Datasets:
Create load_viddiff_dataset.py
Browse files- load_viddiff_dataset.py +368 -0
load_viddiff_dataset.py
ADDED
@@ -0,0 +1,368 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import ipdb
|
2 |
+
import pdb
|
3 |
+
import os
|
4 |
+
import numpy as np
|
5 |
+
import json
|
6 |
+
import re
|
7 |
+
from PIL import Image
|
8 |
+
from pathlib import Path
|
9 |
+
from datasets import load_dataset
|
10 |
+
import decord
|
11 |
+
from tqdm import tqdm
|
12 |
+
import logging
|
13 |
+
import hashlib
|
14 |
+
|
15 |
+
|
16 |
+
def load_viddiff_dataset(splits=["easy"], subset_mode="0", cache_dir=None, test_new=False):
|
17 |
+
"""
|
18 |
+
splits in ['easy', 'medium', 'hard']
|
19 |
+
"""
|
20 |
+
if not test_new:
|
21 |
+
dataset = load_dataset("viddiff/VidDiffBench_2", cache_dir=cache_dir)
|
22 |
+
dataset = dataset['test']
|
23 |
+
valid_splits = set(dataset['split'])
|
24 |
+
else:
|
25 |
+
dataset = load_dataset("viddiff/VidDiffBench_2", cache_dir=cache_dir)
|
26 |
+
dataset = dataset['test']
|
27 |
+
dataset = dataset.map(lambda example: example.update({'split': example['domain']}) or example)
|
28 |
+
valid_splits = set(dataset['split'])
|
29 |
+
|
30 |
+
def _filter_splits(example):
|
31 |
+
return example["split"] in splits
|
32 |
+
|
33 |
+
dataset = dataset.filter(_filter_splits)
|
34 |
+
if len(dataset) == 0:
|
35 |
+
raise ValueError(
|
36 |
+
f"Dataset empty for splits {splits}. Valid splits {valid_splits}")
|
37 |
+
|
38 |
+
def _map_elements_to_json(example):
|
39 |
+
example["videos"] = json.loads(example["videos"])
|
40 |
+
example["differences_annotated"] = json.loads(
|
41 |
+
example["differences_annotated"])
|
42 |
+
example["differences_gt"] = json.loads(example["differences_gt"])
|
43 |
+
return example
|
44 |
+
|
45 |
+
dataset = dataset.map(_map_elements_to_json)
|
46 |
+
# dataset = dataset.map(_clean_annotations)
|
47 |
+
dataset = apply_subset_mode(dataset, subset_mode)
|
48 |
+
|
49 |
+
dataset = _get_difficulty_splits(dataset)
|
50 |
+
|
51 |
+
return dataset
|
52 |
+
|
53 |
+
|
54 |
+
def _get_difficulty_splits(dataset):
|
55 |
+
with open("data/lookup_action_to_split.json", "r") as fp:
|
56 |
+
lookup_action_to_split = json.load(fp)
|
57 |
+
|
58 |
+
def add_split_difficulty(example):
|
59 |
+
example['split_difficulty'] = lookup_action_to_split[example['action']]
|
60 |
+
return example
|
61 |
+
|
62 |
+
dataset = dataset.map(add_split_difficulty)
|
63 |
+
return dataset
|
64 |
+
|
65 |
+
|
66 |
+
def load_all_videos(dataset,
|
67 |
+
cache=True,
|
68 |
+
cache_dir="cache/cache_data",
|
69 |
+
overwrite_cache=False,
|
70 |
+
test_samevideo=0,
|
71 |
+
test_flipvids=0,
|
72 |
+
do_tqdm=True):
|
73 |
+
"""
|
74 |
+
Return a 2-element tuple. Each element is a list of length len(datset).
|
75 |
+
First list is video A for each datapoint as a dict with elements
|
76 |
+
path: original path to video
|
77 |
+
fps: frames per second
|
78 |
+
video: numpy array of the video shape (nframes,H,W,3)
|
79 |
+
Second list is the same but for video B.
