-
Notifications
You must be signed in to change notification settings - Fork 1
/
dataset_utils.py
278 lines (227 loc) · 11.6 KB
/
dataset_utils.py
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
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
from collections import OrderedDict
import whisper
import glob
import torch
from torch.utils.data import Dataset, DataLoader, random_split
# import torchaudio
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, ToPILImage, Normalize
import cv2
import time
import numpy as np
import matplotlib.pyplot as plt
from tqdm.auto import tqdm
import os
import json
def _convert_image_to_rgb(image):
return image.convert("RGB")
def _transform(n_px):
return Compose([
ToPILImage(),
Resize(n_px),
CenterCrop(n_px),
_convert_image_to_rgb,
ToTensor(),
# Normalization params from 1/6th of GreatestHits dataset
Normalize((0.4738, 0.4298, 0.3702),
(0.1994, 0.1913, 0.1805)),
])
class GreatestHitsDataset(Dataset):
def __init__(self, root_dir="./vis-data-256", audio_length=5, use_cached=True, test=False):
self.root_dir = root_dir
self.transform = _transform(224)
# self.transform_to_tensor = T.ToTensor()
self.audio_rate = 96000
self.audio_length = audio_length
self.use_cached = use_cached
self.dataset_cache_dir = root_dir + "-processed"
if use_cached:
# Load times_info from cached json
with open(f"{self.dataset_cache_dir}/times_info.json", "r") as f:
times_info = json.load(f)
if test:
with open(f"test_set_times_info.json", "r") as f:
self.times_info = json.load(f)
else:
self.times_info = times_info
else:
self.times_info = OrderedDict()
for file in sorted(glob.glob(root_dir + "/*_times.txt")):
with open(file, "r") as f:
data = f.readlines()
data = [line.strip().split() for line in data]
data = [(float(line[0]), line[1]) for line in data if (line[1] != "None") and (line[2] == "hit")]
self.times_info[file.split("/")[-1].split("_")[0]] = data
self.all_material_names = set()
for _, frame_info in times_info.items():
self.all_material_names.update([frame[1] for frame in frame_info])
self.all_material_names = sorted(list(self.all_material_names))
self.mat_to_ind = {name: i for i, name in enumerate(self.all_material_names)}
self.ind_to_mat = {i: name for i, name in enumerate(self.all_material_names)}
def __len__(self):
total_len = 0
for key in self.times_info.keys():
total_len += len(self.times_info[key])
return total_len
def _dataset_idx_to_times_info_key_idx(self, idx):
for key in self.times_info.keys():
if idx < len(self.times_info[key]):
break
else:
idx -= len(self.times_info[key])
date_time = key
return date_time, idx
def __getitem__(self, idx):
date_time, idx = self._dataset_idx_to_times_info_key_idx(idx)
frames_info = self.times_info[date_time]
frame_timestamp = frames_info[idx][0]
material_name = frames_info[idx][1]
material_index = self.mat_to_ind[material_name]
material_index = torch.tensor(material_index)
if self.use_cached:
mel = np.load(f"{self.dataset_cache_dir}/{date_time}_{frame_timestamp}_mel.npy")
mel = torch.tensor(mel)
# Load jpg
frame = cv2.imread(f"{self.dataset_cache_dir}/{date_time}_{frame_timestamp}_frame.jpg")
frame = self.transform(frame)
else:
# frames, audio, metadata = torchvision.io.read_video(f"./vis-data-256/{date_time}_denoised.mp4")
# Loading single frame with cv2 for faster frame access compared to loading whole video
cap = cv2.VideoCapture(self.root_dir + f"/{date_time}_denoised.mp4")
audio = whisper.load_audio(self.root_dir + f"/{date_time}_denoised.wav", self.audio_rate)
audio_start_time = frame_timestamp - self.audio_length / 2
audio_start_idx = int(audio_start_time * self.audio_rate)
audio = audio[audio_start_idx : audio_start_idx + self.audio_rate * self.audio_length]
audio = whisper.pad_or_trim(audio, self.audio_rate * self.audio_length)
# print(date_time, frame_timestamp, audio.shape)
mel = whisper.log_mel_spectrogram(audio)
cap.set(cv2.CAP_PROP_POS_MSEC, frame_timestamp * 1000)
ret, frame = cap.read()
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
# frame = (self.transform_to_tensor(frame) * 255).to(torch.uint8)
frame = self.transform(frame)
cap.release()
# return {"images": frame, "audios": mel, "audios_raw":audio, "materials": material_index}
return {"images": frame, "audios": mel, "materials": material_index}
# Create DataLoader
def create_dataloaders(root_dir="./vis-data-256", batch_size=4, val_ratio=0.05, use_cached=True, test=False):
dataset = GreatestHitsDataset(root_dir, use_cached=use_cached, test=test)
print(f"\n{'Test' if test else 'Training'} dataset size: {len(dataset)}\n")
if test:
train_set = dataset
val_set = None
else:
val_size = int(val_ratio * len(dataset))
train_size = len(dataset) - val_size
train_set, val_set = random_split(dataset, [train_size, val_size])
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=True) if val_set is not None else None
return train_loader, val_loader
def test_cached_vs_uncached_datasets(n_test=5, batch_size=4, compare_diff_images=False):
print("Loading cached dataset...")
start_time = time.time()
cached_dataset = GreatestHitsDataset(root_dir=dataset_dir, use_cached=True)
print(f"Time taken to initialize cached dataset: {time.time() - start_time:.2f} seconds")
print("Loading uncached dataset...")
