more description + bins slider

app.py

@@ -44,21 +44,21 @@ def build_compressed_plot(compressed_image, img_array, sampling: str) -> plt.Fig
     return compressed_text, fig
 
 
-def infer(img_array: np.ndarray, sampling: str):
+def infer(img_array: np.ndarray, sampling: str, number_of_bins: int):
     # greyscale_image = input_image.convert("L")
     # img_array = np.array(greyscale_image)
 
     #raccoon_face = face(gray=True)
     init_text, init_fig = build_init_plot(img_array)
     
-    n_bins = 8
+    n_bins = number_of_bins
     encoder = KBinsDiscretizer(
         n_bins=n_bins, encode="ordinal", strategy=sampling, random_state=0
     )
     compressed_image = encoder.fit_transform(img_array.reshape(-1, 1)).reshape(
         img_array.shape
     )
-
+    compressed_image = compressed_image.astype(np.uint8)
     compressed_text, compressed_fig = build_compressed_plot(compressed_image,
                                                             img_array,
                                                             sampling)
@@ -83,13 +83,19 @@ Demo by <a href='https://huggingface.co/johko' target='_blank'>Johannes (johko)
 gr.Interface(
     title="Vector Quantization with scikit-learn",
     description="""<p style="text-align: center;">This is an interactive demo for the <a href="https://scikit-learn.org/stable/auto_examples/cluster/plot_face_compress.html">Vector Quantization Tutorial</a> from scikit-learn.
-    </br>You can simply upload an image and choose from two sampling methods - *uniform* and *kmeans*.
-    </br>Then click 'submit' and the image will the be quantized. You will get information about the histogram, pixles distribution and other image statistics.
+    </br><b>Vector Quantization</b> is a compression technique to reduce the number of color values that are used in an image and with this save memory while trying to keep a good quality.
+    In this demo this can be done naively via <i>uniform</i> sampling, which just uses <i>N</i> color values (specified via slider) uniformly sampled from the whole spectrum or via <i>k-means</i> which pays closer attention 
+    to the actual pixel distribution and potentially leads to a better quality of the compressed image.
+    In this demo we actually won't see a compression effect, because we cannot go smaller than <i>uint8</i> in datatype size here.
+    </br>
+    </br><b>Usage</b>: To run the demo you can simply upload an image and choose from two sampling methods - <i>uniform</i> and <i>kmeans</i>. Choose the number of bins and then click 'submit'. 
+    You will get information about the histogram, pixels distribution and other image statistics for your orginial image as grayscale and the quantized version of it.
     </p>""",
     article=article,
     fn=infer, 
     inputs=[gr.Image(image_mode="L", label="Input Image"),
-            gr.Dropdown(choices=["uniform", "kmeans"], label="Sampling Method")], 
+            gr.Dropdown(choices=["uniform", "kmeans"], label="Sampling Method"),
+            gr.Slider(minimum=2, maximum=50, value=8, step=1, label="Number of Bins")], 
     outputs=[gr.Text(label="Original Image Stats"),
              gr.Plot(label="Original Image Histogram"),
              gr.Text(label="Compressed Image Stats"),