testadmin
/
Parakeet
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #129 from iclementine/speedyspeech 

add speedyspeech model and example with baker dataset.
This commit is contained in:
Hui Zhang 2021-07-19 03:32:51 -05:00 committed by GitHub
parent 124dedbd7b 4ba8e7e342
commit 25788ab2ca
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
36 changed files with 2125 additions and 59 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
examples/parallelwave_gan/baker/conf/default.yaml
View File

@ -6,18 +6,18 @@
########################################################### ###########################################################
# FEATURE EXTRACTION SETTING # # FEATURE EXTRACTION SETTING #
########################################################### ###########################################################
sr: 24000 # Sampling rate. sr: 24000 # Sampling rate.
n_fft: 2048 # FFT size. n_fft: 2048 # FFT size (in samples).
hop_length: 300 # Hop size. hop_length: 300 # Hop size (in samples).
win_length: 1200 # Window length. win_length: 1200 # Window length (in samples).
# If set to null, it will be the same as fft_size. # If set to null, it will be the same as fft_size.
window: "hann" # Window function. window: "hann" # Window function.
n_mels: 80 # Number of mel basis. n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
# global_gain_scale: 1.0 # Will be multiplied to all of waveform. # global_gain_scale: 1.0 # Will be multiplied to all of waveform.
trim_silence: false # Whether to trim the start and end of silence. trim_silence: false # Whether to trim the start and end of silence.
top_db: 60 # Need to tune carefully if the recording is not good. top_db: 60 # Need to tune carefully if the recording is not good.
trim_frame_length: 2048 # Frame size in trimming.(in samples) trim_frame_length: 2048 # Frame size in trimming.(in samples)
trim_hop_length: 512 # Hop size in trimming.(in samples) trim_hop_length: 512 # Hop size in trimming.(in samples)

6
examples/parallelwave_gan/baker/preprocess.py
View File

@ -202,14 +202,12 @@ def process_sentences(config,
def main(): def main():
# parse config and args # parse config and args
parser = argparse.ArgumentParser( parser = argparse.ArgumentParser(
description="Preprocess audio and then extract features (See detail in parallel_wavegan/bin/preprocess.py)." description="Preprocess audio and then extract features .")
)
parser.add_argument( parser.add_argument(
"--rootdir", "--rootdir",
default=None, default=None,
type=str, type=str,
help="directory including wav files. you need to specify either scp or rootdir." help="directory to baker dataset.")
)
parser.add_argument( parser.add_argument(
"--dumpdir", "--dumpdir",
type=str, type=str,

6
examples/parallelwave_gan/baker/preprocess.sh Normal file
View File

@ -0,0 +1,6 @@
python preprocess.py --rootdir=~/datasets/BZNSYP/ --dumpdir=dump --num_cpu=20
python compute_statistics.py --metadata=dump/train/raw/metadata.jsonl --field-name="feats" --dumpdir=dump/train
python normalize.py --metadata=dump/train/raw/metadata.jsonl --dumpdir=dump/train/norm --stats=dump/train/stats.npy
python normalize.py --metadata=dump/dev/raw/metadata.jsonl --dumpdir=dump/dev/norm --stats=dump/train/stats.npy
python normalize.py --metadata=dump/test/raw/metadata.jsonl --dumpdir=dump/test/norm --stats=dump/train/stats.npy

9
examples/parallelwave_gan/baker/run.sh Normal file
View File

@ -0,0 +1,9 @@
FLAGS_cudnn_exhaustive_search=true \
FLAGS_conv_workspace_size_limit=4000 \
python train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=conf/default.yaml \
--output-dir=exp/default \
--nprocs=1

43
examples/speedyspeech/baker/batch_fn.py Normal file
View File

@ -0,0 +1,43 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
from parakeet.data.batch import batch_sequences
def collate_baker_examples(examples):
# fields = ["phones", "tones", "num_phones", "num_frames", "feats", "durations"]
phones = [np.array(item["phones"], dtype=np.int64) for item in examples]
tones = [np.array(item["tones"], dtype=np.int64) for item in examples]
feats = [np.array(item["feats"], dtype=np.float32) for item in examples]
durations = [
np.array(
item["durations"], dtype=np.int64) for item in examples
]
num_phones = np.array([item["num_phones"] for item in examples])
num_frames = np.array([item["num_frames"] for item in examples])
phones = batch_sequences(phones)
tones = batch_sequences(tones)
feats = batch_sequences(feats)
durations = batch_sequences(durations)
batch = {
"phones": phones,
"tones": tones,
"num_phones": num_phones,
"num_frames": num_frames,
"feats": feats,
"durations": durations,
}
return batch

109
examples/speedyspeech/baker/compute_statistics.py Normal file
View File

@ -0,0 +1,109 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Calculate statistics of feature files."""
import argparse
import logging
import os
from pathlib import Path
import numpy as np
import yaml
import json
import jsonlines
from sklearn.preprocessing import StandardScaler
from tqdm import tqdm
from parakeet.datasets.data_table import DataTable
from parakeet.utils.h5_utils import read_hdf5
from parakeet.utils.h5_utils import write_hdf5
from config import get_cfg_default
def main():
"""Run preprocessing process."""
parser = argparse.ArgumentParser(
description="Compute mean and variance of dumped raw features.")
parser.add_argument(
"--metadata", type=str, help="json file with id and file paths ")
parser.add_argument(
"--field-name",
type=str,
help="name of the field to compute statistics for.")
parser.add_argument(
"--config", type=str, help="yaml format configuration file.")
parser.add_argument(
"--output",
type=str,
help="path to save statistics. if not provided, "
"stats will be saved in the above root directory with name stats.npy")
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)")
args = parser.parse_args()
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
)
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
)
else:
logging.basicConfig(
level=logging.WARN,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
)
logging.warning('Skip DEBUG/INFO messages')
config = get_cfg_default()
# load config
if args.config:
config.merge_from_file(args.config)
# check directory existence
if args.output is None:
args.output = Path(args.metadata).parent.with_name("stats.npy")
else:
args.output = Path(args.output)
args.output.parent.mkdir(parents=True, exist_ok=True)
with jsonlines.open(args.metadata, 'r') as reader:
metadata = list(reader)
dataset = DataTable(
metadata,
fields=[args.field_name],
converters={args.field_name: np.load}, )
logging.info(f"The number of files = {len(dataset)}.")
# calculate statistics
scaler = StandardScaler()
for datum in tqdm(dataset):
# StandardScalar supports (*, num_features) by default
scaler.partial_fit(datum[args.field_name])
stats = np.stack([scaler.mean_, scaler.scale_], axis=0)
np.save(str(args.output), stats.astype(np.float32), allow_pickle=False)
if __name__ == "__main__":
main()

