Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import numpy as np
|
3 |
+
import zipfile
|
4 |
+
import imageio
|
5 |
+
|
6 |
+
import tensorflow as tf
|
7 |
+
from tensorflow import keras
|
8 |
+
|
9 |
+
from utils import TubeMaskingGenerator
|
10 |
+
from utils import read_video, frame_sampling, denormalize, reconstrunction
|
11 |
+
from utils import IMAGENET_MEAN, IMAGENET_STD, num_frames, patch_size, input_size
|
12 |
+
from labels import K400_label_map, SSv2_label_map, UCF_label_map
|
13 |
+
|
14 |
+
|
15 |
+
LABEL_MAPS = {
|
16 |
+
'K400': K400_label_map,
|
17 |
+
'SSv2': SSv2_label_map,
|
18 |
+
'UCF' : UCF_label_map
|
19 |
+
}
|
20 |
+
|
21 |
+
ALL_MODELS = [
|
22 |
+
'TFVideoMAE_L_K400_16x224',
|
23 |
+
'TFVideoMAE_B_SSv2_16x224',
|
24 |
+
'TFVideoMAE_B_UCF_16x224',
|
25 |
+
]
|
26 |
+
|
27 |
+
sample_example = [
|
28 |
+
["examples/k400.mp4", ALL_MODELS[0], 0.9],
|
29 |
+
["examples/ssv2.mp4", ALL_MODELS[1], 0.8],
|
30 |
+
["examples/ucf.mp4", ALL_MODELS[2], 0.7],
|
31 |
+
]
|
32 |
+
|
33 |
+
def tube_mask_generator(mask_ratio):
|
34 |
+
window_size = (
|
35 |
+
num_frames // 2,
|
36 |
+
input_size // patch_size[0],
|
37 |
+
input_size // patch_size[1]
|
38 |
+
)
|
39 |
+
tube_mask = TubeMaskingGenerator(
|
40 |
+
input_size=window_size,
|
41 |
+
mask_ratio=mask_ratio
|
42 |
+
)
|
43 |
+
make_bool = tube_mask()
|
44 |
+
bool_masked_pos_tf = tf.constant(make_bool, dtype=tf.int32)
|
45 |
+
bool_masked_pos_tf = tf.expand_dims(bool_masked_pos_tf, axis=0)
|
46 |
+
bool_masked_pos_tf = tf.cast(bool_masked_pos_tf, tf.bool)
|
47 |
+
return bool_masked_pos_tf
|
48 |
+
|
49 |
+
|
50 |
+
def get_model(model_type):
|
51 |
+
ft_path = keras.utils.get_file(
|
52 |
+
origin=f'https://github.com/innat/VideoMAE/releases/download/v1.1/{model_type}_FT.zip',
|
53 |
+
)
|
54 |
+
pt_path = keras.utils.get_file(
|
55 |
+
origin=f'https://github.com/innat/VideoMAE/releases/download/v1.1/{model_type}_PT.zip',
|
56 |
+
)
|
57 |
+
|
58 |
+
with zipfile.ZipFile(ft_path, 'r') as zip_ref:
|
59 |
+
zip_ref.extractall('./')
|
60 |
+
|
61 |
+
with zipfile.ZipFile(pt_path, 'r') as zip_ref:
|
62 |
+
zip_ref.extractall('./')
|
63 |
+
|
64 |
+
ft_model = keras.models.load_model(model_type + '_FT')
|
65 |
+
pt_model = keras.models.load_model(model_type + '_PT')
|
66 |
+
|
67 |
+
if 'K400' in model_type:
|
68 |
+
data_type = 'K400'
|
69 |
+
elif 'SSv2' in model_type:
|
70 |
+
data_type = 'SSv2'
|
71 |
+
else:
|
72 |
+
data_type = 'UCF'
|
73 |
+
|
74 |
+
label_map = LABEL_MAPS.get(data_type)
|
75 |
+
label_map = {v: k for k, v in label_map.items()}
|
76 |
+
|
77 |
+
return ft_model, pt_model, label_map
|
78 |
+
|
79 |
+
|
80 |
+
def inference(video_file, model_type, mask_ratio):
|
81 |
+
# get sample data
|
82 |
+
container = read_video(video_file)
|
83 |
+
frames = frame_sampling(container, num_frames=num_frames)
|
84 |
+
|
85 |
+
# get models
|
86 |
+
bool_masked_pos_tf = tube_mask_generator(mask_ratio)
|
87 |
+
ft_model, pt_model, label_map = get_model(model_type)
|
88 |
+
ft_model.trainable = False
|
89 |
+
pt_model.trainable = False
|
90 |
+
|
91 |
+
# inference on fine-tune model
|
92 |
+
outputs_ft = ft_model(frames[None, ...], training=False)
|
93 |
+
probabilities = tf.nn.softmax(outputs_ft).numpy().squeeze(0)
|
94 |
+
confidences = {
|
95 |
+
label_map[i]: float(probabilities[i]) for i in np.argsort(probabilities)[::-1]
|
96 |
+
}
|
97 |
+
|
98 |
+
# inference on pre-trained model
|
99 |
+
outputs_pt = pt_model(frames[None, ...], bool_masked_pos_tf, training=False)
|
100 |
+
reconstruct_output, mask = reconstrunction(
|
101 |
+
frames[None, ...], bool_masked_pos_tf, outputs_pt
|
102 |
+
)
|
103 |
+
|
104 |
+
# post process
|
105 |
+
input_frame = denormalize(frames)
|
106 |
+
input_mask = denormalize(mask[0] * frames)
|
107 |
+
output_frame = denormalize(reconstruct_output)
|
108 |
+
|
109 |
+
frames = []
|
110 |
+
for frame_a, frame_b, frame_c in zip(input_frame, input_mask, output_frame):
|
111 |
+
combined_frame = np.hstack([frame_a, frame_b, frame_c])
|
112 |
+
frames.append(combined_frame)
|
113 |
+
|
114 |
+
combined_gif = 'combined.gif'
|
115 |
+
imageio.mimsave(combined_gif, frames, duration=300, loop=0)
|
116 |
+
return confidences, combined_gif
|
117 |
+
|
118 |
+
|
119 |
+
def main():
|
120 |
+
iface = gr.Interface(
|
121 |
+
fn=inference,
|
122 |
+
inputs=[
|
123 |
+
gr.Video(type="file", label="Input Video"),
|
124 |
+
gr.Dropdown(
|
125 |
+
choices=ALL_MODELS,
|
126 |
+
default="TFVideoMAE_L_K400_16x224",
|
127 |
+
label="Model"
|
128 |
+
),
|
129 |
+
gr.Slider(
|
130 |
+
0.5,
|
131 |
+
1.0,
|
132 |
+
step=0.1,
|
133 |
+
default=0.5,
|
134 |
+
label='Mask Ratio'
|
135 |
+
)
|
136 |
+
],
|
137 |
+
outputs=[
|
138 |
+
gr.Label(num_top_classes=3, label='scores'),
|
139 |
+
gr.Image(type="filepath", label='reconstructed')
|
140 |
+
],
|
141 |
+
examples=sample_example,
|
142 |
+
title="VideoMAE",
|
143 |
+
description="Keras reimplementation of <a href='https://github.com/innat/VideoMAE'>VideoMAE</a> is presented here."
|
144 |
+
)
|
145 |
+
|
146 |
+
iface.launch()
|
147 |
+
|
148 |
+
if __name__ == '__main__':
|
149 |
+
main()
|