Skip to content

Commit 77dad06

Browse files
nvpstrnvpstr
committed
[FastPitch] Updating for Ampere
1 parent 9f4678d commit 77dad06

30 files changed

Lines changed: 569 additions & 615 deletions

‎PyTorch/SpeechSynthesis/FastPitch/Dockerfile‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,4 @@
1-
ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:20.03-py3
1+
ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:20.06-py3
22
FROM ${FROM_IMAGE_NAME}
33

44
ADD requirements.txt .

‎PyTorch/SpeechSynthesis/FastPitch/README.md‎

Lines changed: 189 additions & 211 deletions
Large diffs are not rendered by default.

‎PyTorch/SpeechSynthesis/FastPitch/audio/sample_fp32.wav‎

-459 KB
Binary file not shown.

‎PyTorch/SpeechSynthesis/FastPitch/common/log_helper.py‎

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,7 @@
11
import atexit
2+
import glob
23
import os
4+
import re
35
import numpy as np
46

57
from tensorboardX import SummaryWriter
@@ -8,6 +10,19 @@
810
from dllogger import StdOutBackend, JSONStreamBackend, Verbosity
911

1012

13+
def unique_dllogger_fpath(log_fpath):
14+
15+
if not os.path.isfile(log_fpath):
16+
return log_fpath
17+
18+
# Avoid overwriting old logs
19+
saved = sorted([int(re.search('\.(\d+)', f).group(1))
20+
for f in glob.glob(f'{log_fpath}.*')])
21+
22+
log_num = (saved[-1] if saved else 0) + 1
23+
return f'{log_fpath}.{log_num}'
24+
25+
1126
def stdout_step_format(step):
1227
if isinstance(step, str):
1328
return step

‎PyTorch/SpeechSynthesis/FastPitch/common/stft.py‎

Lines changed: 6 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -108,11 +108,12 @@ def inverse(self, magnitude, phase):
108108
recombine_magnitude_phase = torch.cat(
109109
[magnitude*torch.cos(phase), magnitude*torch.sin(phase)], dim=1)
110110

111-
inverse_transform = F.conv_transpose1d(
112-
recombine_magnitude_phase,
113-
Variable(self.inverse_basis, requires_grad=False),
114-
stride=self.hop_length,
115-
padding=0)
111+
with torch.no_grad():
112+
inverse_transform = F.conv_transpose2d(
113+
recombine_magnitude_phase.unsqueeze(-1),
114+
self.inverse_basis.unsqueeze(-1),
115+
stride=self.hop_length,
116+
padding=0).squeeze(-1)
116117

117118
if self.window is not None:
118119
window_sum = window_sumsquare(

‎PyTorch/SpeechSynthesis/FastPitch/export_torchscript.py‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -37,7 +37,7 @@ def parse_args(parser):
3737
help='full path to the generator checkpoint file')
3838
parser.add_argument('-o', '--output', type=str, default="trtis_repo/tacotron/1/model.pt",
3939
help='filename for the Tacotron 2 TorchScript model')
40-
parser.add_argument('--amp-run', action='store_true',
40+
parser.add_argument('--amp', action='store_true',
4141
help='inference with AMP')
4242
return parser
4343

@@ -49,7 +49,7 @@ def main():
4949

5050
model = load_and_setup_model(
5151
args.generator_name, parser, args.generator_checkpoint,
52-
args.amp_run, device='cpu', forward_is_infer=True, polyak=False,
52+
args.amp, device='cpu', forward_is_infer=True, polyak=False,
5353
jitable=True)
5454

5555
torch.jit.save(torch.jit.script(model), args.output)

‎PyTorch/SpeechSynthesis/FastPitch/extract_mels.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -194,7 +194,7 @@ def main():
194194
DLLogger.log(step="PARAMETER", data={k:v})
195195

196196
model = load_and_setup_model(
197-
'Tacotron2', parser, args.tacotron2_checkpoint, amp_run=False,
197+
'Tacotron2', parser, args.tacotron2_checkpoint, amp=False,
198198
device=torch.device('cuda' if args.cuda else 'cpu'),
199199
forward_is_infer=False, ema=False)
200200

‎PyTorch/SpeechSynthesis/FastPitch/img/loss.png‎

47.9 KB
Loading

‎PyTorch/SpeechSynthesis/FastPitch/inference.py‎

Lines changed: 37 additions & 29 deletions
Original file line numberDiff line numberDiff line change
@@ -40,10 +40,10 @@
4040
from torch.nn.utils.rnn import pad_sequence
4141

4242
import dllogger as DLLogger
43-
from apex import amp
4443
from dllogger import StdOutBackend, JSONStreamBackend, Verbosity
4544