|
80 |
+
|
81 |
+
Args:
|
82 |
+
cache_dir (str): Directory to store cached video data. Defaults to "cache/cache_data"
|
83 |
+
"""
|
84 |
+
|
85 |
+
all_videos = ([], [])
|
86 |
+
# make iterator, with or without tqdm based on `do_tqdm`
|
87 |
+
if do_tqdm:
|
88 |
+
it = tqdm(dataset)
|
89 |
+
else:
|
90 |
+
it = dataset
|
91 |
+
|
92 |
+
# load each video
|
93 |
+
for row in it:
|
94 |
+
videos = get_video_data(row['videos'],
|
95 |
+
cache=cache,
|
96 |
+
cache_dir=cache_dir,
|
97 |
+
overwrite_cache=overwrite_cache)
|
98 |
+
|
99 |
+
video0, video1 = videos[0], videos[1]
|
100 |
+
|
101 |
+
if test_flipvids:
|
102 |
+
video0, video1 = video1, video0
|
103 |
+
|
104 |
+
if not test_samevideo:
|
105 |
+
all_videos[0].append(video0)
|
106 |
+
all_videos[1].append(video1)
|
107 |
+
else:
|
108 |
+
all_videos[0].append(video1)
|
109 |
+
all_videos[1].append(video1)
|
110 |
+
|
111 |
+
return all_videos
|
112 |
+
|
113 |
+
|
114 |
+
def _clean_annotations(example):
|
115 |
+
# Not all differences in the taxonomy may have a label available, so filter them.
|
116 |
+
|
117 |
+
differences_gt_labeled = {
|
118 |
+
k: v
|
119 |
+
for k, v in example['differences_gt'].items() if v is not None
|
120 |
+
}
|
121 |
+
differences_annotated = {
|
122 |
+
k: v
|
123 |
+
for k, v in example['differences_annotated'].items()
|
124 |
+
if k in differences_gt_labeled.keys()
|
125 |
+
}
|
126 |
+
|
127 |
+
# Directly assign to the example without deepcopy
|
128 |
+
example['differences_gt'] = differences_gt_labeled
|
129 |
+
example['differences_annotated'] = differences_annotated
|
130 |
+
|
131 |
+
return example
|
132 |
+
|
133 |
+
|
134 |
+
def get_video_data(videos: dict, cache=True, cache_dir="cache/cache_data", overwrite_cache=False):
|
135 |
+
"""
|
136 |
+
Pass in the videos dictionary from the dataset, like dataset[idx]['videos'].
|
137 |
+
Load the 2 videos represented as numpy arrays.
|
138 |
+
By default, cache the arrays ... so the second time through, the dataset
|
139 |
+
loading will be faster.
|
140 |
+
|
141 |
+
returns: video0, video1
|
142 |
+
"""
|
143 |
+
video_dicts = []
|
144 |
+
|
145 |
+
for i in [0, 1]:
|
146 |
+
path = videos[i]['path']
|
147 |
+
assert Path(path).exists(
|
148 |
+
), f"Video not downloaded [{path}]\nCheck dataset README about downloading videos"
|
149 |
+
frames_trim = slice(*videos[i]['frames_trim'])
|
150 |
+
|
151 |
+
video_dict = videos[i].copy()
|
152 |
+
|
153 |
+
if cache:
|
154 |
+
dir_cache = Path(cache_dir)
|
155 |
+
dir_cache.mkdir(exist_ok=True, parents=True)
|
156 |
+
hash_key = get_hash_key(path + str(frames_trim))
|
157 |
+
memmap_filename = dir_cache / f"memmap_{hash_key}.npy"
|
158 |
+
|
159 |
+
# if not in the cache, and not overwriting, then get OG video
|
160 |
+
if os.path.exists(memmap_filename) and not overwrite_cache:
|
161 |
+
video_info = np.load(f"{memmap_filename}.info.npy",
|
162 |
+
allow_pickle=True).item()
|
163 |
+
video = np.memmap(memmap_filename,
|
164 |
+
dtype=video_info['dtype'],
|
165 |
+
mode='r',
|
166 |
+
shape=video_info['shape'])
|
167 |
+
video_dict['video'] = video
|
168 |
+
video_dict['fps'] = video_dict['fps_original'] # since we don't downsample here
|
169 |
+
video_dicts.append(video_dict)
|
170 |
+
continue
|
171 |
+
|
172 |
+
is_dir = Path(path).is_dir()
|
173 |
+
if is_dir:
|
174 |
+
video = _load_video_from_directory_of_images(
|
175 |
+
path, frames_trim=frames_trim)
|
176 |
+
|
177 |
+
else:
|
178 |
+
assert Path(path).suffix in (".mp4", ".mov")
|
179 |
+
video, fps = _load_video(path, frames_trim=frames_trim)
|
180 |
+
assert fps == videos[i]['fps_original']
|
181 |
+
|
182 |
+
if cache:
|
183 |
+
np.save(f"{memmap_filename}.info.npy", {
|
184 |
+
'shape': video.shape,
|
185 |
+
'dtype': video.dtype
|
186 |
+
})
|
187 |
+
memmap = np.memmap(memmap_filename,
|
188 |
+
dtype=video.dtype,
|
189 |
+
mode='w+',
|
190 |
+
shape=video.shape)
|
191 |
+
memmap[:] = video[:]
|
192 |
+
memmap.flush()
|
193 |
+
video = memmap
|
194 |
+
|
195 |
+
video_dict['video'] = video
|
196 |
+
video_dict['fps'] = video_dict['fps_original']
|
197 |
+
video_dicts.append(video_dict)
|
198 |
+
|
199 |
+
return video_dicts
|
200 |
+
|
201 |
+
|
202 |
+
def _load_video(f, return_fps=True, frames_trim: slice = None) -> np.ndarray:
|
203 |
+
"""
|
204 |
+
mp4 video to frames numpy array shape (N,H,W,3).
