-
Notifications
You must be signed in to change notification settings - Fork 0
/
extract_features_only_test.py
170 lines (153 loc) · 7.48 KB
/
extract_features_only_test.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
import torch
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
from PIL import Image
from open_clip import create_model_from_pretrained, get_tokenizer
from urllib.request import urlopen
import os
import cv2
import pandas as pd
import numpy as np
import librosa
import torch.nn as nn
def read_video(path: str):
video, audio, info = torchvision.io.read_video(path, pts_unit="sec")
video = video.permute(0, 3, 1, 2) / 255
audio = audio.permute(1, 0)
return video, audio, info
base_dir = "/media/data1/FakeAVCeleb/"
# partition = "train"
# max_len = 256
# frame_rate = 25 # Number of frames to extract per second
# real_dir = os.path.join(base_dir, f"{partition}/realOriginal")
# fake_dir = os.path.join(base_dir, f"{partition}/fake")
# file_list = []
# print(real_dir)
# for root, _, files in os.walk(real_dir):
# for file in files:
# if file.endswith(("mp4", "avi", "mov")):
# file_list.append({"path":os.path.join(root, file), "label":1}) # 1 for real
# for root, _, files in os.walk(fake_dir):
# for file in files:
# if file.endswith(("mp4", "avi", "mov")):
# file_list.append({"path":os.path.join(root, file), "label":0}) # 0 for fake
# print(f"Total files are in {partition} are {len(file_list)}")
# Load the model and tokenizer from Hugging Face Hub
device = "cuda"
model, preprocess = create_model_from_pretrained('hf-hub:apple/DFN5B-CLIP-ViT-H-14', device=device)
tokenizer = get_tokenizer('ViT-H-14')
# Helper function to process video frames
def process_video_frames(video):
# frames = video.permute(0, 3, 1, 2) / 255.0 # Normalize frames
frames = video
frame_embeddings = []
for frame in frames:
frame = frame.permute(1,2,0)
frame = frame.numpy().astype('uint8')
# print(frame.shape)
frame_image = Image.fromarray(frame)
processed_frame = preprocess(frame_image).unsqueeze(0)
with torch.no_grad(), torch.cuda.amp.autocast():
processed_frame= processed_frame.to(device)
frame_features = model.encode_image(processed_frame)
frame_features = F.normalize(frame_features, dim=-1)
frame_embeddings.append(frame_features.cpu().numpy())
return frame_embeddings
def fft_to_image(f):
fshift = np.fft.fftshift(f)
magnitude_spectrum = 20 * np.log(np.abs(fshift) + 1) # Log scale, add 1 to avoid log(0)
magnitude_image = np.uint8(magnitude_spectrum / magnitude_spectrum.max() * 255)
return magnitude_image
def diagnose_audio(audio_np):
print("Audio Shape:", audio_np.shape)
print("Max Value in Audio:", np.max(audio_np))
print("Min Value in Audio:", np.min(audio_np))
# Assuming stereo audio
if audio_np.shape[0] == 2:
print("Stereo audio confirmed")
for i, channel in enumerate(['Left', 'Right']):
max_val = np.max(audio_np[i])
min_val = np.min(audio_np[i])
print(f"{channel} channel: Max={max_val}, Min={min_val}")
if max_val == 0 and min_val == 0:
print(f"Warning: {channel} channel is silent!")
