chop videos in chunks
Browse files- yttemporal180m.py +55 -33
yttemporal180m.py
CHANGED
@@ -20,15 +20,16 @@ _URL_BASE = "https://rowanzellers.com/merlot/#data"
|
|
20 |
|
21 |
url_numbers = ["00" + str(i) if i < 10 else "0" + str(i) for i in range(100)]
|
22 |
_DL_URLS = [
|
23 |
-
f"https://storage.googleapis.com/merlot/yttemporal180m/yttemporal180m_{num}of100.jsonl.gz"
|
|
|
|
|
24 |
|
25 |
|
26 |
def json_serializer(o):
|
27 |
if isinstance(o, datetime):
|
28 |
return str(o)
|
29 |
|
30 |
-
raise TypeError(
|
31 |
-
f"Object of type {o.__class__.__name__} is not JSON serializable")
|
32 |
|
33 |
|
34 |
class yttemporal180mConfig(datasets.BuilderConfig):
|
@@ -36,7 +37,8 @@ class yttemporal180mConfig(datasets.BuilderConfig):
|
|
36 |
|
37 |
def __init__(self, **kwargs):
|
38 |
super(yttemporal180mConfig, self).__init__(
|
39 |
-
version=datasets.Version("2.1.0", ""), **kwargs
|
|
|
40 |
|
41 |
|
42 |
class yttemporal180m(datasets.GeneratorBasedBuilder):
|
@@ -44,7 +46,8 @@ class yttemporal180m(datasets.GeneratorBasedBuilder):
|
|
44 |
DEFAULT_CONFIG_NAME = "default"
|
45 |
BUILDER_CONFIGS = [
|
46 |
yttemporal180mConfig(
|
47 |
-
name="default", description="Default full yttemporal180m dataset"
|
|
|
48 |
]
|
49 |
|
50 |
def _info(self):
|
@@ -54,8 +57,9 @@ class yttemporal180m(datasets.GeneratorBasedBuilder):
|
|
54 |
{
|
55 |
"video_id": datasets.Value("string"),
|
56 |
"video_url": datasets.Value("string"),
|
57 |
-
"
|
58 |
-
"
|
|
|
59 |
"meta": datasets.Value("string"),
|
60 |
}
|
61 |
),
|
@@ -65,15 +69,12 @@ class yttemporal180m(datasets.GeneratorBasedBuilder):
|
|
65 |
)
|
66 |
|
67 |
def _split_generators(self, dl_manager):
|
68 |
-
archive_paths = [dl_manager.download_and_extract(
|
69 |
-
url) for url in _DL_URLS]
|
70 |
|
71 |
train_split = [
|
72 |
datasets.SplitGenerator(
|
73 |
name=datasets.Split.TRAIN,
|
74 |
-
gen_kwargs={
|
75 |
-
"jsonl_files": archive_paths
|
76 |
-
},
|
77 |
)
|
78 |
]
|
79 |
|
@@ -88,26 +89,47 @@ class yttemporal180m(datasets.GeneratorBasedBuilder):
|
|
88 |
for json_str in json_list:
|
89 |
infos = json.loads(json_str)
|
90 |
|
91 |
-
id = infos[
|
92 |
url = "https://www.youtube.com/watch?v=" + id
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
}
|
102 |
-
yield idx, {
|
103 |
-
"video_id": id,
|
104 |
-
"video_url": url,
|
105 |
-
"asr": asr,
|
106 |
-
"title": infos['info']['title'],
|
107 |
-
"meta": json.dumps(
|
108 |
-
metadata_dict,
|
109 |
-
default=json_serializer,
|
110 |
-
indent=2
|
111 |
)
|
112 |
-
|
113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
url_numbers = ["00" + str(i) if i < 10 else "0" + str(i) for i in range(100)]
|
22 |
_DL_URLS = [
|
23 |
+
f"https://storage.googleapis.com/merlot/yttemporal180m/yttemporal180m_{num}of100.jsonl.gz"
|
24 |
+
for num in url_numbers
|
25 |
+
]
|
26 |
|
27 |
|
28 |
def json_serializer(o):
|
29 |