|
205 |
+
Do not use for long videos
|
206 |
+
frames_trim: (s,e) is start and end int frames to include (warning, the range
|
207 |
+
is inclusive, unlike in list indexing.)
|
208 |
+
"""
|
209 |
+
vid = decord.VideoReader(str(f))
|
210 |
+
fps = vid.get_avg_fps()
|
211 |
+
|
212 |
+
if len(vid) > 50000:
|
213 |
+
raise ValueError(
|
214 |
+
"Video probably has too many frames to convert to a numpy")
|
215 |
+
|
216 |
+
if frames_trim is None:
|
217 |
+
frames_trim = slice(0, None, None)
|
218 |
+
video_np = vid[frames_trim].asnumpy()
|
219 |
+
|
220 |
+
if not return_fps:
|
221 |
+
return video_np
|
222 |
+
else:
|
223 |
+
assert fps > 0
|
224 |
+
return video_np, fps
|
225 |
+
|
226 |
+
|
227 |
+
def _load_video_from_directory_of_images(
|
228 |
+
path_dir: str,
|
229 |
+
frames_trim: slice = None,
|
230 |
+
downsample_time: int = None,
|
231 |
+
) -> np.ndarray:
|
232 |
+
"""
|
233 |
+
|
234 |
+
`path_dir` is a directory path with images that, when arranged in alphabetical
|
235 |
+
order, make a video.
|
236 |
+
This function returns the a numpy array shape (N,H,W,3) where N is the
|
237 |
+
number of frames.
|
238 |
+
"""
|
239 |
+
files = sorted(os.listdir(path_dir))
|
240 |
+
|
241 |
+
if frames_trim is not None:
|
242 |
+
files = files[frames_trim]
|
243 |
+
|
244 |
+
if downsample_time is not None:
|
245 |
+
files = files[::downsample_time]
|
246 |
+
|
247 |
+
files = [f"{path_dir}/{f}" for f in files]
|
248 |
+
images = [Image.open(f) for f in files]
|
249 |
+
|
250 |
+
video_array = np.stack(images)
|
251 |
+
|
252 |
+
return video_array
|
253 |
+
|
254 |
+
|
255 |
+
def _subsample_video(video: np.ndarray,
|
256 |
+
fps_original: int,
|
257 |
+
fps_target: int,
|
258 |
+
fps_warning: bool = True):
|
259 |
+
"""
|
260 |
+
video: video as numby array (nframes, h, w, 3)
|
261 |
+
fps_original: original fps of the video
|
262 |
+
fps_target: target fps to downscale to
|
263 |
+
fps_warning: if True, then log warnings to logger if the target fps is
|
264 |
+
higher than original fps, or if the target fps isn't possible because
|
265 |
+
it isn't divisible by the original fps.