def process_video_audio_fft(video, audio, output_dir):
# Define the dimensions for the output video, e.g., 224x224 for simplicity
out_height, out_width = 224, 224
fourcc = cv2.VideoWriter_fourcc(*'mp4v') # Define the codec
out = cv2.VideoWriter(f'{output_dir}_video_fft.mp4', fourcc, 25.0, (out_width, out_height))
for i in range(len(video)):
frame = video[i].numpy().transpose(1, 2, 0) # Convert to HxWxC
gray_frame = cv2.cvtColor((frame * 255).astype(np.uint8), cv2.COLOR_RGB2GRAY)
f = np.fft.fft2(gray_frame)
magnitude_image = fft_to_image(f)
magnitude_image = cv2.resize(magnitude_image, (out_width, out_height))
color_mapped_image = cv2.applyColorMap(magnitude_image, cv2.COLORMAP_JET)
out.write(color_mapped_image)
# Audio FFT processing
audio_np = audio.numpy()
# diagnose_audio(audio_np)
out = cv2.VideoWriter(f'{output_dir}_audio_fft.mp4', fourcc, 25.0, (out_width, out_height))
window_size = 1024 # Size of the window to apply FFT on
hop_length = window_size // 2 # 50% overlap
# We assume audio has two channels; process each channel separately
if audio_np.shape[0] == 1: # Stereo audio
audio_channel = audio_np[0] # Extract the single audio channel
for start in range(0, len(audio_channel) - window_size, hop_length):
# Compute Mel Spectrogram
S = librosa.feature.melspectrogram(y=audio_channel[start:start + window_size], sr=44100, n_fft=window_size, hop_length=hop_length, n_mels=128)
S_dB = librosa.power_to_db(S, ref=np.max)
# Compute MFCC
mfccs = librosa.feature.mfcc(S=S_dB, n_mfcc=13)
# Optionally, apply FFT to MFCCs
mfccs_fft = np.fft.fft(mfccs, axis=1)
mfccs_fft_shift = np.fft.fftshift(mfccs_fft, axes=1)
magnitude_mfcc_fft = 20 * np.log10(np.abs(mfccs_fft_shift) + 1e-6) # Log scale
# Map the magnitude to an image
magnitude_image = np.uint8(255 * (magnitude_mfcc_fft - magnitude_mfcc_fft.min()) / (magnitude_mfcc_fft.max() - magnitude_mfcc_fft.min()))
color_mapped_image = cv2.applyColorMap(magnitude_image, cv2.COLORMAP_JET)
color_mapped_image = cv2.resize(color_mapped_image, (out_width, out_height))
out.write(color_mapped_image)
out.release() # Release the video writer
# Example text and video path
def save_features( output_dir,row):
labels_list = [f"These video and audio frames are of {row['race']} {row['gender']}, which is {row['method']} and it has {row['type']}, your job is to learn the patterns of the audio and video. "]
# Tokenize tex
print(labels_list[0])
text = tokenizer(labels_list, context_length=model.context_length)
# Process the video and generate frame embeddings
# frame_embeddings = process_video_frames(video)
# Convert list of arrays into a single tensor
# frame_embeddings_tensor = torch.tensor(np.vstack(frame_embeddings))
with torch.no_grad(), torch.cuda.amp.autocast():
text = text.to(device)
text_features = model.encode_text(text)
text_features = F.normalize(text_features, dim=-1)
text_features = F.normalize(text_features, dim=-1)
# frame_embeddings_tensor = frame_embeddings_tensor.cpu()
text_features = text_features.cpu()
# np.save(f'{output_dir}_video_embds.npy', frame_embeddings_tensor)
np.save(f'{output_dir}_text_embds_row.npy', text_features)
# video, _, _ = torchvision.io.read_video(video_path, pts_unit='sec')
def save_ffts(video, audio, output_dir):
audio = audio.permute(1,0)
process_video_audio_fft(video, audio, output_dir)
def save(video_path,output_dir, row):
# video, audio, info = read_video(video_path)
# print(video.shape)
# save_ffts(video, audio,output_dir)
save_features(output_dir,row)
def main():
base_dir = "/home/PromptCLIP/FakeAVCeleb_v1.2/"
csv_path = f'{base_dir}meta_data.csv' # Change this to the actual path of your CSV file
data = pd.read_csv(csv_path)
# data = data[1348*15:1348*16]
for index, row in data.iterrows():
path = os.path.join(base_dir,row['Unnamed: 9'].replace("FakeAVCeleb/", ""))
video_path = os.path.join(path, row["path"])
print(video_path)
output_dir = os.path.join(path, row["path"].replace(".mp4", ""))
save(video_path, output_dir,row)
if __name__ == "__main__":
main()