-
Notifications
You must be signed in to change notification settings - Fork 21
/
generate_for_reflow.py
191 lines (167 loc) · 7.21 KB
/
generate_for_reflow.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
import os
import torch
from kaldiio import WriteHelper
from torch.utils.data import DataLoader
from tqdm import tqdm
import tempfile
import torch.multiprocessing as mp
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
import tools
def run(rank, n_gpus, hps, args, ckpt, feats_dir, temp_dir):
logger = tools.get_logger(hps.model_dir, f"inference.{rank}.log") # NOTE: cannot delete this line.
device = torch.device('cpu' if not torch.cuda.is_available() else f"cuda:{rank}")
torch.manual_seed(hps.train.seed) # NOTE: control seed
setattr(hps.data, "train_utts" if args.dataset == "train" else "val_utts", f"{temp_dir}/{rank}.txt")
train_dataset, collate_fn, model = tools.get_correct_class(hps)
val_dataset, _, _ = tools.get_correct_class(hps, train=False)
batch_collate = collate_fn
train_loader = DataLoader(
dataset=train_dataset,
batch_size=1,
collate_fn=batch_collate,
drop_last=True,
num_workers=4,
shuffle=False,
)
val_loader = DataLoader(
dataset=val_dataset,
batch_size=1,
collate_fn=batch_collate,
drop_last=True,
num_workers=4,
shuffle=False,
)
model = model(**hps.model).to(device)
tools.load_checkpoint(ckpt, model, None)
print(f"Loaded checkpoint from {ckpt}")
model.eval()
print(f"Number of parameters: {model.nparams}")
print(f"Number of encoder parameters: {model.encoder.nparams}")
print(f"Number of decoder parameters: {model.decoder.nparams}")
if args.dataset == "val":
which_loader = val_loader # NOTE: specify the dataset: train or val?
which_set = val_dataset
else:
which_loader = train_loader
which_set = train_dataset
met = False
with torch.no_grad():
with WriteHelper(
f"ark,scp:{os.getcwd()}/{feats_dir}/feats.{rank}.ark,{feats_dir}/feats.{rank}.scp"
) as feats, WriteHelper(
f"ark,scp:{os.getcwd()}/{feats_dir}/noise.{rank}.ark,{feats_dir}/noise.{rank}.scp"
) as noise_feats, open(
f"{feats_dir}/duration.{rank}", "w"
) as duration_writer:
# NOTE: its necessary to add "os.getcwd" here.
for batch_idx, batch in tqdm(enumerate(which_loader), total=len(which_loader)):
utts = batch["utt"]
# ============== Loop Controlling block ============
if met:
break
if args.specify_utt_name is not None:
if not utts[0] == args.specify_utt_name:
continue
else:
met = True
elif batch_idx >= args.max_utt_num:
break
# ==================================================
x, x_lengths = batch["text_padded"].to(device), batch["input_lengths"].to(
device
)
dur = batch["dur_padded"].to(device) if args.gt_dur else None
# ================== Decode ======================
if hps.xvector:
if args.use_control_spk:
xvector = which_set.spk2xvector[args.control_spk_name]
spk = torch.FloatTensor(xvector).squeeze().unsqueeze(0).to(device)
else:
spk = batch["xvector"].to(device)
else:
if args.use_control_spk:
spk = torch.LongTensor([args.control_spk_id]).to(device)
else:
spk = batch["spk_ids"].to(device)
y_enc, y_dec, attn, z, pred_dur = model.inference(
x,
x_lengths,
n_timesteps=args.timesteps,
temperature=1.5,
spk=spk,
length_scale=1.0,
solver=args.solver,
gt_dur=dur,
)
# =================================================
if args.use_control_spk:
save_utt_name = f"[spk_{args.control_spk_name if hps.xvector else args.control_spk_id}]{utts[0]}"
else:
save_utt_name = f"{utts[0]}"
feats(
save_utt_name, y_dec.squeeze().cpu().numpy().T
) # save to ark and scp, mel: (L, 80)
noise_feats(save_utt_name, z.squeeze().cpu().numpy().T)
dur_seq = pred_dur.long().squeeze().cpu().numpy().tolist()
if isinstance(dur_seq, int):
dur_seq = [dur_seq]
dur_seq = " ".join(list(map(str, dur_seq)))
duration_writer.write(f"{save_utt_name} {dur_seq}\n")
if __name__ == '__main__':
"""Assume Single Node Multi GPUs Training Only"""
assert torch.cuda.is_available(), "CPU training is not allowed."
n_gpus = torch.cuda.device_count()
print(f"============> using {n_gpus} GPUS")
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "80000"
hps, args = tools.get_hparams_decode()
ckpt = tools.latest_checkpoint_path(hps.model_dir, "grad_*.pt" if not args.EMA else "EMA_grad_*.pt")
if args.use_control_spk:
feats_dir = f"synthetic_wav/{args.model}/tts_other_spk"
else:
feats_dir = f"synthetic_wav/{args.model}/generate_for_reflow/{args.dataset}"
if not os.path.exists(feats_dir):
os.makedirs(feats_dir)
with tempfile.TemporaryDirectory() as temp_dir:
print("Temporary dir is", temp_dir)
# split hps.data.{dataset}_utts into multiple copies
which_file_to_split = hps.data.train_utts if args.dataset == "train" else hps.data.val_utts
with open(which_file_to_split, 'r') as fr:
lines = fr.readlines()
total_lines = len(lines)
lines_per_copy = total_lines // n_gpus
remaining_lines = total_lines % n_gpus
for i in range(n_gpus):
output_file = f"{temp_dir}/{i}.txt"
with open(output_file, 'w') as fw:
start = i * lines_per_copy + min(i, remaining_lines)
end = start + lines_per_copy + (1 if i < remaining_lines else 0)
fw.writelines(lines[start:end])
mp.spawn(
run,
nprocs=n_gpus,
args=(
n_gpus,
hps,
args,
ckpt,
feats_dir,
temp_dir
),
)
with open(f"{feats_dir}/feats.scp", 'w') as output:
for i in range(n_gpus):
input_file = f"{feats_dir}/feats.{i}.scp"
with open(input_file, "r") as fr:
output.write(fr.read())
with open(f"{feats_dir}/noise.scp", 'w') as output:
for i in range(n_gpus):
input_file = f"{feats_dir}/noise.{i}.scp"
with open(input_file, "r") as fr:
output.write(fr.read())
with open(f"{feats_dir}/duration", 'w') as output:
for i in range(n_gpus):
input_file = f"{feats_dir}/duration.{i}"
with open(input_file, "r") as fr:
output.write(fr.read())