.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "auto_examples/cluster/plot_face_compress.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note :ref:`Go to the end ` to download the full example code or to run this example in your browser via JupyterLite or Binder .. rst-class:: sphx-glr-example-title .. _sphx_glr_auto_examples_cluster_plot_face_compress.py: =========================== Vector Quantization Example =========================== This example shows how one can use :class:`~sklearn.preprocessing.KBinsDiscretizer` to perform vector quantization on a set of toy image, the raccoon face. .. GENERATED FROM PYTHON SOURCE LINES 9-14 .. code-block:: Python # Authors: Gael Varoquaux # Jaques Grobler # License: BSD 3 clause .. GENERATED FROM PYTHON SOURCE LINES 15-25 Original image -------------- We start by loading the raccoon face image from SciPy. We will additionally check a couple of information regarding the image, such as the shape and data type used to store the image. Note that depending of the SciPy version, we have to adapt the import since the function returning the image is not located in the same module. Also, SciPy >= 1.10 requires the package `pooch` to be installed. .. GENERATED FROM PYTHON SOURCE LINES 25-36 .. code-block:: Python try: # Scipy >= 1.10 from scipy.datasets import face except ImportError: from scipy.misc import face raccoon_face = face(gray=True) print(f"The dimension of the image is {raccoon_face.shape}") print(f"The data used to encode the image is of type {raccoon_face.dtype}") print(f"The number of bytes taken in RAM is {raccoon_face.nbytes}") .. rst-class:: sphx-glr-script-out .. code-block:: none The dimension of the image is (768, 1024) The data used to encode the image is of type uint8 The number of bytes taken in RAM is 786432 .. GENERATED FROM PYTHON SOURCE LINES 37-44 Thus the image is a 2D array of 768 pixels in height and 1024 pixels in width. Each value is a 8-bit unsigned integer, which means that the image is encoded using 8 bits per pixel. The total memory usage of the image is 786 kilobytes (1 byte equals 8 bits). Using 8-bit unsigned integer means that the image is encoded using 256 different shades of gray, at most. We can check the distribution of these values. .. GENERATED FROM PYTHON SOURCE LINES 44-57 .. code-block:: Python import matplotlib.pyplot as plt fig, ax = plt.subplots(ncols=2, figsize=(12, 4)) ax[0].imshow(raccoon_face, cmap=plt.cm.gray) ax[0].axis("off") ax[0].set_title("Rendering of the image") ax[1].hist(raccoon_face.ravel(), bins=256) ax[1].set_xlabel("Pixel value") ax[1].set_ylabel("Count of pixels") ax[1].set_title("Distribution of the pixel values") _ = fig.suptitle("Original image of a raccoon face") .. image-sg:: /auto_examples/cluster/images/sphx_glr_plot_face_compress_001.png :alt: Original image of a raccoon face, Rendering of the image, Distribution of the pixel values :srcset: /auto_examples/cluster/images/sphx_glr_plot_face_compress_001.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 58-75 Compression via vector quantization ----------------------------------- The idea behind compression via vector quantization is to reduce the number of gray levels to represent an image. For instance, we can use 8 values instead of 256 values. Therefore, it means that we could efficiently use 3 bits instead of 8 bits to encode a single pixel and therefore reduce the memory usage by a factor of approximately 2.5. We will later discuss about this memory usage. Encoding strategy """"""""""""""""" The compression can be done using a :class:`~sklearn.preprocessing.KBinsDiscretizer`. We need to choose a strategy to define the 8 gray values to sub-sample. The simplest strategy is to define them equally spaced, which correspond to setting `strategy="uniform"`. From the previous histogram, we know that this strategy is certainly not optimal. .. GENERATED FROM PYTHON SOURCE LINES 75-100 .. code-block:: Python from sklearn.preprocessing import KBinsDiscretizer n_bins = 8 encoder = KBinsDiscretizer( n_bins=n_bins, encode="ordinal", strategy="uniform", random_state=0, subsample=200_000, ) compressed_raccoon_uniform = encoder.fit_transform(raccoon_face.reshape(-1, 1)).reshape( raccoon_face.shape ) fig, ax = plt.subplots(ncols=2, figsize=(12, 4)) ax[0].imshow(compressed_raccoon_uniform, cmap=plt.cm.gray) ax[0].axis("off") ax[0].set_title("Rendering of the image") ax[1].hist(compressed_raccoon_uniform.ravel(), bins=256) ax[1].set_xlabel("Pixel value") ax[1].set_ylabel("Count of pixels") ax[1].set_title("Sub-sampled distribution of the pixel values") _ = fig.suptitle("Raccoon face compressed using 3 bits and a uniform strategy") .. image-sg:: /auto_examples/cluster/images/sphx_glr_plot_face_compress_002.png :alt: Raccoon face compressed using 3 bits and a uniform strategy, Rendering of the image, Sub-sampled distribution of the pixel values :srcset: /auto_examples/cluster/images/sphx_glr_plot_face_compress_002.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 101-108 Qualitatively, we can spot some small regions where we see the effect of the compression (e.g. leaves on the bottom right corner). But after all, the resulting image is still looking good. We observe that the distribution of pixels values have been mapped to 8 different values. We can check the correspondence between such values and the original pixel values. .. GENERATED FROM PYTHON SOURCE LINES 108-113 .. code-block:: Python bin_edges = encoder.bin_edges_[0] bin_center = bin_edges[:-1] + (bin_edges[1:] - bin_edges[:-1]) / 2 bin_center .. rst-class:: sphx-glr-script-out .. code-block:: none array([ 15.625, 46.875, 78.125, 109.375, 140.625, 171.875, 203.125, 234.375]) .. GENERATED FROM PYTHON SOURCE LINES 114-121 .. code-block:: Python _, ax = plt.subplots() ax.hist(raccoon_face.ravel(), bins=256) color = "tab:orange" for center in bin_center: ax.axvline(center, color=color) ax.text(center - 10, ax.get_ybound()[1] + 100, f"{center:.1f}", color=color) .. image-sg:: /auto_examples/cluster/images/sphx_glr_plot_face_compress_003.png :alt: plot face compress :srcset: /auto_examples/cluster/images/sphx_glr_plot_face_compress_003.