60
examples/speedyspeech/baker/conf/default.yaml Normal file
View File

@ -0,0 +1,60 @@
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
sr: 24000 # Sampling rate.
n_fft: 2048 # FFT size.
hop_length: 300 # Hop size.
win_length: 1200 # Window length.
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation.
fmax: 7600 # Maximum frequency in mel basis calculation.
# global_gain_scale: 1.0 # Will be multiplied to all of waveform.
trim_silence: false # Whether to trim the start and end of silence.
top_db: 60 # Need to tune carefully if the recording is not good.
trim_frame_length: 2048 # Frame size in trimming.(in samples)
trim_hop_length: 512 # Hop size in trimming.(in samples)
###########################################################
# DATA SETTING #
###########################################################
batch_size: 32
num_workers: 4
###########################################################
# MODEL SETTING #
###########################################################
model:
vocab_size: 101 # 99 + 2
tone_size: 8 # 6 + 2
encoder_hidden_size: 128
encoder_kernel_size: 3
encoder_dilations: [1, 3, 9, 27, 1, 3, 9, 27, 1, 1]
duration_predictor_hidden_size: 128
decoder_hidden_size: 128
decoder_output_size: 80
decoder_kernel_size: 3
decoder_dilations: [1, 3, 9, 27, 1, 3, 9, 27, 1, 3, 9, 27, 1, 3, 9, 27, 1, 1]
###########################################################
# OPTIMIZER SETTING #
###########################################################
###########################################################
# TRAINING SETTING #
###########################################################
max_epoch: 300
num_snapshots: 5
###########################################################
# OTHER SETTING #
###########################################################
seed: 10086

25
examples/speedyspeech/baker/config.py Normal file
View File

@ -0,0 +1,25 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import yaml
from yacs.config import CfgNode as Configuration
with open("conf/default.yaml", 'rt') as f:
_C = yaml.safe_load(f)
_C = Configuration(_C)
def get_cfg_default():
config = _C.clone()
return config

92
examples/speedyspeech/baker/frontend.py Normal file
View File

@ -0,0 +1,92 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import re
import numpy as np
import paddle
import pypinyin
from pypinyin import lazy_pinyin, Style
import jieba
import phkit
phkit.initialize()
from parakeet.frontend.vocab import Vocab
with open("phones.txt", 'rt') as f:
phones = [line.strip() for line in f.readlines()]
with open("tones.txt", 'rt') as f:
tones = [line.strip() for line in f.readlines()]
voc_phones = Vocab(phones, start_symbol=None, end_symbol=None)
voc_tones = Vocab(tones, start_symbol=None, end_symbol=None)
def segment(sentence):
segments = re.split(r'[：，；。？！]', sentence)
segments = [seg for seg in segments if len(seg)]
return segments
def g2p(sentence):
segments = segment(sentence)
phones = []
phones.append('sil')
tones = []
tones.append('0')
for seg in segments:
seg = jieba.lcut(seg)
initials = lazy_pinyin(
seg, neutral_tone_with_five=True, style=Style.INITIALS)
finals = lazy_pinyin(
seg, neutral_tone_with_five=True, style=Style.FINALS_TONE3)
for c, v in zip(initials, finals):
# NOTE: post process for pypinyin outputs
# we discriminate i, ii and iii
if re.match(r'i\d', v):
if c in ['z', 'c', 's']:
v = re.sub('i', 'ii', v)
elif c in ['zh', 'ch', 'sh', 'r']:
v = re.sub('i', 'iii', v)
if c:
phones.append(c)
tones.append('0')
if v:
phones.append(v[:-1])
tones.append(v[-1])
phones.append('sp')
tones.append('0')
phones[-1] = 'sil'
tones[-1] = '0'
return (phones, tones)
def p2id(voc, phonemes):
phone_ids = [voc.lookup(item) for item in phonemes]
return np.array(phone_ids, np.int64)
def t2id(voc, tones):
tone_ids = [voc.lookup(item) for item in tones]
return np.array(tone_ids, np.int64)
def text_analysis(sentence):
phonemes, tones = g2p(sentence)
print(sentence)
print([p + t if t != '0' else p for p, t in zip(phonemes, tones)])
phone_ids = p2id(voc_phones, phonemes)
tone_ids = t2id(voc_tones, tones)
phones = paddle.to_tensor(phone_ids)
tones = paddle.to_tensor(tone_ids)
return phones, tones

150
examples/speedyspeech/baker/normalize.py Normal file
View File

@ -0,0 +1,150 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Normalize feature files and dump them."""
import argparse
import logging
import os
from copy import copy
from operator import itemgetter
from pathlib import Path
import numpy as np
import yaml
import jsonlines
from sklearn.preprocessing import StandardScaler
from tqdm import tqdm
from parakeet.frontend.vocab import Vocab
from parakeet.datasets.data_table import DataTable
from config import get_cfg_default
def main():
"""Run preprocessing process."""
parser = argparse.ArgumentParser(
description="Normalize dumped raw features (See detail in parallel_wavegan/bin/normalize.py)."
)
parser.add_argument(
"--metadata",
type=str,
required=True,
help="directory including feature files to be normalized. "
"you need to specify either *-scp or rootdir.")
parser.add_argument(
"--dumpdir",
type=str,
required=True,
help="directory to dump normalized feature files.")
parser.add_argument(
"--stats", type=str, required=True, help="statistics file.")
parser.add_argument(
"--phones",
type=str,
default="phones.txt",
help="phone vocabulary file.")
parser.add_argument(
"--tones", type=str, default="tones.txt", help="tone vocabulary file.")
parser.add_argument(
"--config", type=str, help="yaml format configuration file.")
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)")
args = parser.parse_args()
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
)
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
)
else:
logging.basicConfig(
level=logging.WARN,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
)
logging.warning('Skip DEBUG/INFO messages')
# load config
config = get_cfg_default()
if args.config:
config.merge_from_file(args.config)
# check directory existence
dumpdir = Path(args.dumpdir).resolve()
dumpdir.mkdir(parents=True, exist_ok=True)
# get dataset
with jsonlines.open(args.metadata, 'r') as reader:
metadata = list(reader)
dataset = DataTable(metadata, converters={'feats': np.load, })
logging.info(f"The number of files = {len(dataset)}.")
# restore scaler
scaler = StandardScaler()
scaler.mean_ = np.load(args.stats)[0]
scaler.scale_ = np.load(args.stats)[1]
# from version 0.23.0, this information is needed
scaler.n_features_in_ = scaler.mean_.shape[0]
with open(args.phones, 'rt') as f:
phones = [line.strip() for line in f.readlines()]
with open(args.tones, 'rt') as f:
tones = [line.strip() for line in f.readlines()]
voc_phones = Vocab(phones, start_symbol=None, end_symbol=None)
voc_tones = Vocab(tones, start_symbol=None, end_symbol=None)
# process each file
output_metadata = []
for item in tqdm(dataset):
utt_id = item['utt_id']
mel = item['feats']
# normalize
mel = scaler.transform(mel)
# save
mel_path = dumpdir / f"{utt_id}-feats.npy"
np.save(mel_path, mel.astype(np.float32), allow_pickle=False)
phone_ids = [voc_phones.lookup(p) for p in item['phones']]
tone_ids = [voc_tones.lookup(t) for t in item['tones']]
output_metadata.append({
'utt_id': utt_id,
'phones': phone_ids,
'tones': tone_ids,
'num_phones': item['num_phones'],
'num_frames': item['num_frames'],
'durations': item['durations'],
'feats': str(mel_path),
})
output_metadata.sort(key=itemgetter('utt_id'))
output_metadata_path = Path(args.dumpdir) / "metadata.jsonl"
with jsonlines.open(output_metadata_path, 'w') as writer:
for item in output_metadata:
writer.write(item)
logging.info(f"metadata dumped into {output_metadata_path}")
if __name__ == "__main__":
main()