start_time = time.time()
uncached_dataset = GreatestHitsDataset(root_dir=dataset_dir, use_cached=False)
print(f"Time taken to initialize uncached dataset: {time.time() - start_time:.2f} seconds")
num_test_samples = n_test
sample_indices = np.random.choice(len(cached_dataset), num_test_samples, replace=False)
all_passed = True
for i in sample_indices:
cached_data = cached_dataset[i]
uncached_data = uncached_dataset[i]
for key in cached_data.keys():
if not torch.allclose(cached_data[key], uncached_data[key], atol=1e-2):
print(f"Test {i+1} failed")
print(f"{key} not equal for test {i+1} - (index {i})")
print(f"Shapes: {cached_data['images'].shape}, {uncached_data['images'].shape}")
print(f"Max: {cached_data['images'].max()}, {uncached_data['images'].max()}")
print(f"Min: {cached_data['images'].min()}, {uncached_data['images'].min()}")
print(f"Mean: {cached_data['images'].mean()}, {uncached_data['images'].mean()}")
print(f"Std: {cached_data['images'].std()}, {uncached_data['images'].std()}")
if key == "images" and compare_diff_images:
# Save both images for comparison
cached_img = cached_data[key].permute(1, 2, 0).numpy().clip(0,1)
uncached_img = uncached_data[key].permute(1, 2, 0).numpy().clip(0,1)
plt.imsave(f"cached_img_{i+1}.jpg", cached_img)
plt.imsave(f"uncached_img_{i+1}.jpg", uncached_img)
all_passed = False
break
else:
print(f"Test {i+1} passed")
if all_passed:
print(f"All {num_test_samples} tests passed, datasets are same!")
cached_loader = DataLoader(cached_dataset, batch_size=batch_size, shuffle=True)
start_time = time.time()
avg_cached_time = 0
for i, data in enumerate(cached_loader):
print([data[k].shape for k in data])
time_taken = time.time() - start_time
print(f"Time taken to load cached batch {i+1}: {time_taken:.2f} seconds")
avg_cached_time += time_taken
if i == n_test - 1:
break
start_time = time.time()
print(f"\nAverage time taken to load a cached batch: {avg_cached_time / n_test:.2f} seconds\n")
uncached_loader = DataLoader(uncached_dataset, batch_size=batch_size, shuffle=True)
start_time = time.time()
avg_time = 0
for i, data in enumerate(uncached_loader):
print([data[k].shape for k in data])
time_taken = time.time() - start_time
print(f"Time taken to load batch {i+1}: {time_taken:.2f} seconds")
avg_time += time_taken
if i == n_test - 1:
break
start_time = time.time()
print(f"\nAverage time taken to load a batch: {avg_time / n_test:.2f} seconds")
print(f"\nAverage speedup by caching: {avg_time / avg_cached_time:.2f}x")
def test_dataloader():
batch_size = 4
start_time = time.time()
train_loader, _ = create_dataloaders(root_dir=dataset_dir, batch_size=batch_size)
print(f"Time taken to create dataloader (with batch size {batch_size}): {time.time() - start_time:.2f} seconds")
start_time = time.time()
for i, data in enumerate(train_loader):
print([data[k].shape for k in data])
print(f"\nTime taken to load batch {i+1}: {time.time() - start_time:.2f} seconds")
plt.imsave(f"image_{i+1}.jpg", data["images"][0].permute(1, 2, 0).numpy().clip(0, 1))
mel = data["audios"][0].numpy()
plt.imsave(f"mel_{i+1}.jpg", mel, cmap="viridis")
# print(f"Min, Max for mel: {mel.min(), mel.max()}")
# if "audios_raw" in data:
# torchaudio.save(f"audio_{i+1}.wav", data["audios_raw"][0].unsqueeze(0), 96000)
if i == 1:
break
start_time = time.time()
def check_if_entire_dataset_cached():
dataset = GreatestHitsDataset(root_dir=dataset_dir, use_cached=True)
uncached = {"frame": [], "mel": []}
for frame_timestamp, frame_info in tqdm(dataset.times_info.items(), desc="Checking cached dataset"):
for frame in frame_info:
mel_path = f"{dataset.dataset_cache_dir}/{frame_timestamp}_{frame[0]}_mel.npy"
frame_path = f"{dataset.dataset_cache_dir}/{frame_timestamp}_{frame[0]}_frame.jpg"
if not os.path.exists(mel_path):
print(f"Mel not cached for {frame_timestamp}_{frame[0]}")
uncached["mel"].append((frame_timestamp, frame[0]))
if not os.path.exists(frame_path):
print(f"Frame not cached for {frame_timestamp}_{frame[0]}")
uncached["frame"].append((frame_timestamp, frame[0]))
if len(uncached["mel"]) == 0 and len(uncached["frame"]) == 0:
print("Entire cached dataset is available!")
else:
print(f"Mel uncached: {len(uncached['mel'])}, Frame uncached: {len(uncached['frame'])}")
print(f"Mel uncached: {uncached['mel']}, \nFrame uncached: {uncached['frame']}")
def test_all_batch_sizes_shape():
_, val_loader = create_dataloaders(root_dir=dataset_dir, batch_size=4, use_cached=True)
for i, data in tqdm(enumerate(val_loader)):
if i < len(val_loader) - 5:
continue
print(f"Batch {i+1} shapes:")
print([data[k].shape for k in data])
if __name__ == '__main__':
dataset_dir = "./vis-data-256"
# test_cached_vs_uncached_datasets()
check_if_entire_dataset_cached()
# test_all_batch_sizes_shape()