if isinstance(o, datetime):
|
30 |
return str(o)
|
31 |
|
32 |
+
raise TypeError(f"Object of type {o.__class__.__name__} is not JSON serializable")
|
|
|
33 |
|
34 |
|
35 |
class yttemporal180mConfig(datasets.BuilderConfig):
|
|
|
37 |
|
38 |
def __init__(self, **kwargs):
|
39 |
super(yttemporal180mConfig, self).__init__(
|
40 |
+
version=datasets.Version("2.1.0", ""), **kwargs
|
41 |
+
)
|
42 |
|
43 |
|
44 |
class yttemporal180m(datasets.GeneratorBasedBuilder):
|
|
|
46 |
DEFAULT_CONFIG_NAME = "default"
|
47 |
BUILDER_CONFIGS = [
|
48 |
yttemporal180mConfig(
|
49 |
+
name="default", description="Default full yttemporal180m dataset"
|
50 |
+
),
|
51 |
]
|
52 |
|
53 |
def _info(self):
|
|
|
57 |
{
|
58 |
"video_id": datasets.Value("string"),
|
59 |
"video_url": datasets.Value("string"),
|
60 |
+
"caption": datasets.Value("string"),
|
61 |
+
"timestamp_start": datasets.Value("float32"),
|
62 |
+
"timestamp_stop": datasets.Value("float32"),
|
63 |
"meta": datasets.Value("string"),
|
64 |
}
|
65 |
),
|
|
|
69 |
)
|
70 |
|
71 |
def _split_generators(self, dl_manager):
|
72 |
+
archive_paths = [dl_manager.download_and_extract(url) for url in _DL_URLS]
|
|
|
73 |
|
74 |
train_split = [
|
75 |
datasets.SplitGenerator(
|
76 |
name=datasets.Split.TRAIN,
|
77 |
+
gen_kwargs={"jsonl_files": archive_paths},
|
|
|
|
|
78 |
)
|
79 |
]
|
80 |
|
|
|
89 |
for json_str in json_list:
|
90 |
infos = json.loads(json_str)
|
91 |
|
92 |
+
id = infos["info"]["display_id"]
|
93 |
url = "https://www.youtube.com/watch?v=" + id
|
94 |
+
|
95 |
+
# Divide video by segments of 15 sec
|
96 |
+
max_sec_per_segment = 15
|
97 |
+
num_chunks = (
|
98 |
+
int(
|
99 |
+
divmod(infos["subtitles"][-1]["time"], max_sec_per_segment)[
|
100 |
+
0
|
101 |
+
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
102 |
)
|
103 |
+
+ 1
|
104 |
+
)
|
105 |
+
time_chunks = [
|
106 |
+
i * max_sec_per_segment for i in range(num_chunks + 1)
|
107 |
+
]
|
108 |
+
time_chunk_idx = 0
|
109 |
+
caption = ""
|
110 |
+
for el in infos["subtitles"]:
|
111 |
+
if el["time"] <= time_chunks[time_chunk_idx + 1]:
|
112 |
+
caption += el["word"] + " "
|
113 |
+
else:
|
114 |
+
timestamp_start = float(time_chunks[time_chunk_idx])
|
115 |
+
timestamp_stop = float(time_chunks[time_chunk_idx + 1])
|
116 |
+
time_chunk_idx += 1
|
117 |
+
|
118 |
+
metadata_dict = {
|
119 |
+
"asr_info": infos["denoised"],
|
120 |
+
"info": infos["info"],
|
121 |
+
"subtitles": infos["subtitles"],
|
122 |
+
"title": infos["info"]["title"],
|
123 |
+
}
|
124 |
+
yield idx, {
|
125 |
+
"video_id": id,
|
126 |
+
"video_url": url,
|
127 |
+
"caption": caption,
|
128 |
+
"timestamp_start": timestamp_start,
|
129 |
+
"timestamp_stop": timestamp_stop,
|
130 |
+
"meta": json.dumps(
|
131 |
+
metadata_dict, default=json_serializer, indent=2
|
132 |
+
),
|
133 |
+
}
|
134 |
+
idx += 1
|
135 |
+
caption = ""
|