test.py

n_masks = 100
n_channels = 100
# do not sample more masks than this
n_max_samples = 50

target_corr = 0.1
target_corr_eps = 1.0e-2

rate = target_corr

min_corr = target_corr - target_corr_eps
max_corr = target_corr + target_corr_eps

n_indices = int(n_channels * rate)

masks = np.zeros((n_masks, n_channels))
indices = np.random.choice(n_channels, n_indices, replace=False)
masks[0,indices] = 1.0
n_samples = 1

def _sample_mask():
    indices = np.random.choice(n_channels, n_indices, replace=False)
    mask = np.zeros((n_channels))
    mask[indices] = 1.0
    return mask
    
for i in range(1, n_masks):
    # generating the next mask
    prev_masks = masks[:n_samples]
    for j in range(n_max_samples):
        mask = _sample_mask()
        mean_corr = np.mean(np.dot(prev_masks, mask) / n_indices)
        if min_corr < mean_corr and mean_corr < max_corr:
            break

    # checking if it satisfies our constraints
    masks[n_samples,:] = mask
    n_samples += 1
    
corr = np.dot(masks, masks.T) / n_indices
np.mean(corr)