4645
from common import utils
46+
from common.log_helper import unique_dllogger_fpath
4747
from common.text import text_to_sequence
4848
from waveglow import model as glow
4949
from waveglow.denoiser import Denoiser
@@ -59,8 +59,8 @@ def parse_args(parser):
5959
help='Full path to the input text (phareses separated by newlines)')
6060
parser.add_argument('-o', '--output', default=None,
6161
help='Output folder to save audio (file per phrase)')
62-
parser.add_argument('--log-file', type=str, default='nvlog.json',
63-
help='Filename for logging')
62+
parser.add_argument('--log-file', type=str, default=None,
63+
help='Path to a DLLogger log file')
6464
parser.add_argument('--cuda', action='store_true',
6565
help='Run inference on a GPU using CUDA')
6666
parser.add_argument('--fastpitch', type=str,
@@ -75,7 +75,7 @@ def parse_args(parser):
7575
help='Sampling rate')
7676
parser.add_argument('--stft-hop-length', type=int, default=256,
7777
help='STFT hop length for estimating audio length from mel size')
78-
parser.add_argument('--amp-run', action='store_true',
78+
parser.add_argument('--amp', action='store_true',
7979
help='Inference with AMP')
8080
parser.add_argument('--batch-size', type=int, default=64)
8181
parser.add_argument('--include-warmup', action='store_true',
@@ -105,7 +105,7 @@ def parse_args(parser):
105105
return parser
106106

107107

108-
def load_and_setup_model(model_name, parser, checkpoint, amp_run, device,
108+
def load_and_setup_model(model_name, parser, checkpoint, amp, device,
109109
unk_args=[], forward_is_infer=False, ema=True,
110110
jitable=False):
111111
model_parser = models.parse_model_args(model_name, parser, add_help=False)
@@ -139,7 +139,7 @@ def load_and_setup_model(model_name, parser, checkpoint, amp_run, device,
139139

140140
if model_name == "WaveGlow":
141141
model = model.remove_weightnorm(model)
142-
if amp_run:
142+
if amp:
143143
model.half()
144144
model.eval()
145145
return model.to(device)
@@ -232,25 +232,28 @@ def main():
232232
Launches text to speech (inference).
233233
Inference is executed on a single GPU.
234234
"""
235+
236+
torch.backends.cudnn.benchmark = True
237+
235238
parser = argparse.ArgumentParser(description='PyTorch FastPitch Inference',
236239
allow_abbrev=False)
237240
parser = parse_args(parser)
238241
args, unk_args = parser.parse_known_args()
239242

240-
DLLogger.init(backends=[JSONStreamBackend(Verbosity.DEFAULT, args.log_file),
241-
StdOutBackend(Verbosity.VERBOSE)])
242-
for k,v in vars(args).items():
243-
DLLogger.log(step="PARAMETER", data={k:v})
244-
DLLogger.log(step="PARAMETER", data={'model_name': 'FastPitch_PyT'})
245-
246243
if args.output is not None:
247244
Path(args.output).mkdir(parents=False, exist_ok=True)
248245

246+
log_fpath = args.log_file or str(Path(args.output, 'nvlog_infer.json'))
247+
log_fpath = unique_dllogger_fpath(log_fpath)
248+
DLLogger.init(backends=[JSONStreamBackend(Verbosity.DEFAULT, log_fpath),
249+
StdOutBackend(Verbosity.VERBOSE)])
250+
[DLLogger.log("PARAMETER", {k:v}) for k,v in vars(args).items()]
251+
249252
device = torch.device('cuda' if args.cuda else 'cpu')
250253

251254
if args.fastpitch is not None:
252255
generator = load_and_setup_model(
253-
'FastPitch', parser, args.fastpitch, args.amp_run, device,
256+
'FastPitch', parser, args.fastpitch, args.amp, device,
254257
unk_args=unk_args, forward_is_infer=True, ema=args.ema,
255258
jitable=args.torchscript)
256259

@@ -263,7 +266,7 @@ def main():
263266
with warnings.catch_warnings():
264267
warnings.simplefilter("ignore")
265268
waveglow = load_and_setup_model(
266-
'WaveGlow', parser, args.waveglow, args.amp_run, device,
269+
'WaveGlow', parser, args.waveglow, args.amp, device,
267270
unk_args=unk_args, forward_is_infer=True, ema=args.ema)
268271
denoiser = Denoiser(waveglow).to(device)
269272
waveglow = getattr(waveglow, 'infer', waveglow)
@@ -305,13 +308,14 @@ def main():
305308
all_frames = 0
306309

307310
reps = args.repeats
308-
log_enabled = reps == 1
311+
log_enabled = True # reps == 1
309312
log = lambda s, d: DLLogger.log(step=s, data=d) if log_enabled else None
310313

