-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinfer.py
146 lines (120 loc) · 7.37 KB
/
infer.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
import argparse
import os
import torch
import numpy as np
from yacs.config import CfgNode as CN
from config import _C as config
from data_utils import RMS, VideoAnnotation
from util import load_config, load_models, prepare_dataloaders, get_audio, save_audio, save_video_with_audio, interpolate_rms_for_rms2sound, set_seed
@torch.no_grad
def infer(epoch:int=500, video2rms_ckpt_dir:str='', rms2sound_ckpt_dir:str = '',
prompt_type:str='audio', config:CN=None, output_dir:str='./infer') -> None:
if video2rms_ckpt_dir == '' or rms2sound_ckpt_dir == '':
raise ValueError("ckpt_dir is empty")
if config is None:
raise ValueError("config is None")
if prompt_type not in ['audio', 'text']:
raise ValueError("prompt_type must be 'audio' or 'text'")
print(f"Inference with epoch {epoch}...")
print(f'Setting seed: {config.train.seed}')
set_seed(config.train.seed)
# load ckpt and prepare model
print('Loading model...')
video2rms_model, audio_ldm_controlnet = load_models(epoch, video2rms_ckpt_dir, rms2sound_ckpt_dir, config,
torch.device('cuda:0' if torch.cuda.is_available() else 'cpu'))
# prepare dataloader
print('Preparing dataset...')
test_loader = prepare_dataloaders(config.data, batch_size=16, train=False, test=True)
# evaluate
print('Inference...')
os.makedirs(os.path.join(output_dir, f'{prompt_type}_prompt', 'audio'), exist_ok=True)
os.makedirs(os.path.join(output_dir, f'{prompt_type}_prompt', 'video'), exist_ok=True)
video2rms_model.eval()
assert config.data.rms_discretize, 'RMS must be discretized'
mu_bins:torch.tensor = RMS.get_mu_bins(config.data.rms_mu, config.data.rms_num_bins, config.data.rms_min)
with torch.no_grad():
for i, batch in enumerate(test_loader):
video2rms_model.parse_batch(batch)
video2rms_model.forward()
for j in range(len(video2rms_model.pred_rms)):
pred_rms = video2rms_model.pred_rms[j].detach().cpu().numpy()
pred_rms_undiscretized:torch.tensor =RMS.undiscretize_rms(torch.from_numpy(pred_rms.argmax(axis=0)),
mu_bins, ignore_min=True)
pred_rms_undiscretized = pred_rms_undiscretized.detach().cpu().unsqueeze(0)
pred_rms_undiscretized = interpolate_rms_for_rms2sound(pred_rms_undiscretized,
audio_len=config.data.audio_samples ,
sr=config.data.audio_sample_rate,
frame_len=1024,
hop_len=160)
if prompt_type == 'audio':
gt_audio:np.ndarray = get_audio(os.path.join(config.data.audio_src_dir.replace('*', video2rms_model.video_class[j]),
video2rms_model.video_name[j] + '.wav'),
sr=config.data.audio_sample_rate)
gt_audio = torch.from_numpy(gt_audio).unsqueeze(0)
generated_audio:np.ndarray = audio_ldm_controlnet.generate(
waveform=gt_audio,
rms=pred_rms_undiscretized
)
else:
videoname, index = video2rms_model.video_name[j].split('_')
index = int(index)
text_prompt:str = VideoAnnotation.get_text_prompt(
annot_dir=config.data.annotation_dir,
videoname=videoname,
index=index,
length=config.data.audio_samples
)
generated_audio:np.ndarray = audio_ldm_controlnet.generate(
text_prompt=text_prompt,
rms=pred_rms_undiscretized
)
save_audio(audio=generated_audio,
output_path=os.path.join(output_dir, f'{prompt_type}_prompt', 'audio',
video2rms_model.video_name[j] + '.wav'),
sr=config.data.audio_sample_rate)
save_video_with_audio(video_path=os.path.join(config.data.video_src_dir.replace('*', video2rms_model.video_class[j]),
video2rms_model.video_name[j] + '.mp4'),
audio=generated_audio,
output_path=os.path.join(output_dir, f'{prompt_type}_prompt', 'video',
video2rms_model.video_name[j] + '.mp4'),
sr=config.data.audio_sample_rate)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-v', '--video2rms_ckpt_dir', type=str,
default='./ckpt/video-foley-model',
help='directory for model checkpoint')
parser.add_argument('-r', '--rms2sound_ckpt_dir', type=str,
default='./ckpt/video-foley-model',
help='directory for model checkpoint')
parser.add_argument('-e', '--epoch', type=int, default=500,
help='number of epochs of Video2RMS model')
parser.add_argument('-d', '--data_dir', type=str, default=None,
help='input directory for dataset')
parser.add_argument('-o', '--output_dir', type=str,
default='./video2sound/Video-Foley',
help='output directory to save generated results')
parser.add_argument('-p', '--prompt_type', type=str,
default='text',
choices=['text', 'audio'],
help='prompt type for audio generation')
args = parser.parse_args()
config = load_config(os.path.join(args.video2rms_ckpt_dir, 'opts.yml'))
config_data_dir = config.data.rgb_feature_dirs[0].split('/features/')[0]
if args.data_dir is None:
args.data_dir = config_data_dir
else:
if not (os.path.exists(os.path.join(args.data_dir, 'features'))
and os.path.exists(os.path.join(args.data_dir, 'features'))):
raise FileNotFoundError(f"Data directory {args.data_dir} not found")
if not os.path.abspath(args.data_dir) == os.path.abspath(config_data_dir):
print(f"Warning: data directory {args.data_dir} is different from data directory in config {config_data_dir}")
config.data.audio_src_dir = os.path.join(args.data_dir, 'features/*/audio_10s_16000hz_muted')
config.data.video_src_dir = os.path.join(args.data_dir, 'features/*/videos_10s_30fps')
config.data.annotation_dir = '/media/daftpunk4/dataset/GreatestHits/vis-data' # os.path.join(args.data_dir, 'vis-data')
config.freeze()
infer(epoch=args.epoch,
video2rms_ckpt_dir=args.video2rms_ckpt_dir,
rms2sound_ckpt_dir=args.rms2sound_ckpt_dir,
prompt_type=args.prompt_type,
config=config,
output_dir=args.output_dir)