convertPb2Tfrecord.py
3.92 KB
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
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
import tensorflow as tf
import glob, os
import numpy
def _make_bytes(int_array):
if bytes == str: # Python2
return ''.join(map(chr, int_array))
else:
return bytes(int_array)
def quantize(features, min_quantized_value=-2.0, max_quantized_value=2.0):
"""Quantizes float32 `features` into string."""
assert features.dtype == 'float32'
assert len(features.shape) == 1 # 1-D array
features = numpy.clip(features, min_quantized_value, max_quantized_value)
quantize_range = max_quantized_value - min_quantized_value
features = (features - min_quantized_value) * (255.0 / quantize_range)
features = [int(round(f)) for f in features]
return _make_bytes(features)
# for parse feature.pb
contexts = {
'AUDIO/feature/dimensions': tf.io.FixedLenFeature([], tf.int64),
'AUDIO/feature/rate': tf.io.FixedLenFeature([], tf.float32),
'RGB/feature/dimensions': tf.io.FixedLenFeature([], tf.int64),
'RGB/feature/rate': tf.io.FixedLenFeature([], tf.float32),
'clip/data_path': tf.io.FixedLenFeature([], tf.string),
'clip/end/timestamp': tf.io.FixedLenFeature([], tf.int64),
'clip/start/timestamp': tf.io.FixedLenFeature([], tf.int64)
}
features = {
'AUDIO/feature/floats': tf.io.VarLenFeature(dtype=tf.float32),
'AUDIO/feature/timestamp': tf.io.VarLenFeature(tf.int64),
'RGB/feature/floats': tf.io.VarLenFeature(dtype=tf.float32),
'RGB/feature/timestamp': tf.io.VarLenFeature(tf.int64)
}
def parse_exmp(serial_exmp):
_, sequence_parsed = tf.io.parse_single_sequence_example(
serialized=serial_exmp,
context_features=contexts,
sequence_features=features)
sequence_parsed = tf.contrib.learn.run_n(sequence_parsed)[0]
audio = sequence_parsed['AUDIO/feature/floats'].values
rgb = sequence_parsed['RGB/feature/floats'].values
# print(audio.values)
# print(type(audio.values))
# audio is 128 8bit, rgb is 1024 8bit for every second
audio_slices = [audio[128 * i: 128 * (i + 1)] for i in range(len(audio) // 128)]
rgb_slices = [rgb[1024 * i: 1024 * (i + 1)] for i in range(len(rgb) // 1024)]
byte_audio = []
byte_rgb = []
for seg in audio_slices:
audio_seg = quantize(seg)
byte_audio.append(audio_seg)
for seg in rgb_slices:
rgb_seg = quantize(seg)
byte_rgb.append(rgb_seg)
return byte_audio, byte_rgb
def make_exmp(id, labels, audio, rgb):
audio_features = []
rgb_features = []
for embedding in audio:
embedding_feature = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[embedding]))
audio_features.append(embedding_feature)
for embedding in rgb:
embedding_feature = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[embedding]))
rgb_features.append(embedding_feature)
# for construct yt8m data
seq_exmp = tf.train.SequenceExample(
context=tf.train.Features(
feature={
'id': tf.train.Feature(bytes_list=tf.train.BytesList(
value=[id.encode('utf-8')])),
'labels': tf.train.Feature(int64_list=tf.train.Int64List(
value=[labels]))
}),
feature_lists=tf.train.FeatureLists(
feature_list={
'audio': tf.train.FeatureList(
feature=audio_features
),
'rgb': tf.train.FeatureList(
feature=rgb_features
)
})
)
serialized = seq_exmp.SerializeToString()
return serialized
if __name__ == '__main__':
filename = '/tmp/mediapipe/features.pb'
sequence_example = open(filename, 'rb').read()
audio, rgb = parse_exmp(sequence_example)
id = 'test_001'
labels = 1
tmp_example = make_exmp(id, labels, audio, rgb)
decoded = tf.train.SequenceExample.FromString(tmp_example)
print(decoded)
# then you can write tmp_example to tfrecord files