histogram2d(pixel_ab[:, 0], pixel_ab[:, 1], bins=(ab_values[:-1], ab_values[:-1]), density=True)
    smoothed_hist = gaussian_filter(hist, sigma)

    p = smoothed_hist / np.sum(smoothed_hist)
    p_tilde = (1 - lam) * p + lam / (ab_bins * ab_bins)

    weights = (1 / p_tilde).flatten()
    weights /= np.sum(p * weights)

    return ab_space, weights

The following answer is going to address the exception, and may not fix the entire code...

Following the documentation of numpy.histogram2d.

> numpy.histogram2d(x, y, bins=10, range=None, density=None, weights=None)
>Compute the bi-dimensional histogram of two data samples.
>Parameters
> - x: array_like, shape (N,)
> An array containing the x coordinates of the points to be histogrammed.
> - y array_like, shape (N,)
>An array containing the y coordinates of the points to be histogrammed.

According to the documentation the shape of x argument is (N,), and the shape if y argument is also (N,).
That means that x and y arguments are 1 dimensional arrays with N elements.

Following your code (using the debugger), we can see that:
pixel_ab[:, 0].shape = (N, 2) and pixel_ab[:, 1].shape = (N, 2).

When executing np.histogram2d(pixel_ab[:, 0], pixel_ab[:, 1]... the x and y arguments are two dimensional arrays instead of 1 dimensional (as expected by np.histogram2d).

The result is the exception ValueError('too many values to unpack (expected 2).

Fixing the code is outside the scope of the answer.

The following code sample is an example that doesn't rise an exception (and with some luck gets closer to your desired functionality):

import numpy as np
from skimage.color import rgb2lab
from scipy.ndimage import gaussian_filter

def compute_color_distribution(images, ab_bins=32, sigma=5, lam=0.5):
    ab_values = np.linspace(-1, 1, ab_bins)
    ab_space = np.stack(np.meshgrid(ab_values, ab_values), axis=-1).reshape(-1, 2)

    #pixel_ab = []
    #for img in images:
    #    lab_img = rgb2lab(img / 255.0)
    #    pixel_ab.append(lab_img[:, :, 1:] / 128)
    #pixel_ab = np.vstack(pixel_ab)

    pixel_a = []
    pixel_b = []
    for img in images:
        lab_img = rgb2lab(img / 255.0)
        a_img = lab_img[:, :, 1] / 128
        b_img = lab_img[:, :, 2] / 128
        a_img_as_row = np.ravel(a_img)
        b_img_as_row = np.ravel(b_img)
        pixel_a.append(a_img_as_row)
        pixel_b.append(b_img_as_row)
    pixel_a = np.hstack(pixel_a)
    pixel_b = np.hstack(pixel_b)

    #hist, x_edges, y_edges = np.histogram2d(pixel_ab[:, 0], pixel_ab[:, 1], bins=(ab_values[:-1], ab_values[:-1]), density=True)
    hist, x_edges, y_edges = np.histogram2d(pixel_a, pixel_b, bins=ab_bins, density=True)  # hist.shape = (32, 32)
    smoothed_hist = gaussian_filter(hist, sigma)

    p = smoothed_hist / np.sum(smoothed_hist)
    p_tilde = (1 - lam) * p + lam / (ab_bins * ab_bins)

    #weights = (1 / p_tilde).flatten()
    weights = 1 / p_tilde  # If we flatten() then we can multiply p*weights in the next statement.
    weights /= np.sum(p * weights)

    return ab_space, weights


images = [np.random.randint(0, 255, (100, 100, 3), np.uint8), np.random.randint(0, 255, (100, 100, 3), np.uint8), np.random.randint(0, 255, (100, 100, 3), np.uint8)]

ab_space, weights = compute_color_distribution(images)