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 122-127 As previously stated, the uniform sampling strategy is not optimal. Notice for instance that the pixels mapped to the value 7 will encode a rather small amount of information, whereas the mapped value 3 will represent a large amount of counts. We can instead use a clustering strategy such as k-means to find a more optimal mapping. .. GENERATED FROM PYTHON SOURCE LINES 127-149 .. code-block:: Python encoder = KBinsDiscretizer( n_bins=n_bins, encode="ordinal", strategy="kmeans", random_state=0, subsample=200_000, ) compressed_raccoon_kmeans = encoder.fit_transform(raccoon_face.reshape(-1, 1)).reshape( raccoon_face.shape ) fig, ax = plt.subplots(ncols=2, figsize=(12, 4)) ax[0].imshow(compressed_raccoon_kmeans, cmap=plt.cm.gray) ax[0].axis("off") ax[0].set_title("Rendering of the image") ax[1].hist(compressed_raccoon_kmeans.ravel(), bins=256) ax[1].set_xlabel("Pixel value") ax[1].set_ylabel("Number of pixels") ax[1].set_title("Distribution of the pixel values") _ = fig.suptitle("Raccoon face compressed using 3 bits and a K-means strategy") .. image-sg:: /auto_examples/cluster/images/sphx_glr_plot_face_compress_004.png :alt: Raccoon face compressed using 3 bits and a K-means strategy, Rendering of the image, Distribution of the pixel values :srcset: /auto_examples/cluster/images/sphx_glr_plot_face_compress_004.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 150-154 .. code-block:: Python bin_edges = encoder.bin_edges_[0] bin_center = bin_edges[:-1] + (bin_edges[1:] - bin_edges[:-1]) / 2 bin_center .. rst-class:: sphx-glr-script-out .. code-block:: none array([ 18.90885631, 53.34346583, 82.64447187, 109.28225276, 134.70763101, 159.78681467, 185.17226834, 224.02069427]) .. GENERATED FROM PYTHON SOURCE LINES 155-162 .. code-block:: Python _, ax = plt.subplots() ax.hist(raccoon_face.ravel(), bins=256) color = "tab:orange" for center in bin_center: ax.axvline(center, color=color) ax.text(center - 10, ax.get_ybound()[1] + 100, f"{center:.1f}", color=color) .. image-sg:: /auto_examples/cluster/images/sphx_glr_plot_face_compress_005.png :alt: plot face compress :srcset: /auto_examples/cluster/images/sphx_glr_plot_face_compress_005.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 163-171 The counts in the bins are now more balanced and their centers are no longer equally spaced. Note that we could enforce the same number of pixels per bin by using the `strategy="quantile"` instead of `strategy="kmeans"`. Memory footprint """""""""""""""" We previously stated that we should save 8 times less memory. Let's verify it. .. GENERATED FROM PYTHON SOURCE LINES 171-175 .. code-block:: Python print(f"The number of bytes taken in RAM is {compressed_raccoon_kmeans.nbytes}") print(f"Compression ratio: {compressed_raccoon_kmeans.nbytes / raccoon_face.nbytes}") .. rst-class:: sphx-glr-script-out .. code-block:: none The number of bytes taken in RAM is 6291456 Compression ratio: 8.0 .. GENERATED FROM PYTHON SOURCE LINES 176-180 It is quite surprising to see that our compressed image is taking x8 more memory than the original image. This is indeed the opposite of what we expected. The reason is mainly due to the type of data used to encode the image. .. GENERATED FROM PYTHON SOURCE LINES 180-183 .. code-block:: Python print(f"Type of the compressed image: {compressed_raccoon_kmeans.dtype}") .. rst-class:: sphx-glr-script-out .. code-block:: none Type of the compressed image: float64 .. GENERATED FROM PYTHON SOURCE LINES 184-194 Indeed, the output of the :class:`~sklearn.preprocessing.KBinsDiscretizer` is an array of 64-bit float. It means that it takes x8 more memory. However, we use this 64-bit float representation to encode 8 values. Indeed, we will save memory only if we cast the compressed image into an array of 3-bits integers. We could use the method `numpy.ndarray.astype`. However, a 3-bits integer representation does not exist and to encode the 8 values, we would need to use the 8-bit unsigned integer representation as well. In practice, observing a memory gain would require the original image to be in a 64-bit float representation. .. rst-class:: sphx-glr-timing **Total running time of the script:** (0 minutes 1.980 seconds) .. _sphx_glr_download_auto_examples_cluster_plot_face_compress.py: .. only:: html .. container:: sphx-glr-footer sphx-glr-footer-example .. container:: binder-badge .. image:: images/binder_badge_logo.svg :target: https://mybinder.org/v2/gh/scikit-learn/scikit-learn/main?urlpath=lab/tree/notebooks/auto_examples/cluster/plot_face_compress.ipynb :alt: Launch binder :width: 150 px .. container:: lite-badge .. image:: images/jupyterlite_badge_logo.svg :target: ../../lite/lab/?path=auto_examples/cluster/plot_face_compress.ipynb :alt: Launch JupyterLite :width: 150 px .. container:: sphx-glr-download sphx-glr-download-jupyter :download:`Download Jupyter notebook: plot_face_compress.ipynb ` .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: plot_face_compress.py ` .. include:: plot_face_compress.recommendations .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_