99
examples/speedyspeech/baker/phones.txt Normal file
View File

@ -0,0 +1,99 @@
b
p
m
f
d
t
n
l
g
k
h
zh
ch
sh
r
z
c
s
j
q
x
a
ar
ai
air
ao
aor
an
anr
ang
angr
e
er
ei
eir
en
enr
eng
engr
o
or
ou
our
ong
ongr
ii
iir
iii
iiir
i
ir
ia
iar
iao
iaor
ian
ianr
iang
iangr
ie
ier
io
ior
iou
iour
iong
iongr
in
inr
ing
ingr
u
ur
ua
uar
uai
uair
uan
uanr
uang
uangr
uei
ueir
uo
uor
uen
uenr
ueng
uengr
v
vr
ve
ver
van
vanr
vn
vnr
sil
sp

309
examples/speedyspeech/baker/preprocess.py Normal file
View File

@ -0,0 +1,309 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import List, Dict, Any
import soundfile as sf
import librosa
import numpy as np
import argparse
import yaml
import json
import re
import jsonlines
import concurrent.futures
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
from pathlib import Path
import tqdm
from operator import itemgetter
from praatio import tgio
import logging
from config import get_cfg_default
from tg_utils import validate_textgrid
def logmelfilterbank(audio,
sr,
n_fft=1024,
hop_length=256,
win_length=None,
window="hann",
n_mels=80,
fmin=None,
fmax=None,
eps=1e-10):
"""Compute log-Mel filterbank feature.
Parameters
----------
audio : ndarray
Audio signal (T,).
sr : int
Sampling rate.
n_fft : int
FFT size. (Default value = 1024)
hop_length : int
Hop size. (Default value = 256)
win_length : int
Window length. If set to None, it will be the same as fft_size. (Default value = None)
window : str
Window function type. (Default value = "hann")
n_mels : int
Number of mel basis. (Default value = 80)
fmin : int
Minimum frequency in mel basis calculation. (Default value = None)
fmax : int
Maximum frequency in mel basis calculation. (Default value = None)
eps : float
Epsilon value to avoid inf in log calculation. (Default value = 1e-10)
Returns
-------
np.ndarray
Log Mel filterbank feature (#frames, num_mels).
"""
# get amplitude spectrogram
x_stft = librosa.stft(
audio,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
pad_mode="reflect")
spc = np.abs(x_stft) # (#bins, #frames,)
# get mel basis
fmin = 0 if fmin is None else fmin
fmax = sr / 2 if fmax is None else fmax
mel_basis = librosa.filters.mel(sr, n_fft, n_mels, fmin, fmax)
return np.log10(np.maximum(eps, np.dot(mel_basis, spc)))
def process_sentence(config: Dict[str, Any],
fp: Path,
alignment_fp: Path,
output_dir: Path):
utt_id = fp.stem
# reading
y, sr = librosa.load(fp, sr=config.sr) # resampling may occur
assert len(y.shape) == 1, f"{utt_id} is not a mono-channel audio."
assert np.abs(y).max(
) <= 1.0, f"{utt_id} is seems to be different that 16 bit PCM."
duration = librosa.get_duration(y, sr=sr)
# intervals with empty lables are ignored
alignment = tgio.openTextgrid(alignment_fp)
# validate text grid against audio file
num_samples = y.shape[0]
validate_textgrid(alignment, num_samples, sr)
# only with baker's annotation
intervals = alignment.tierDict[alignment.tierNameList[0]].entryList
first, last = intervals[0], intervals[-1]
if not (first.label == "sil" and first.end < duration):
logging.warning(
f" There is something wrong with the fisrt interval {first} in utterance: {utt_id}"
)
if not (last.label == "sil" and last.start < duration):
logging.warning(
f" There is something wrong with the last interval {last} in utterance: {utt_id}"
)
logmel = logmelfilterbank(
y,
sr=sr,
n_fft=config.n_fft,
window=config.window,
win_length=config.win_length,
hop_length=config.hop_length,
n_mels=config.n_mels,
fmin=config.fmin,
fmax=config.fmax)
# extract phone and duration
phones = []
tones = []
ends = []
durations_sec = []
for interval in intervals:
label = interval.label
label = label.replace("sp1", "sp") # Baker has sp1 rather than sp
# split tone from finals
match = re.match(r'^(\w+)([012345])$', label)
if match:
phones.append(match.group(1))
tones.append(match.group(2))
else:
phones.append(label)
tones.append('0')
end = min(duration, interval.end)
ends.append(end)
durations_sec.append(end - interval.start) # duration in seconds
frame_pos = librosa.time_to_frames(
ends, sr=sr, hop_length=config.hop_length)
durations_frame = np.diff(frame_pos, prepend=0)
num_frames = logmel.shape[-1] # number of frames of the spectrogram
extra = np.sum(durations_frame) - num_frames
assert extra <= 0, (
f"Number of frames inferred from alignemnt is "
f"larger than number of frames of the spectrogram by {extra} frames")
durations_frame[-1] += (-extra)
assert np.sum(durations_frame) == num_frames
durations_frame = durations_frame.tolist()
mel_path = output_dir / (utt_id + "_feats.npy")
np.save(mel_path, logmel.T) # (num_frames, n_mels)
record = {
"utt_id": utt_id,
"phones": phones,
"tones": tones,
"num_phones": len(phones),
"num_frames": num_frames,
"durations": durations_frame,
"feats": str(mel_path.resolve()), # use absolute path
}
return record
def process_sentences(config,
fps: List[Path],
alignment_fps: List[Path],
output_dir: Path,
nprocs: int=1):
if nprocs == 1:
results = []
for fp, alignment_fp in tqdm.tqdm(
zip(fps, alignment_fps), total=len(fps)):
results.append(
process_sentence(config, fp, alignment_fp, output_dir))
else:
with ThreadPoolExecutor(nprocs) as pool:
futures = []
with tqdm.tqdm(total=len(fps)) as progress:
for fp, alignment_fp in zip(fps, alignment_fps):
future = pool.submit(process_sentence, config, fp,
alignment_fp, output_dir)
future.add_done_callback(lambda p: progress.update())
futures.append(future)
results = []
for ft in futures:
results.append(ft.result())
results.sort(key=itemgetter("utt_id"))
with jsonlines.open(output_dir / "metadata.jsonl", 'w') as writer:
for item in results:
writer.write(item)
print("Done")
def main():
# parse config and args
parser = argparse.ArgumentParser(
description="Preprocess audio and then extract features.")
parser.add_argument(
"--rootdir",
default=None,
type=str,
help="directory to baker dataset.")
parser.add_argument(
"--dumpdir",
type=str,
required=True,
help="directory to dump feature files.")
parser.add_argument(
"--config", type=str, help="yaml format configuration file.")
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)")
parser.add_argument(
"--num_cpu", type=int, default=1, help="number of process.")
args = parser.parse_args()
C = get_cfg_default()
if args.config:
C.merge_from_file(args.config)
C.freeze()
if args.verbose > 1:
print(vars(args))
print(C)
root_dir = Path(args.rootdir).expanduser()
dumpdir = Path(args.dumpdir).expanduser()
dumpdir.mkdir(parents=True, exist_ok=True)
wav_files = sorted(list((root_dir / "Wave").rglob("*.wav")))
alignment_files = sorted(
list((root_dir / "PhoneLabeling").rglob("*.interval")))
# filter out several files that have errors in annotation
exclude = {'000611', '000662', '002365', '005107'}
wav_files = [f for f in wav_files if f.stem not in exclude]
alignment_files = [f for f in alignment_files if f.stem not in exclude]
# split data into 3 sections
num_train = 9800
num_dev = 100
train_wav_files = wav_files[:num_train]
dev_wav_files = wav_files[num_train:num_train + num_dev]
test_wav_files = wav_files[num_train + num_dev:]
train_alignment_files = alignment_files[:num_train]
dev_alignment_files = alignment_files[num_train:num_train + num_dev]
test_alignment_files = alignment_files[num_train + num_dev:]
train_dump_dir = dumpdir / "train" / "raw"
train_dump_dir.mkdir(parents=True, exist_ok=True)
dev_dump_dir = dumpdir / "dev" / "raw"
dev_dump_dir.mkdir(parents=True, exist_ok=True)
test_dump_dir = dumpdir / "test" / "raw"
test_dump_dir.mkdir(parents=True, exist_ok=True)
# process for the 3 sections
process_sentences(
C,
train_wav_files,
train_alignment_files,
train_dump_dir,
nprocs=args.num_cpu)
process_sentences(
C,
dev_wav_files,
dev_alignment_files,
dev_dump_dir,
nprocs=args.num_cpu)
process_sentences(
C,
test_wav_files,
test_alignment_files,
test_dump_dir,
nprocs=args.num_cpu)
if __name__ == "__main__":
main()

