Thback/CSI

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    "nvidia/segformer-b3-finetuned-cityscapes-1024-1024"12)13model = TFSegformerForSemanticSegmentation.from_pretrained(14    "nvidia/segformer-b3-finetuned-cityscapes-1024-1024"15)16 17def ade_palette():18    """ADE20K palette that maps each class to RGB values."""19    return [20        [234, 234, 234],21        [0, 0, 0],22        [255, 0, 0],23        [255, 255, 0],24        [255, 255, 255],25        [0, 255, 255],26        [0, 0, 255],27        [255, 0, 255],28        [243, 97, 220],29        [155, 0, 67],30        [50, 130, 255],31        [255, 130, 50],32        [53, 53, 53],33        [177, 177, 177],34        [95, 0, 255],35        [29, 255, 22],36        [255, 0, 95],37        [100, 100, 100],38        [92, 209, 229],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 49def label_to_color_image(label):50    if label.ndim != 2:51        raise ValueError("Expect 2-D input label")52 53    if np.max(label) >= len(colormap):54        raise ValueError("label value too large.")55    return colormap[label]56 57def draw_plot(pred_img, seg):58    fig = plt.figure(figsize=(20, 15))59 60    grid_spec = gridspec.GridSpec(1, 2, width_ratios=[6, 1])61 62    plt.subplot(grid_spec[0])63    plt.imshow(pred_img)64    plt.axis('off')65    LABEL_NAMES = np.asarray(labels_list)66    FULL_LABEL_MAP = np.arange(len(LABEL_NAMES)).reshape(len(LABEL_NAMES), 1)67    FULL_COLOR_MAP = label_to_color_image(FULL_LABEL_MAP)68 69    unique_labels = np.unique(seg.numpy().astype("uint8"))70    ax = plt.subplot(grid_spec[1])71    plt.imshow(FULL_COLOR_MAP[unique_labels].astype(np.uint8), interpolation="nearest")72    ax.yaxis.tick_right()73    plt.yticks(range(len(unique_labels)), LABEL_NAMES[unique_labels])74    plt.xticks([], [])75    ax.tick_params(width=0.0, labelsize=25)76    return fig77 78def sepia(input_img):79    input_img = Image.fromarray(input_img)80 81    inputs = feature_extractor(images=input_img, return_tensors="tf")82    outputs = model(**inputs)83    logits = outputs.logits84 85    logits = tf.transpose(logits, [0, 2, 3, 1])86    logits = tf.image.resize(87        logits, input_img.size[::-1]88    )  # We reverse the shape of `image` because `image.size` returns width and height.89    seg = tf.math.argmax(logits, axis=-1)[0]90 91    color_seg = np.zeros(92        (seg.shape[0], seg.shape[1], 3), dtype=np.uint893    )  # height, width, 394    for label, color in enumerate(colormap):95        color_seg[seg.numpy() == label, :] = color96 97    # Show image + mask98    pred_img = np.array(input_img) * 0.5 + color_seg * 0.599    pred_img = pred_img.astype(np.uint8)100 101    fig = draw_plot(pred_img, seg)102    return fig103 104demo = gr.Interface(fn=sepia,105                    inputs=gr.Image(shape=(400, 600)),106                    outputs=['plot'],107                    examples=["cityscapes-1.jpg", "cityscapes-2.jpg", "cityscapes-3.jpg", "cityscapes-4.jpg", "cityscapes-5.jpg", "cityscapes-6.jpg"],108                    allow_flagging='never')109 110 111demo.launch()112