taewon99/face-segmentation

1import gradio as gr2 3from matplotlib import gridspec4import matplotlib.pyplot as plt5import numpy as np6from PIL import Image7import tensorflow as tf8from transformers import SegformerFeatureExtractor, TFSegformerForSemanticSegmentation9 10feature_extractor = SegformerFeatureExtractor.from_pretrained(11    "jonathandinu/face-parsing"12)13model = TFSegformerForSemanticSegmentation.from_pretrained("jonathandinu/face-parsing")14 15 16def ade_palette():17    """ADE20K palette that maps each class to RGB values."""18    return [19        [125, 237, 123],20        [25, 97, 48],21        [59, 11, 81],22        [163, 123, 42],23        [239, 41, 136],24        [224, 4, 115],25        [114, 84, 169],26        [16, 137, 208],27        [153, 91, 30],28        [48, 90, 221],29        [91, 245, 206],30        [108, 87, 175],31        [232, 181, 231],32        [153, 70, 176],33        [32, 25, 179],34        [118, 177, 239],35        [246, 75, 15],36        [183, 17, 190],37        [79, 235, 51],38    ]39 40 41labels_list = []42 43with open(r"labels.txt", "r") as fp:44    for line in fp:45        labels_list.append(line[:-1])46 47colormap = np.asarray(ade_palette())48 49 50def label_to_color_image(label):51    if label.ndim != 2:52        raise ValueError("Expect 2-D input label")53 54    if np.max(label) >= len(colormap):55        raise ValueError("label value too large.")56    return colormap[label]57 58 59def draw_plot(pred_img, seg):60    fig = plt.figure(figsize=(20, 15))61 62    grid_spec = gridspec.GridSpec(1, 2, width_ratios=[6, 1])63 64    plt.subplot(grid_spec[0])65    plt.imshow(pred_img)66    plt.axis("off")67    LABEL_NAMES = np.asarray(labels_list)68    FULL_LABEL_MAP = np.arange(len(LABEL_NAMES)).reshape(len(LABEL_NAMES), 1)69    FULL_COLOR_MAP = label_to_color_image(FULL_LABEL_MAP)70 71    unique_labels = np.unique(seg.numpy().astype("uint8"))72    ax = plt.subplot(grid_spec[1])73    plt.imshow(FULL_COLOR_MAP[unique_labels].astype(np.uint8), interpolation="nearest")74    ax.yaxis.tick_right()75    plt.yticks(range(len(unique_labels)), LABEL_NAMES[unique_labels])76    plt.xticks([], [])77    ax.tick_params(width=0.0, labelsize=25)78    return fig79 80 81def sepia(input_img):82    input_img = Image.fromarray(input_img)83 84    inputs = feature_extractor(images=input_img, return_tensors="tf")85    outputs = model(**inputs)86    logits = outputs.logits87 88    logits = tf.transpose(logits, [0, 2, 3, 1])89    logits = tf.image.resize(90        logits, input_img.size[::-1]91    )  # We reverse the shape of `image` because `image.size` returns width and height.92    seg = tf.math.argmax(logits, axis=-1)[0]93 94    color_seg = np.zeros(95        (seg.shape[0], seg.shape[1], 3), dtype=np.uint896    )  # height, width, 397    for label, color in enumerate(colormap):98        color_seg[seg.numpy() == label, :] = color99 100    # Show image + mask101    pred_img = np.array(input_img) * 0.5 + color_seg * 0.5102    pred_img = pred_img.astype(np.uint8)103 104    fig = draw_plot(pred_img, seg)105    return fig106 107 108demo = gr.Interface(109    fn=sepia,110    inputs=gr.Image(shape=(400, 600)),111    outputs=["plot"],112    examples=[113        "elon.jpg",114        "biden.jpeg",115        "bezos.jpeg",116        "zuckerberg.jpeg",117    ],118    allow_flagging="never",119)120 121 122demo.launch()123