6
examples/speedyspeech/baker/preprocess.sh Normal file
View File

@ -0,0 +1,6 @@
python preprocess.py --rootdir=~/datasets/BZNSYP/ --dumpdir=dump --num_cpu=20
python compute_statistics.py --metadata=dump/train/raw/metadata.jsonl --field-name="feats" --output=dump/train/stats.npy
python normalize.py --metadata=dump/train/raw/metadata.jsonl --dumpdir=dump/train/norm --stats=dump/train/stats.npy
python normalize.py --metadata=dump/dev/raw/metadata.jsonl --dumpdir=dump/dev/norm --stats=dump/train/stats.npy
python normalize.py --metadata=dump/test/raw/metadata.jsonl --dumpdir=dump/test/norm --stats=dump/train/stats.npy

6
examples/speedyspeech/baker/run.sh Normal file
View File

@ -0,0 +1,6 @@
python train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=conf/default.yaml \
--output-dir=exp/default \
--nprocs=1

16
examples/speedyspeech/baker/sentences.txt Normal file
View File

@ -0,0 +1,16 @@
001 凯莫瑞安联合体的经济崩溃，迫在眉睫。
002 对于所有想要离开那片废土，去寻找更美好生活的人来说。
003 克哈，是你们所有人安全的港湾。
004 为了保护尤摩扬人民不受异虫的残害，我所做的，比他们自己的领导委员会都多。
005 无论他们如何诽谤我，我将继续为所有泰伦人的最大利益，而努力奋斗。
006 身为你们的元首，我带领泰伦人实现了人类统治领地和经济的扩张。
007 我们将继续成长，用行动回击那些只会说风凉话，不愿意和我们相向而行的害群之马。
008 帝国武装力量，无数的优秀儿女，正时刻守卫着我们的家园大门，但是他们孤木难支。
009 凡是今天应征入伍者，所获的所有刑罚罪责，减半。
010 激进分子和异见者希望你们一听见枪声，就背弃多年的和平与繁荣。
011 他们没有勇气和能力，带领人类穿越一个充满危险的星系。
012 法治是我们的命脉，然而它却受到前所未有的挑战。
013 我将恢复我们帝国的荣光，绝不会向任何外星势力低头。
014 我已经驯服了异虫，荡平了星灵。如今它们的创造者，想要夺走我们拥有的一切。
015 永远记住，谁才是最能保护你们的人。
016 不要听信别人的谗言，我不是什么克隆人。

110
examples/speedyspeech/baker/speedyspeech_updater.py Normal file
View File

@ -0,0 +1,110 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from paddle.nn import functional as F
from paddle.fluid.layers import huber_loss
from parakeet.modules.ssim import ssim
from parakeet.modules.losses import masked_l1_loss, weighted_mean
from parakeet.training.reporter import report
from parakeet.training.updaters.standard_updater import StandardUpdater
from parakeet.training.extensions.evaluator import StandardEvaluator
from parakeet.models.speedyspeech import SpeedySpeech
class SpeedySpeechUpdater(StandardUpdater):
def update_core(self, batch):
decoded, predicted_durations = self.model(
text=batch["phones"],
tones=batch["tones"],
plens=batch["num_phones"],
durations=batch["durations"])
target_mel = batch["feats"]
spec_mask = F.sequence_mask(
batch["num_frames"], dtype=target_mel.dtype).unsqueeze(-1)
text_mask = F.sequence_mask(
batch["num_phones"], dtype=predicted_durations.dtype)
# spec loss
l1_loss = masked_l1_loss(decoded, target_mel, spec_mask)
# duration loss
target_durations = batch["durations"]
target_durations = paddle.maximum(
target_durations.astype(predicted_durations.dtype),
paddle.to_tensor([1.0]))
duration_loss = weighted_mean(
huber_loss(
predicted_durations, paddle.log(target_durations), delta=1.0),
text_mask, )
# ssim loss
ssim_loss = 1.0 - ssim((decoded * spec_mask).unsqueeze(1),
(target_mel * spec_mask).unsqueeze(1))
loss = l1_loss + ssim_loss + duration_loss
optimizer = self.optimizer
optimizer.clear_grad()
loss.backward()
optimizer.step()
report("train/loss", float(loss))
report("train/l1_loss", float(l1_loss))
report("train/duration_loss", float(duration_loss))
report("train/ssim_loss", float(ssim_loss))
class SpeedySpeechEvaluator(StandardEvaluator):
def evaluate_core(self, batch):
print("fire")
decoded, predicted_durations = self.model(
text=batch["phones"],
tones=batch["tones"],
plens=batch["num_phones"],
durations=batch["durations"])
target_mel = batch["feats"]
spec_mask = F.sequence_mask(
batch["num_frames"], dtype=target_mel.dtype).unsqueeze(-1)
text_mask = F.sequence_mask(
batch["num_phones"], dtype=predicted_durations.dtype)
# spec loss
l1_loss = masked_l1_loss(decoded, target_mel, spec_mask)
# duration loss
target_durations = batch["durations"]
target_durations = paddle.maximum(
target_durations.astype(predicted_durations.dtype),
paddle.to_tensor([1.0]))
duration_loss = weighted_mean(
huber_loss(
predicted_durations, paddle.log(target_durations), delta=1.0),
text_mask, )
# ssim loss
ssim_loss = 1.0 - ssim((decoded * spec_mask).unsqueeze(1),
(target_mel * spec_mask).unsqueeze(1))
loss = l1_loss + ssim_loss + duration_loss
# import pdb; pdb.set_trace()
report("eval/loss", float(loss))
report("eval/l1_loss", float(l1_loss))
report("eval/duration_loss", float(duration_loss))
report("eval/ssim_loss", float(ssim_loss))