|
266 |
+
"""
|
267 |
+
subsample_time = fps_original / fps_target
|
268 |
+
|
269 |
+
if subsample_time < 1 and fps_warning:
|
270 |
+
logging.warning(f"Trying to subsample frames to fps {fps_target}, which "\
|
271 |
+
"is higher than the fps of the original video which is "\
|
272 |
+
"{video['fps']}. The video fps won't be changed for {video['path']}. "\
|
273 |
+
f"\nSupress this warning by setting config fps_warning=False")
|
274 |
+
return video, fps_original, 1
|
275 |
+
|
276 |
+
subsample_time_int = int(subsample_time)
|
277 |
+
fps_new = int(fps_original / subsample_time_int)
|
278 |
+
if fps_new != fps_target and fps_warning:
|
279 |
+
logging.warning(f"Config lmm.fps='{fps_target}' but the original fps is {fps_original} " \
|
280 |
+
f"so we downscale to fps {fps_new} instead. " \
|
281 |
+
f"\nSupress this warning by setting config fps_warning=False")
|
282 |
+
|
283 |
+
video = video[::subsample_time_int]
|
284 |
+
|
285 |
+
return video, fps_new, subsample_time_int
|
286 |
+
|
287 |
+
def downsample_videos(dataset, videos, args_fps_inference, fps_warning=True):
|
288 |
+
"""To fix some hacky - oOnly called by viddiff_method.run_viddiff.py """
|
289 |
+
for i in range(len(dataset)):
|
290 |
+
row = dataset[i]
|
291 |
+
domain = row['domain']
|
292 |
+
fps_inference = args_fps_inference[domain]
|
293 |
+
video0, video1 = videos[0][i], videos[1][i]
|
294 |
+
for video in (video0, video1):
|
295 |
+
video['video'], fps_new, subsample_time_int = _subsample_video(
|
296 |
+
video['video'], video['fps_original'], fps_inference, fps_warning)
|
297 |
+
video['fps'] = fps_new
|
298 |
+
|
299 |
+
return videos
|
300 |
+
|
301 |
+
|
302 |
+
def apply_subset_mode(dataset, subset_mode):
|
303 |
+
"""
|
304 |
+
For example if subset_mode is "3_per_action" then just get the first 3 rows
|
305 |
+
for each unique action.
|
306 |
+
Useful for working with subsets.
|
307 |
+
"""
|
308 |
+
match = re.match(r"(\d+)_per_action", subset_mode)
|
309 |
+
if match:
|
310 |
+
instances_per_action = int(match.group(1))
|
311 |
+
action_counts = {}
|
312 |
+
subset_indices = []
|
313 |
+
|
314 |
+
for idx, example in enumerate(dataset):
|
315 |
+
action = example['action']
|
316 |
+
if action not in action_counts:
|
317 |
+
action_counts[action] = 0
|
318 |
+
|
319 |
+
if action_counts[action] < instances_per_action:
|
320 |
+
subset_indices.append(idx)
|
321 |
+
action_counts[action] += 1
|
322 |
+
|
323 |
+
return dataset.select(subset_indices)
|
324 |
+
else:
|
325 |
+
return dataset
|
326 |
+
|
327 |
+
|
328 |
+
def get_hash_key(key: str) -> str:
|
329 |
+
return hashlib.sha256(key.encode()).hexdigest()
|
330 |
+
|
331 |
+
|
332 |
+
def dataset_metrics(dataset):
|
333 |
+
import pandas as pd
|
334 |
+
df = pd.DataFrame(dataset)
|
335 |
+
print("Number of actions ")
|
336 |
+
print(df.groupby(['split'])['action'].nunique())
|
337 |
+
print("Total actions", df['action'].nunique())
|
338 |
+
|
339 |
+
print("Samples by category")
|
340 |
+
print(df.groupby(["split"])['split'].count())
|
341 |
+
print("Total ", len(df))
|
342 |
+
print()
|
343 |
+
|
344 |
+
diffs = []
|
345 |
+
for row in dataset:
|
346 |
+
diff = {
|
347 |
+
k: v
|
348 |
+
for k, v in row['differences_gt'].items() if v is not None
|
349 |
+
}
|
350 |
+
diffs.append(diff)
|
351 |
+
cnts = [len(d) for d in diffs]
|
352 |
+
df['variation_cnts'] = cnts
|
353 |
+
print("Variation counts by category")
|
354 |
+
print(df.groupby(['split'])['variation_cnts'].sum())
|
355 |
+
print("total ", df['variation_cnts'].sum())
|
356 |
+
|
357 |
+
|
358 |
+
print()
|
359 |
+
|
360 |
+
|
361 |
+
if __name__ == "__main__":
|
362 |
+
# these are the 3 data loading commands
|
363 |
+
splits = ['ballsports', 'fitness', 'diving', 'music', 'surgery']
|
364 |
+
dataset = load_viddiff_dataset(splits=splits)
|
365 |
+
metrics = dataset_metrics(dataset)
|
366 |
+
|
367 |
+
videos = load_all_videos(dataset)
|
368 |
+
n_differences = lvd.get_n_differences(dataset, "data/n_differences.json")
|