311-
for repeat in (tqdm.tqdm(range(reps)) if reps > 1 else range(reps)):
314+
# for repeat in (tqdm.tqdm(range(reps)) if reps > 1 else range(reps)):
315+
for rep in range(reps):
312316
for b in batches:
313317
if generator is None:
314-
log(0, {'Synthesizing from ground truth mels'})
318+
log(rep, {'Synthesizing from ground truth mels'})
315319
mel, mel_lens = b['mel'], b['mel_lens']
316320
else:
317321
with torch.no_grad(), gen_measures:
@@ -321,8 +325,8 @@ def main():
321325
gen_infer_perf = mel.size(0) * mel.size(2) / gen_measures[-1]
322326
all_letters += b['text_lens'].sum().item()
323327
all_frames += mel.size(0) * mel.size(2)
324-
log(0, {"generator_frames_per_sec": gen_infer_perf})
325-
log(0, {"generator_latency": gen_measures[-1]})
328+
log(rep, {"fastpitch_frames_per_sec": gen_infer_perf})
329+
log(rep, {"fastpitch_latency": gen_measures[-1]})
326330

327331
if waveglow is not None:
328332
with torch.no_grad(), waveglow_measures:
@@ -336,8 +340,8 @@ def main():
336340
waveglow_infer_perf = (
337341
audios.size(0) * audios.size(1) / waveglow_measures[-1])
338342

339-
log(0, {"waveglow_samples_per_sec": waveglow_infer_perf})
340-
log(0, {"waveglow_latency": waveglow_measures[-1]})
343+
log(rep, {"waveglow_samples_per_sec": waveglow_infer_perf})
344+
log(rep, {"waveglow_latency": waveglow_measures[-1]})
341345

342346
if args.output is not None and reps == 1:
343347
for i, audio in enumerate(audios):
@@ -354,27 +358,31 @@ def main():
354358
write(audio_path, args.sampling_rate, audio.cpu().numpy())
355359

356360
if generator is not None and waveglow is not None:
357-
log(0, {"latency": (gen_measures[-1] + waveglow_measures[-1])})
361+
log(rep, {"latency": (gen_measures[-1] + waveglow_measures[-1])})
358362

359363
log_enabled = True
360364
if generator is not None:
361365
gm = np.sort(np.asarray(gen_measures))
362-
log('avg', {"generator letters/s": all_letters / gm.sum()})
363-
log('avg', {"generator_frames/s": all_frames / gm.sum()})
364-
log('avg', {"generator_latency": gm.mean()})
365-
log('90%', {"generator_latency": gm.mean() + norm.ppf((1.0 + 0.90) / 2) * gm.std()})
366-
log('95%', {"generator_latency": gm.mean() + norm.ppf((1.0 + 0.95) / 2) * gm.std()})
367-
log('99%', {"generator_latency": gm.mean() + norm.ppf((1.0 + 0.99) / 2) * gm.std()})
366+
rtf = all_samples / (all_utterances * gm.mean() * args.sampling_rate)
367+
log('avg', {"fastpitch letters/s": all_letters / gm.sum()})
368+
log('avg', {"fastpitch_frames/s": all_frames / gm.sum()})
369+
log('avg', {"fastpitch_latency": gm.mean()})
370+
log('avg', {"fastpitch RTF": rtf})
371+
log('90%', {"fastpitch_latency": gm.mean() + norm.ppf((1.0 + 0.90) / 2) * gm.std()})
372+
log('95%', {"fastpitch_latency": gm.mean() + norm.ppf((1.0 + 0.95) / 2) * gm.std()})
373+
log('99%', {"fastpitch_latency": gm.mean() + norm.ppf((1.0 + 0.99) / 2) * gm.std()})
368374
if waveglow is not None:
369375
wm = np.sort(np.asarray(waveglow_measures))
376+
rtf = all_samples / (all_utterances * wm.mean() * args.sampling_rate)
370377
log('avg', {"waveglow_samples/s": all_samples / wm.sum()})
371378
log('avg', {"waveglow_latency": wm.mean()})
379+
log('avg', {"waveglow RTF": rtf})
372380
log('90%', {"waveglow_latency": wm.mean() + norm.ppf((1.0 + 0.90) / 2) * wm.std()})
373381
log('95%', {"waveglow_latency": wm.mean() + norm.ppf((1.0 + 0.95) / 2) * wm.std()})
374382
log('99%', {"waveglow_latency": wm.mean() + norm.ppf((1.0 + 0.99) / 2) * wm.std()})
375383
if generator is not None and waveglow is not None:
376384
m = gm + wm
377-
rtf = all_samples / (len(batches) * all_utterances * m.mean() * args.sampling_rate)
385+
rtf = all_samples / (all_utterances * m.mean() * args.sampling_rate)
378386
log('avg', {"samples/s": all_samples / m.sum()})
379387
log('avg', {"letters/s": all_letters / m.sum()})
380388
log('avg', {"latency": m.mean()})

‎PyTorch/SpeechSynthesis/FastPitch/inference_perf.py‎

Lines changed: 0 additions & 107 deletions
This file was deleted.

0 commit comments

Comments
 (0)