146
examples/speedyspeech/baker/synthesize.py Normal file
View File

@ -0,0 +1,146 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
import logging
import argparse
import dataclasses
from pathlib import Path
import yaml
import jsonlines
import paddle
import numpy as np
import soundfile as sf
import paddle
from paddle import nn
from paddle.nn import functional as F
from paddle import distributed as dist
from yacs.config import CfgNode
from parakeet.datasets.data_table import DataTable
from parakeet.models.speedyspeech import SpeedySpeech, SpeedySpeechInference
from parakeet.models.parallel_wavegan import PWGGenerator, PWGInference
from parakeet.modules.normalizer import ZScore
def evaluate(args, speedyspeech_config, pwg_config):
# dataloader has been too verbose
logging.getLogger("DataLoader").disabled = True
# construct dataset for evaluation
with jsonlines.open(args.test_metadata, 'r') as reader:
test_metadata = list(reader)
test_dataset = DataTable(
data=test_metadata, fields=["utt_id", "phones", "tones"])
model = SpeedySpeech(**speedyspeech_config["model"])
model.set_state_dict(
paddle.load(args.speedyspeech_checkpoint)["main_params"])
model.eval()
vocoder = PWGGenerator(**pwg_config["generator_params"])
vocoder.set_state_dict(paddle.load(args.pwg_params))
vocoder.remove_weight_norm()
vocoder.eval()
print("model done!")
stat = np.load(args.speedyspeech_stat)
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
speedyspeech_normalizer = ZScore(mu, std)
stat = np.load(args.pwg_stat)
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
pwg_normalizer = ZScore(mu, std)
speedyspeech_inferencce = SpeedySpeechInference(speedyspeech_normalizer,
model)
pwg_inference = PWGInference(pwg_normalizer, vocoder)
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
for datum in test_dataset:
utt_id = datum["utt_id"]
phones = paddle.to_tensor(datum["phones"])
tones = paddle.to_tensor(datum["tones"])
with paddle.no_grad():
wav = pwg_inference(speedyspeech_inferencce(phones, tones))
sf.write(
output_dir / (utt_id + ".wav"),
wav.numpy(),
samplerate=speedyspeech_config.sr)
print(f"{utt_id} done!")
def main():
# parse args and config and redirect to train_sp
parser = argparse.ArgumentParser(
description="Synthesize with speedyspeech & parallel wavegan.")
parser.add_argument(
"--speedyspeech-config",
type=str,
help="config file to overwrite default config")
parser.add_argument(
"--speedyspeech-checkpoint",
type=str,
help="speedyspeech checkpoint to load.")
parser.add_argument(
"--speedyspeech-stat",
type=str,
help="mean and standard deviation used to normalize spectrogram when training speedyspeech."
)
parser.add_argument(
"--pwg-config",
type=str,
help="mean and standard deviation used to normalize spectrogram when training speedyspeech."
)
parser.add_argument(
"--pwg-params",
type=str,
help="parallel wavegan generator parameters to load.")
parser.add_argument(
"--pwg-stat",
type=str,
help="mean and standard deviation used to normalize spectrogram when training speedyspeech."
)
parser.add_argument("--test-metadata", type=str, help="test metadata")
parser.add_argument("--output-dir", type=str, help="output dir")
parser.add_argument(
"--device", type=str, default="gpu", help="device type to use")
parser.add_argument("--verbose", type=int, default=1, help="verbose")
args = parser.parse_args()
with open(args.speedyspeech_config) as f:
speedyspeech_config = CfgNode(yaml.safe_load(f))
with open(args.pwg_config) as f:
pwg_config = CfgNode(yaml.safe_load(f))
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(speedyspeech_config)
print(pwg_config)
evaluate(args, speedyspeech_config, pwg_config)
if __name__ == "__main__":
main()

10
examples/speedyspeech/baker/synthesize.sh Normal file
View File

@ -0,0 +1,10 @@
python synthesize.py \
--speedyspeech-config=conf/default.yaml \
--speedyspeech-checkpoint=exp/debug/checkpoints/snapshot_iter_91800.pdz \
--speedyspeech-stat=dump/train/stats.npy \
--pwg-config=../../parallelwave_gan/baker/conf/default.yaml \
--pwg-params=../../parallelwave_gan/baker/converted.pdparams \
--pwg-stat=../../parallelwave_gan/baker/dump/train/stats.npy \
--test-metadata=dump/test/norm/metadata.jsonl \
--output-dir=exp/debug/test \
--device="gpu"

