han-byeol commited on
Commit
352a01b
1 Parent(s): c1715d9

Upload 8 files

Browse files
app.txt ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+
3
+ from matplotlib import gridspec
4
+ import matplotlib.pyplot as plt
5
+ import numpy as np
6
+ from PIL import Image
7
+ import tensorflow as tf
8
+ from transformers import SegformerFeatureExtractor, TFSegformerForSemanticSegmentation
9
+
10
+ feature_extractor = SegformerFeatureExtractor.from_pretrained(
11
+ "nvidia/segformer-b5-finetuned-cityscapes-1024-1024"
12
+ )
13
+ model = TFSegformerForSemanticSegmentation.from_pretrained(
14
+ "nvidia/segformer-b5-finetuned-cityscapes-1024-1024"
15
+ )
16
+
17
+ def ade_palette():
18
+ """ADE20K palette that maps each class to RGB values."""
19
+ return [
20
+ [204, 87, 92],
21
+ [112, 185, 212],
22
+ [45, 189, 106],
23
+ [234, 123, 67],
24
+ [78, 56, 123],
25
+ [210, 32, 89],
26
+ [90, 180, 56],
27
+ [155, 102, 200],
28
+ [33, 147, 176],
29
+ [255, 183, 76],
30
+ [67, 123, 89],
31
+ [190, 60, 45],
32
+ [134, 112, 200],
33
+ [56, 45, 189],
34
+ [200, 56, 123],
35
+ [87, 92, 204],
36
+ [120, 56, 123],
37
+ [45, 78, 123]
38
+ ]
39
+
40
+ labels_list = []
41
+
42
+ with open(r'labels.txt', 'r') as fp:
43
+ for line in fp:
44
+ labels_list.append(line[:-1])
45
+
46
+ colormap = np.asarray(ade_palette())
47
+
48
+ def label_to_color_image(label):
49
+ if label.ndim != 2:
50
+ raise ValueError("Expect 2-D input label")
51
+
52
+ if np.max(label) >= len(colormap):
53
+ raise ValueError("label value too large.")
54
+ return colormap[label]
55
+
56
+ def draw_plot(pred_img, seg):
57
+ fig = plt.figure(figsize=(20, 15))
58
+
59
+ grid_spec = gridspec.GridSpec(1, 2, width_ratios=[6, 1])
60
+
61
+ plt.subplot(grid_spec[0])
62
+ plt.imshow(pred_img)
63
+ plt.axis('off')
64
+ LABEL_NAMES = np.asarray(labels_list)
65
+ FULL_LABEL_MAP = np.arange(len(LABEL_NAMES)).reshape(len(LABEL_NAMES), 1)
66
+ FULL_COLOR_MAP = label_to_color_image(FULL_LABEL_MAP)
67
+
68
+ unique_labels = np.unique(seg.numpy().astype("uint8"))
69
+ ax = plt.subplot(grid_spec[1])
70
+ plt.imshow(FULL_COLOR_MAP[unique_labels].astype(np.uint8), interpolation="nearest")
71
+ ax.yaxis.tick_right()
72
+ plt.yticks(range(len(unique_labels)), LABEL_NAMES[unique_labels])
73
+ plt.xticks([], [])
74
+ ax.tick_params(width=0.0, labelsize=25)
75
+ return fig
76
+
77
+ def sepia(input_img):
78
+ input_img = Image.fromarray(input_img)
79
+
80
+ inputs = feature_extractor(images=input_img, return_tensors="tf")
81
+ outputs = model(**inputs)
82
+ logits = outputs.logits
83
+
84
+ logits = tf.transpose(logits, [0, 2, 3, 1])
85
+ logits = tf.image.resize(
86
+ logits, input_img.size[::-1]
87
+ ) # We reverse the shape of `image` because `image.size` returns width and height.
88
+ seg = tf.math.argmax(logits, axis=-1)[0]
89
+
90
+ color_seg = np.zeros(
91
+ (seg.shape[0], seg.shape[1], 3), dtype=np.uint8
92
+ ) # height, width, 3
93
+ for label, color in enumerate(colormap):
94
+ color_seg[seg.numpy() == label, :] = color
95
+
96
+ # Show image + mask
97
+ pred_img = np.array(input_img) * 0.5 + color_seg * 0.5
98
+ pred_img = pred_img.astype(np.uint8)
99
+
100
+ fig = draw_plot(pred_img, seg)
101
+ return fig
102
+
103
+ demo = gr.Interface(fn=sepia,
104
+ inputs=gr.Image(shape=(400, 600)),
105
+ outputs=['plot'],
106
+ examples=["cityscapes-1.jpg", "cityscapes-2.jpg", "cityscapes-3.jpg"],
107
+ allow_flagging='never')
108
+
109
+
110
+ demo.launch()
cityscapes-1.jpg ADDED
cityscapes-2.jpg ADDED
cityscapes-3.jpg ADDED
cityscapes-4.jpg ADDED
cityscapes-5.jpg ADDED
labels.txt ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ road
2
+ sidewalk
3
+ building
4
+ wall
5
+ fence
6
+ pole
7
+ traffic light
8
+ traffic sign
9
+ vegetation
10
+ terrain
11
+ sky
12
+ person
13
+ rider
14
+ car
15
+ truck
16
+ bus
17
+ train
18
+ motorcycle
requirements.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ torch
2
+ transformers
3
+ tensorflow
4
+ numpy
5
+ Image
6
+ matplotlib