latent features in u-net

unet_kmeans.py

from sklearn.cluster import KMeans
import numpy as np
from keras.datasets import mnist
import os, tarfile, datetime

def kmeans(path, state):
    if path == "original":
        (latent, gt), (_, _) = mnist.load_data()
        latent = (latent / 255.0).reshape(-1, 784)
        gt = np.ravel(gt)
    else:
        data = np.load(path)
        latent = data["latent"]
        gt = np.ravel(data["ground_truth"])

    km = KMeans(n_clusters=10, random_state=state)
    km_label = km.fit_predict(latent)

    # row:kmeans, col:ground_truth
    matrix = np.zeros((10, 10), np.int32)
    for i in range(60000):
        row = km_label[i]
        col = gt[i]
        matrix[row, col] += 1

    # purity
    row_purity = np.max(matrix, axis=-1) / np.sum(matrix, axis=-1)
    weights = np.sum(matrix, axis=-1) / np.sum(matrix)
    total_purity = np.sum(row_purity * weights)

    return total_purity

def kmeans_multiple(path):
    results = np.zeros(10)
    print("Starts...", path)
    for i in range(10):
        print("... i = ", i, datetime.datetime.now())
        results[i] = kmeans(path, i)
    print(path)
    print(results)
    print(np.mean(results))

    if not os.path.exists("results"):
        os.mkdir("results")
    np.savez("results/"+path.replace(".npz", ""), results=results)

def kmeans_all():
    paths = ["original",
            "latent_skip_False_bottleneck_False.npz",
             "latent_skip_False_bottleneck_True.npz",
             "latent_skip_True_bottleneck_False.npz",
             "latent_skip_True_bottleneck_True.npz"]
    for p in paths:
        kmeans_multiple(p)

    with tarfile.open("results.tar.gz", mode="w:gz") as tar:
        tar.add("results")


kmeans_all()