150
examples/speedyspeech/baker/synthesize_e2e.py Normal file
View File

@ -0,0 +1,150 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
import logging
import argparse
import dataclasses
from pathlib import Path
import yaml
import jsonlines
import paddle
import numpy as np
import soundfile as sf
import paddle
from paddle import nn
from paddle.nn import functional as F
from paddle import distributed as dist
from yacs.config import CfgNode
from parakeet.datasets.data_table import DataTable
from parakeet.models.speedyspeech import SpeedySpeech, SpeedySpeechInference
from parakeet.models.parallel_wavegan import PWGGenerator, PWGInference
from parakeet.modules.normalizer import ZScore
from frontend import text_analysis
def evaluate(args, speedyspeech_config, pwg_config):
# dataloader has been too verbose
logging.getLogger("DataLoader").disabled = True
# construct dataset for evaluation
sentences = []
with open(args.text, 'rt') as f:
for line in f:
utt_id, sentence = line.strip().split()
sentences.append((utt_id, sentence))
model = SpeedySpeech(**speedyspeech_config["model"])
model.set_state_dict(
paddle.load(args.speedyspeech_checkpoint)["main_params"])
model.eval()
vocoder = PWGGenerator(**pwg_config["generator_params"])
vocoder.set_state_dict(paddle.load(args.pwg_params))
vocoder.remove_weight_norm()
vocoder.eval()
print("model done!")
stat = np.load(args.speedyspeech_stat)
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
speedyspeech_normalizer = ZScore(mu, std)
stat = np.load(args.pwg_stat)
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
pwg_normalizer = ZScore(mu, std)
speedyspeech_inferencce = SpeedySpeechInference(speedyspeech_normalizer,
model)
pwg_inference = PWGInference(pwg_normalizer, vocoder)
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
for utt_id, sentence in sentences:
phones, tones = text_analysis(sentence)
with paddle.no_grad():
wav = pwg_inference(speedyspeech_inferencce(phones, tones))
sf.write(
output_dir / (utt_id + ".wav"),
wav.numpy(),
samplerate=speedyspeech_config.sr)
print(f"{utt_id} done!")
def main():
# parse args and config and redirect to train_sp
parser = argparse.ArgumentParser(
description="Synthesize with speedyspeech & parallel wavegan.")
parser.add_argument(
"--speedyspeech-config",
type=str,
help="config file to overwrite default config")
parser.add_argument(
"--speedyspeech-checkpoint",
type=str,
help="speedyspeech checkpoint to load.")
parser.add_argument(
"--speedyspeech-stat",
type=str,
help="mean and standard deviation used to normalize spectrogram when training speedyspeech."
)
parser.add_argument(
"--pwg-config",
type=str,
help="mean and standard deviation used to normalize spectrogram when training speedyspeech."
)
parser.add_argument(
"--pwg-params",
type=str,
help="parallel wavegan generator parameters to load.")
parser.add_argument(
"--pwg-stat",
type=str,
help="mean and standard deviation used to normalize spectrogram when training speedyspeech."
)
parser.add_argument(
"--text",
type=str,
help="text to synthesize, a 'utt_id sentence' pair per line")
parser.add_argument("--output-dir", type=str, help="output dir")
parser.add_argument(
"--device", type=str, default="gpu", help="device type to use")
parser.add_argument("--verbose", type=int, default=1, help="verbose")
args = parser.parse_args()
with open(args.speedyspeech_config) as f:
speedyspeech_config = CfgNode(yaml.safe_load(f))
with open(args.pwg_config) as f:
pwg_config = CfgNode(yaml.safe_load(f))
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(speedyspeech_config)
print(pwg_config)
evaluate(args, speedyspeech_config, pwg_config)
if __name__ == "__main__":
main()

10
examples/speedyspeech/baker/synthesize_e2e.sh Normal file
View File

@ -0,0 +1,10 @@
python synthesize_e2e.py \
--speedyspeech-config=conf/default.yaml \
--speedyspeech-checkpoint=exp/debug/checkpoints/snapshot_iter_91800.pdz \
--speedyspeech-stat=dump/train/stats.npy \
--pwg-config=../../parallelwave_gan/baker/conf/default.yaml \
--pwg-params=../../parallelwave_gan/baker/converted.pdparams \
--pwg-stat=../../parallelwave_gan/baker/dump/train/stats.npy \
--text=sentences.txt \
--output-dir=exp/e2e \
--device="gpu"

27
examples/speedyspeech/baker/tg_utils.py Normal file
View File

@ -0,0 +1,27 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import librosa
from praatio import tgio
def validate_textgrid(text_grid, num_samples, sr):
"""Validate Text Grid to make sure that the time interval annotated
by the tex grid file does not go beyond the audio file.
"""
start = text_grid.minTimestamp
end = text_grid.maxTimestamp
end_audio = librosa.samples_to_time(num_samples, sr)
return start == 0.0 and end <= end_audio

6
examples/speedyspeech/baker/tones.txt Normal file
View File

@ -0,0 +1,6 @@
0
1
2
3
4
5

186
examples/speedyspeech/baker/train.py Normal file
View File

@ -0,0 +1,186 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
import logging
import argparse
import dataclasses
from pathlib import Path
import yaml
import jsonlines
import paddle
import numpy as np
from paddle import nn
from paddle.nn import functional as F
from paddle import distributed as dist
from paddle.io import DataLoader, DistributedBatchSampler
from paddle.optimizer import Adam # No RAdaom
from paddle.optimizer.lr import StepDecay
from paddle import DataParallel
from visualdl import LogWriter
from parakeet.datasets.data_table import DataTable
from parakeet.models.speedyspeech import SpeedySpeech
from parakeet.training.updater import UpdaterBase
from parakeet.training.trainer import Trainer
from parakeet.training.reporter import report
from parakeet.training import extension
from parakeet.training.extensions.snapshot import Snapshot
from parakeet.training.extensions.visualizer import VisualDL
from parakeet.training.seeding import seed_everything
from batch_fn import collate_baker_examples
from speedyspeech_updater import SpeedySpeechUpdater, SpeedySpeechEvaluator
from config import get_cfg_default
def train_sp(args, config):
# decides device type and whether to run in parallel
# setup running environment correctly
if not paddle.is_compiled_with_cuda:
paddle.set_device("cpu")
else:
paddle.set_device("gpu")
world_size = paddle.distributed.get_world_size()
if world_size > 1:
paddle.distributed.init_parallel_env()
# set the random seed, it is a must for multiprocess training
seed_everything(config.seed)
print(
f"rank: {dist.get_rank()}, pid: {os.getpid()}, parent_pid: {os.getppid()}",
)
# dataloader has been too verbose
logging.getLogger("DataLoader").disabled = True
# construct dataset for training and validation
with jsonlines.open(args.train_metadata, 'r') as reader:
train_metadata = list(reader)
train_dataset = DataTable(
data=train_metadata,
fields=[
"phones", "tones", "num_phones", "num_frames", "feats", "durations"
],
converters={"feats": np.load, }, )
with jsonlines.open(args.dev_metadata, 'r') as reader:
dev_metadata = list(reader)
dev_dataset = DataTable(
data=dev_metadata,
fields=[
"phones", "tones", "num_phones", "num_frames", "feats", "durations"
],
converters={"feats": np.load, }, )
# collate function and dataloader
train_sampler = DistributedBatchSampler(
train_dataset,
batch_size=config.batch_size,
shuffle=False,
drop_last=True)
# dev_sampler = DistributedBatchSampler(dev_dataset,
# batch_size=config.batch_size,
# shuffle=False,
# drop_last=False)
print("samplers done!")
train_dataloader = DataLoader(
train_dataset,
batch_sampler=train_sampler,
collate_fn=collate_baker_examples,
num_workers=config.num_workers)
dev_dataloader = DataLoader(
dev_dataset,
shuffle=False,
drop_last=False,
batch_size=config.batch_size,
collate_fn=collate_baker_examples,
num_workers=config.num_workers)
print("dataloaders done!")
# batch = collate_baker_examples([train_dataset[i] for i in range(10)])
# # batch = collate_baker_examples([dev_dataset[i] for i in range(10)])
# import pdb; pdb.set_trace()
model = SpeedySpeech(**config["model"])
if world_size > 1:
model = DataParallel(model) # TODO, do not use vocab size from config
# print(model)
print("model done!")
optimizer = Adam(
0.001,
parameters=model.parameters(),
grad_clip=nn.ClipGradByGlobalNorm(5.0))
print("optimizer done!")
updater = SpeedySpeechUpdater(
model=model, optimizer=optimizer, dataloader=train_dataloader)
output_dir = Path(args.output_dir)
trainer = Trainer(updater, (config.max_epoch, 'epoch'), output_dir)
evaluator = SpeedySpeechEvaluator(model, dev_dataloader)
if dist.get_rank() == 0:
trainer.extend(evaluator, trigger=(1, "epoch"))
writer = LogWriter(str(output_dir))
trainer.extend(VisualDL(writer), trigger=(1, "iteration"))
trainer.extend(
Snapshot(max_size=config.num_snapshots), trigger=(1, 'epoch'))
print(trainer.extensions)
trainer.run()
def main():
# parse args and config and redirect to train_sp
parser = argparse.ArgumentParser(description="Train a ParallelWaveGAN "
"model with Baker Mandrin TTS dataset.")
parser.add_argument(
"--config", type=str, help="config file to overwrite default config")
parser.add_argument("--train-metadata", type=str, help="training data")
parser.add_argument("--dev-metadata", type=str, help="dev data")
parser.add_argument("--output-dir", type=str, help="output dir")
parser.add_argument(
"--device", type=str, default="gpu", help="device type to use")
parser.add_argument(
"--nprocs", type=int, default=1, help="number of processes")
parser.add_argument("--verbose", type=int, default=1, help="verbose")
args = parser.parse_args()
if args.device == "cpu" and args.nprocs > 1:
raise RuntimeError("Multiprocess training on CPU is not supported.")
config = get_cfg_default()
if args.config:
config.merge_from_file(args.config)
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(config)
print(
f"master see the word size: {dist.get_world_size()}, from pid: {os.getpid()}"
)
# dispatch
if args.nprocs > 1:
dist.spawn(train_sp, (args, config), nprocs=args.nprocs)
else:
train_sp(args, config)
if __name__ == "__main__":
main()

