sparse_dict_demo.ipynb

Kernel: Python 3 (ipykernel)

"""
Author : Srikar reddy Jilugu(@always-newbie161)
This code is referenced from
https://github.com/probml/pmtk3/blob/master/demos/sparseDictDemo.m
spams package is from
https://gitlab.inria.fr/thoth/spams-devel
(Can be instaled from PyPi through pip install -qq ...)
"""


import numpy as np

try:
    from spams import trainDL, nmf, displayPatches, im2col_sliding
except ModuleNotFoundError:
    %pip install -qq spams
    from spams import trainDL, nmf, displayPatches, im2col_sliding
from sklearn.decomposition import TruncatedSVD

try:
    import probml_utils as pml
except ModuleNotFoundError:
    %pip install -qq git+https://github.com/probml/probml-utils.git
    import probml_utils as pml

import matplotlib.pyplot as plt
from PIL import Image
import requests
import io

# dimension of latent space or no .of basis vectors.
D = 64
r = requests.get("https://github.com/probml/probml-data/blob/main/data/lena.png?raw=true", stream=True)
img = Image.open(io.BytesIO(r.content))

img = np.array(img) / 255
# spams accepts only column stored arrays i.e fortran-based
X = np.asfortranarray(img)
# preprocessing the image with a sliding window to get the actual observational data.
X = im2col_sliding(X, 12, 12)
print("Shape of observational data: ", X.shape)
print("Shape of the matrix of basis vectors will be: ", (X.shape[0], D))

methods = ["pca", "spca", "dl", "nmf"]

for m in methods:

    # Some preprocess should be done before training.
    # nmf - unitnorm
    # spca, dl - centered and unit norm
    # pca - centered data will be SVD decomposed.

    X_m = X

    # NMf accepts onlt non-negative data, So it should not be centered.
    if m != "nmf":
        X_m = X_m - X_m.mean(axis=0)

    # normalizing..
    X_m = np.asfortranarray(X_m / np.sqrt((X_m * X_m).sum(axis=0)), dtype=np.float64)

    if m == "pca":
        svd = TruncatedSVD(n_components=D, n_iter=7)
        W = svd.fit_transform(X_m)
    else:
        param = {"K": D, "numThreads": -1, "batchsize": 256, "iter": 1000}

        if m == "nmf":
            W = nmf(X_m, **param)
        else:
            param["verbose"] = False
            if m == "spca":
                param["lambda1"] = 0.1
                param["gamma1"] = 0.1
                param["modeD"] = 1
                W = trainDL(X_m, **param)
            elif m == "dl":
                param["lambda1"] = 0.1
                param["mode"] = 2
                W = trainDL(X_m, **param)

    # resulting basisvectors will be visualized as image patches.
    img = displayPatches(W)
    img = np.uint8(img[:, :, 0] * 255.0)
    img = Image.fromarray(img, mode="L")
    plt.title("sparse_dict_demo_{}".format(m))
    plt.imshow(img, cmap="gray")
    pml.savefig("sparse_dict_demo_{}.pdf".format(m))
    plt.show()