-
Notifications
You must be signed in to change notification settings - Fork 0
/
datasets_lavdf.py
275 lines (240 loc) · 12.3 KB
/
datasets_lavdf.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
import numpy as np
import os
from torch.utils.data import Dataset
import librosa
import torch
import torchaudio
import random
import cv2 # OpenCV for video processing
import torchvision
import traceback
import torch.nn.functional as F
from decord import VideoReader
from decord import cpu, gpu
import torchvision.transforms as transforms
import av
import pandas as pd
def pad_random(x: np.ndarray, max_len: int = 64000):
x_len = x.shape[0]
if x_len > max_len:
stt = np.random.randint(x_len - max_len)
return x[stt:stt + max_len]
num_repeats = int(max_len / x_len) + 1
padded_x = np.tile(x, (num_repeats))
return pad_random(padded_x, max_len)
def add_noise(features, noise_level=0.005):
noise = torch.randn(features.size()) * noise_level
return features + noise
def mask_features(features, mask_prob=0.1, mask_value=0.0):
mask = torch.rand(features.size()) < mask_prob
features = features.masked_fill(mask, mask_value)
return features
def time_shift(features, shift_limit=0.2):
shift = int(random.uniform(-shift_limit, shift_limit) * features.size(1))
return torch.roll(features, shifts=shift, dims=1)
def pitch_shift(features, sr=16000, n_steps=2):
shifted = librosa.effects.pitch_shift(features.numpy(), sr=sr, n_steps=n_steps)
return torch.tensor(shifted)
def speed_change(features, rate=1.1):
changed = librosa.effects.time_stretch(features.numpy(), rate=rate)
return torch.tensor(changed)
def apply_augmentation(features):
if random.random() < 0.5:
features = add_noise(features)
if random.random() < 0.5:
features = mask_features(features)
if random.random() < 0.5:
features = time_shift(features)
if random.random() < 0.5:
features = pitch_shift(features)
if random.random() < 0.5:
features = speed_change(features)
return features
def read_video_with_audio(path):
container = av.open(path)
video_frames = []
audio_frames = []
stream = container.streams.video[0]
for frame in container.decode(stream):
img = frame.to_image() # Convert frame to PIL Image
img_tensor = transforms.functional.to_tensor(img)
video_frames.append(img_tensor)
if len(video_frames) >= 40:
break
if container.streams.audio:
audio_stream = container.streams.audio[0]
for frame in container.decode(audio_stream):
audio_data = frame.to_ndarray()
audio_frames.append(audio_data)
video_tensor = torch.stack(video_frames) if video_frames else None
audio_tensor = np.concatenate(audio_frames, axis=1) if audio_frames else None
# del
return video_tensor, torch.tensor(audio_tensor, dtype=torch.float32) if audio_tensor is not None else None
def read_video_decord(path):
vr = VideoReader(path, ctx=cpu(0)) # Use gpu(0) for GPU
video = vr.get_batch(range(40)).asnumpy() # Get first 40 frames
video = torch.tensor(video).permute(0, 3, 1, 2).float() / 255.0 # Convert to torch tensor and normalize
return video
def extract_audio_with_ffmpeg(path, output_format='wav', sample_rate=16000):
# Temporary output file
temp_audio = 'temp_audio.wav'
# Command to extract audio
command = ['ffmpeg', '-i', path, '-ar', str(sample_rate), '-ac', '1', temp_audio, '-y']
subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
# Load the processed audio file
audio, sr = librosa.load(temp_audio, sr=sample_rate)
audio_tensor = torch.tensor(audio).float()
# Delete the temporary audio file after reading
os.remove(temp_audio)
return audio_tensor
def read_video(path: str):
video, audio, info = torchvision.io.read_video(path, pts_unit="sec")
video = video[:128]
video = video.permute(0, 3, 1, 2) / 255
audio = audio.permute(1, 0)
return video, audio, info
class SAMMD_dataset(Dataset):
"""
Dataset class for the SAMMD2024 dataset.