24
parakeet/data/batch.py
View File

@ -161,3 +161,27 @@ def batch_spec(minibatch, pad_value=0., time_major=False, dtype=np.float32):
mode='constant', mode='constant',
constant_values=pad_value)) constant_values=pad_value))
return np.array(batch, dtype=dtype), np.array(lengths, dtype=np.int64) return np.array(batch, dtype=dtype), np.array(lengths, dtype=np.int64)
def batch_sequences(sequences, axis=0, pad_value=0):
# import pdb; pdb.set_trace()
seq = sequences[0]
ndim = seq.ndim
if axis < 0:
axis += ndim
dtype = seq.dtype
pad_value = dtype.type(pad_value)
seq_lengths = [seq.shape[axis] for seq in sequences]
max_length = np.max(seq_lengths)
padded_sequences = []
for seq, length in zip(sequences, seq_lengths):
padding = [(0, 0)] * axis + [(0, max_length - length)] + [(0, 0)] * (
ndim - axis - 1)
padded_seq = np.pad(seq,
padding,
mode='constant',
constant_values=pad_value)
padded_sequences.append(padded_seq)
batch = np.stack(padded_sequences)
return batch

12
parakeet/models/parallel_wavegan.py
View File

@ -768,3 +768,15 @@ class ResidualPWGDiscriminator(nn.Layer):
pass pass
self.apply(_remove_weight_norm) self.apply(_remove_weight_norm)
class PWGInference(nn.Layer):
def __init__(self, normalizer, pwg_generator):
super().__init__()
self.normalizer = normalizer
self.pwg_generator = pwg_generator
def forward(self, logmel):
normalized_mel = self.normalizer(logmel)
wav = self.pwg_generator.inference(normalized_mel)
return wav

