File size: 3,222 Bytes
24cfc1b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fbc7dad
 
24cfc1b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c01c45a
 
 
24cfc1b
 
 
 
 
 
 
 
f811062
 
24cfc1b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4771f9a
24cfc1b
 
 
 
 
 
 
 
fbc7dad
 
24cfc1b
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
import gradio as gr
import numpy as np
import imageio

import tensorflow as tf
from tensorflow import keras

from utils import TubeMaskingGenerator
from utils import read_video, frame_sampling, denormalize, reconstrunction
from utils import IMAGENET_MEAN, IMAGENET_STD, num_frames, patch_size, input_size
from labels import K400_label_map, SSv2_label_map, UCF_label_map


MODELS = {
    'K400': [
        'innat/videomae/TFVideoMAE_S_K400_16x224_FT',
        'innat/videomae/TFVideoMAE_S_K400_16x224_PT'
        ],
    'SSv2': [],
    'UCF' : []
}


def tube_mask_generator():
    window_size = (
        num_frames // 2, 
        input_size // patch_size[0], 
        input_size // patch_size[1]
    )
    tube_mask = TubeMaskingGenerator(
        input_size=window_size, 
        mask_ratio=0.70
    )
    make_bool = tube_mask()
    bool_masked_pos_tf = tf.constant(make_bool, dtype=tf.int32)
    bool_masked_pos_tf = tf.expand_dims(bool_masked_pos_tf, axis=0)
    bool_masked_pos_tf = tf.cast(bool_masked_pos_tf, tf.bool)
    return bool_masked_pos_tf


def video_to_gif(video_array, gif_filename):
    imageio.mimsave(
        gif_filename, video_array, duration=100
    )


def get_model(data_type):
    ft_model = keras.models.load_model(MODELS[data_type][0])
    pt_model = keras.models.load_model(MODELS[data_type][1])
    label_map = {v: k for k, v in K400_label_map.items()}
    return ft_model, pt_model, label_map


def inference(video_file, dataset_type):
    container = read_video(video_file)
    frames = frame_sampling(container, num_frames=num_frames)
    bool_masked_pos_tf = tube_mask_generator()
    ft_model, pt_model, label_map = get_model(dataset_type)
    ft_model.trainable = False
    pt_model.trainable = False

    # inference on fine-tune model
    outputs_ft = ft_model(frames[None, ...], training=False)
    probabilities = tf.nn.softmax(outputs_ft).numpy().squeeze(0)
    confidences = {
        label_map[i]: float(probabilities[i]) for i in np.argsort(probabilities)[::-1]
    }

    # inference on pre-trained model
    outputs_pt = pt_model(frames[None, ...], bool_masked_pos_tf, training=False)
    reconstruct_output, mask = reconstrunction(
        frames[None, ...], bool_masked_pos_tf, outputs_pt
    )

    input_frame = denormalize(frames)
    input_mask = denormalize(mask[0] * frames)
    output_frame = denormalize(reconstruct_output)

    frames = []
    for frame_a, frame_b, frame_c in zip(input_frame, input_mask, output_frame):
        combined_frame = np.hstack([frame_a, frame_b, frame_c])
        frames.append(combined_frame)

    combined_gif = 'combined.gif'
    imageio.mimsave(combined_gif, frames, duration=300, loop=0)
    return confidences, combined_gif


gr.Interface(
    fn=inference,
    inputs=[ 
        gr.Video(type="file"),
        gr.Radio(
            ['K400', 'SSv2', 'UCF'], 
            label='Dataset', value='K400'
        ),
    ],
    outputs=[
        gr.Label(num_top_classes=3, label='confidence scores'),
        gr.Image(type="filepath", label='reconstructed masked autoencoder')
    ],
    examples=[
        ["examples/k400.mp4"],
        ["examples/k400.mp4"],
        ["examples/k400.mp4"],
    ],
    title="VideoMAE",
).launch()