"""
def __init__(self, base_dir, partition="train", max_len=64000, frame_rate=1):
assert partition in ["train", "dev", "test"], "Invalid partition. Must be one of ['train', 'dev', 'test']"
self.base_dir = base_dir
self.partition = partition
self.max_len = max_len
self.frame_rate = frame_rate # Number of frames to extract per second
self.real_dir = os.path.join(base_dir, f"{partition}/realOriginal")
self.fake_dir = os.path.join(base_dir, f"{partition}/fake")
self.file_list = []
# Limit for samples from each category
limit_per_category = 64
# base_dir = "/home/shahid/PromptCLIP/FakeAVCeleb_v1.2/"
csv_path = f'{self.base_dir}/metadata.json' # Change this to the actual path of your CSV file
data = pd.read_json(csv_path)
# Process real videos
count_real = 0
print(self.real_dir)
for index, row in data.iterrows():
path = self.base_dir + "/" + row["file"]
audio_fft = path.replace(".mp4", "_audio_fft.mp4")
video_fft = path.replace(".mp4", "_video_fft.mp4")
text_embeddings = path.replace(".mp4", "_text_embds_clip_vit.npy")
video_embeddings = path.replace(".mp4","_video_embds_clip_vit.npy")
# label = 0 if row["method"]=="real" else 1
label = 0 if row["modify_video"] == False and row["modify_audio"] == False else 1
# if label==0:
# print("True detected")
# if label == 0:
# print(label)
if partition == row["split"]:
if os.path.exists(video_fft) and os.path.exists(path) and os.path.exists(video_embeddings):
self.file_list.append({
"video_path":path,
"audio_fft":audio_fft,
"video_fft":video_fft,
"text_embeddings":text_embeddings,
"video_embeddings":video_embeddings,
"label":label}) # 1 for real
# for root, _, files in os.walk(self.real_dir):
# for file in files:
# if file.endswith(("mp4", "avi", "mov")):
# self.file_list.append({"path":os.path.join(root, file), "label":1}) # 1 for real
# count_real += 1
# # if count_real >= limit_per_category:
# # break
# # if count_real >= limit_per_category:
# # break
# count_fake = 0
# for root, _, files in os.walk(self.fake_dir):
# for file in files:
# if file.endswith(("mp4", "avi", "mov")):
# self.file_list.append({"path":os.path.join(root, file), "label":0}) # 0 for fake
# count_fake += 1
# if count_fake >= limit_per_category:
# break
# if count_fake >= limit_per_category:
# break
print(f"Total files are in {partition} are {len(self.file_list)}")
def __len__(self):
return len(self.file_list)
def __getitem__(self, index):
try:
data = self.file_list[index]
file_path = data["video_path"]
audio_fft = data["audio_fft"]
video_fft = data["video_fft"]
text_embeddings = data["text_embeddings"]
video_embeddings = data["video_embeddings"]
label = data["label"]
# print(file_path)
# if file_path == "/media/data2/FakeAVCeleb_v1.2/FakeAVCeleb/train/real/RealVideo-RealAudio-id00350-00015.mp4":
# self.__getitem__(index+1)
# Extract audio
# audio_waveform, sample_rate = torchaudio.load(file_path)
# audio_waveform = audio_waveform.mean(dim=0) # Convert to mono
# audio_waveform = librosa.resample(audio_waveform.numpy(), sample_rate, 16000)
# audio_waveform = pad_random(audio_waveform, self.max_len)
# audio_waveform = torch.tensor(audio_waveform)
# if self.partition == "train":
# audio_waveform = apply_augmentation(audio_waveform)
# print(file_path)
# edia/data2/FakeAVCeleb_v1.2/FakeAVCeleb/train/real/RealVideo-RealAudio-id00857-00347.mp4
video, audio, info = read_video(file_path)
# audio_fft,_,_ = read_video(audio_fft)
video_fft,_,_ = read_video(video_fft)
text_embeddings = np.load(text_embeddings)
video_embeddings = np.load(video_embeddings)
audio = audio.permute(1,0)
audio = audio.mean(dim=0) # Convert to mono
audio = pad_random(audio, self.max_len)
if isinstance(audio, np.ndarray):
audio = torch.from_numpy(audio)
video_embeddings = torch.tensor(video_embeddings)
# print(video_embeddings.shape)
# padding_size=(pad_width,pad_height)
# video = F.pad(video, (0, 0, 0, 0, 0, 0, *padding_size))
# if self.partition == "train":
# audio_waveform = apply_augmentation(audio_waveform)
# Adjust video to have exactly 40 frames
# num_frames = video.shape[0]
# if num_frames < 1024:
# pad_size = 40 - num_frames # Calculate how many frames are missing
# padding = torch.zeros(pad_size, *video.shape[1:], device=video.device) # Create a padding tensor
# video = torch.cat([video, padding], dim=0) # Concatenate the video and padding tensor along the time dimension
# print(video_embeddings.shape)
# video = video[:40]
# video_embeddings = video_embeddings[:40]
# video_fft = video_fft[:40]
# audio_fft = audio_fft[:40]
# video = video[:40]# Assuming video is your tensor with the shape (40, 3, 244, 244)
video_shape = video.shape[0] # current number of frames, 40 in your case
target_frames = 256 # target number of frames
# print(video.shape)
video = torch.nn.functional.interpolate(video, size=(244, 244), mode='bilinear', align_corners=False)
# audio_fft = torch.nn.functional.interpolate(audio_fft, size=(244, 244), mode='bilinear', align_corners=False)
video_fft = torch.nn.functional.interpolate(video_fft, size=(244, 244), mode='bilinear', align_corners=False)
# assert audio_fft.shape[0] > 0, "audio_fft has zero frames after interpolation, which is unexpected."
audio_fft = torch.randn(0,123)
# if video_fft.shape[0]<40 and video_embeddings.shape[0]<40:
# print(video_fft.shape[0],audio_fft.shape[0], video_embeddings.shape[0])
# return self.__getitem__((index + 1) % len(self.file_list))
if video_shape < target_frames:
# Number of frames to add
pad_size = target_frames - video_shape
# Create a padding tensor of shape (pad_size, 3, 244, 244)
padding = torch.zeros((pad_size, *video.shape[1:]), dtype=video.dtype)
padding_embeddings = torch.zeros((pad_size, *video_embeddings.shape[1:]), dtype=video.dtype)
# Concatenate the original video with the padding tensor
video_fft = torch.cat((video_fft, padding), dim=0) # Append at the end
video = torch.cat((video, padding), dim=0) # Append at the end
video_embeddings = torch.cat((video_embeddings, padding_embeddings), dim=0)
else:
video_fft = video_fft[:target_frames] # or just cut the excess frames
video = video[:target_frames] # or just cut the excess frames
video_embeddings = video_embeddings[:target_frames] # or just cut the excess frames
# print(video.shape,video_fft.shape, audio_fft.shape)
return audio, video, label, audio_fft, video_fft, text_embeddings, video_embeddings, os.path.basename(file_path)
except Exception as e:
# traceback.print_exc()
print(f"Error loading {file_path}: {e}")
return self.__getitem__((index + 1) % len(self.file_list))