226
parakeet/models/speedyspeech.py Normal file
View File

@ -0,0 +1,226 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import numpy as np
import paddle
from paddle import Tensor
from paddle import nn
from paddle.nn import functional as F
from paddle.nn import initializer as I
from parakeet.modules.positional_encoding import sinusoid_position_encoding
from parakeet.modules.expansion import expand
class ResidualBlock(nn.Layer):
def __init__(self, channels, kernel_size, dilation, n=2):
super().__init__()
blocks = [
nn.Sequential(
nn.Conv1D(
channels,
channels,
kernel_size,
dilation=dilation,
padding="same",
data_format="NLC"),
nn.ReLU(),
nn.BatchNorm1D(
channels, data_format="NLC"), ) for _ in range(n)
]
self.blocks = nn.Sequential(*blocks)
def forward(self, x):
return x + self.blocks(x)
class TextEmbedding(nn.Layer):
def __init__(self,
vocab_size: int,
embedding_size: int,
tone_vocab_size: int=None,
tone_embedding_size: int=None,
padding_idx: int=None,
tone_padding_idx: int=None,
concat: bool=False):
super().__init__()
self.text_embedding = nn.Embedding(vocab_size, embedding_size,
padding_idx)
if tone_vocab_size:
tone_embedding_size = tone_embedding_size or embedding_size
if tone_embedding_size != embedding_size and not concat:
raise ValueError(
"embedding size != tone_embedding size, only conat is avaiable."
)
self.tone_embedding = nn.Embedding(
tone_vocab_size, tone_embedding_size, tone_padding_idx)
self.concat = concat
def forward(self, text, tone=None):
text_embed = self.text_embedding(text)
if tone is None:
return text_embed
tone_embed = self.tone_embedding(tone)
if self.concat:
embed = paddle.concat([text_embed, tone_embed], -1)
else:
embed = text_embed + tone_embed
return embed
class SpeedySpeechEncoder(nn.Layer):
def __init__(self, vocab_size, tone_size, hidden_size, kernel_size,
dilations):
super().__init__()
self.embedding = TextEmbedding(
vocab_size,
hidden_size,
tone_size,
padding_idx=0,
tone_padding_idx=0)
self.prenet = nn.Sequential(
nn.Linear(hidden_size, hidden_size),
nn.ReLU(), )
res_blocks = [
ResidualBlock(
hidden_size, kernel_size, d, n=2) for d in dilations
]
self.res_blocks = nn.Sequential(*res_blocks)
self.postnet1 = nn.Sequential(nn.Linear(hidden_size, hidden_size))
self.postnet2 = nn.Sequential(
nn.ReLU(),
nn.BatchNorm1D(
hidden_size, data_format="NLC"),
nn.Linear(hidden_size, hidden_size), )
def forward(self, text, tones):
embedding = self.embedding(text, tones)
embedding = self.prenet(embedding)
x = self.res_blocks(embedding)
x = embedding + self.postnet1(x)
x = self.postnet2(x)
return x
class DurationPredictor(nn.Layer):
def __init__(self, hidden_size):
super().__init__()
self.layers = nn.Sequential(
ResidualBlock(
hidden_size, 4, 1, n=1),
ResidualBlock(
hidden_size, 3, 1, n=1),
ResidualBlock(
hidden_size, 1, 1, n=1),
nn.Linear(hidden_size, 1))
def forward(self, x):
return paddle.squeeze(self.layers(x), -1)
class SpeedySpeechDecoder(nn.Layer):
def __init__(self, hidden_size, output_size, kernel_size, dilations):
super().__init__()
res_blocks = [
ResidualBlock(
hidden_size, kernel_size, d, n=2) for d in dilations
]
self.res_blocks = nn.Sequential(*res_blocks)
self.postnet1 = nn.Sequential(nn.Linear(hidden_size, hidden_size))
self.postnet2 = nn.Sequential(
ResidualBlock(
hidden_size, kernel_size, 1, n=2),
nn.Linear(hidden_size, output_size))
def forward(self, x):
xx = self.res_blocks(x)
x = x + self.postnet1(xx)
x = self.postnet2(x)
return x
class SpeedySpeech(nn.Layer):
def __init__(
self,
vocab_size,
encoder_hidden_size,
encoder_kernel_size,
encoder_dilations,
duration_predictor_hidden_size,
decoder_hidden_size,
decoder_output_size,
decoder_kernel_size,
decoder_dilations,
tone_size=None, ):
super().__init__()
encoder = SpeedySpeechEncoder(vocab_size, tone_size,
encoder_hidden_size, encoder_kernel_size,
encoder_dilations)
duration_predictor = DurationPredictor(duration_predictor_hidden_size)
decoder = SpeedySpeechDecoder(decoder_hidden_size, decoder_output_size,
decoder_kernel_size, decoder_dilations)
self.encoder = encoder
self.duration_predictor = duration_predictor
self.decoder = decoder
def forward(self, text, tones, plens, durations):
encodings = self.encoder(text, tones)
pred_durations = self.duration_predictor(encodings.detach()) # (B, T)
# expand encodings
durations_to_expand = durations
encodings = expand(encodings, durations_to_expand)
# decode
# remove positional encoding here
_, t_dec, feature_size = encodings.shape
encodings += sinusoid_position_encoding(t_dec, feature_size)
decoded = self.decoder(encodings)
return decoded, pred_durations
def inference(self, text, tones):
# text: [T]
# tones: [T]
text = text.unsqueeze(0)
if tones is not None:
tones = tones.unsqueeze(0)
encodings = self.encoder(text, tones)
pred_durations = self.duration_predictor(encodings) # (1, T)
durations_to_expand = paddle.round(pred_durations.exp())
durations_to_expand = (durations_to_expand).astype(paddle.int64)
encodings = expand(encodings, durations_to_expand)
shape = paddle.shape(encodings)
t_dec, feature_size = shape[1], shape[2]
encodings += sinusoid_position_encoding(t_dec, feature_size)
decoded = self.decoder(encodings)
return decoded[0]
class SpeedySpeechInference(nn.Layer):
def __init__(self, normalizer, speedyspeech_model):
super().__init__()
self.normalizer = normalizer
self.acoustic_model = speedyspeech_model
def forward(self, phones, tones):
normalized_mel = self.acoustic_model.inference(phones, tones)
logmel = self.normalizer.inverse(normalized_mel)
return logmel

6
parakeet/models/transformer_tts.py
View File

@ -403,7 +403,7 @@ class TransformerTTS(nn.Layer):
else: else:
self.toned = False self.toned = False
# position encoding matrix may be extended later # position encoding matrix may be extended later
self.encoder_pe = pe.sinusoid_positional_encoding(0, 1000, d_encoder) self.encoder_pe = pe.sinusoid_positional_encoding(1000, d_encoder)
self.encoder_pe_scalar = self.create_parameter( self.encoder_pe_scalar = self.create_parameter(
[1], attr=I.Constant(1.)) [1], attr=I.Constant(1.))
self.encoder = TransformerEncoder(d_encoder, n_heads, d_ffn, self.encoder = TransformerEncoder(d_encoder, n_heads, d_ffn,
@ -411,7 +411,7 @@ class TransformerTTS(nn.Layer):
# decoder # decoder
self.decoder_prenet = MLPPreNet(d_mel, d_prenet, d_decoder, dropout) self.decoder_prenet = MLPPreNet(d_mel, d_prenet, d_decoder, dropout)
self.decoder_pe = pe.sinusoid_positional_encoding(0, 1000, d_decoder) self.decoder_pe = pe.sinusoid_positional_encoding(1000, d_decoder)
self.decoder_pe_scalar = self.create_parameter( self.decoder_pe_scalar = self.create_parameter(
[1], attr=I.Constant(1.)) [1], attr=I.Constant(1.))
self.decoder = TransformerDecoder( self.decoder = TransformerDecoder(
@ -488,7 +488,7 @@ class TransformerTTS(nn.Layer):
# twice its length if needed # twice its length if needed
if x.shape[1] * self.r > self.decoder_pe.shape[0]: if x.shape[1] * self.r > self.decoder_pe.shape[0]:
new_T = max(x.shape[1] * self.r, self.decoder_pe.shape[0] * 2) new_T = max(x.shape[1] * self.r, self.decoder_pe.shape[0] * 2)
self.decoder_pe = pe.sinusoid_positional_encoding(0, new_T, self.decoder_pe = pe.sinusoid_positional_encoding(new_T,
self.d_decoder) self.d_decoder)
pos_enc = self.decoder_pe[:T_dec * self.r:self.r, :] pos_enc = self.decoder_pe[:T_dec * self.r:self.r, :]
x = x.scale(math.sqrt( x = x.scale(math.sqrt(

39
parakeet/modules/expansion.py Normal file
View File

@ -0,0 +1,39 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle
from paddle import Tensor
def expand(encodings: Tensor, durations: Tensor) -> Tensor:
"""
encodings: (B, T, C)
durations: (B, T)
"""
batch_size, t_enc = durations.shape
durations = durations.numpy()
slens = np.sum(durations, -1)
t_dec = np.max(slens)
M = np.zeros([batch_size, t_dec, t_enc])
for i in range(batch_size):
k = 0
for j in range(t_enc):
d = durations[i, j]
M[i, k:k + d, j] = 1
k += d
M = paddle.to_tensor(M, dtype=encodings.dtype)
encodings = paddle.matmul(M, encodings)
return encodings

29
parakeet/modules/normalizer.py Normal file
View File

@ -0,0 +1,29 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle import nn
class ZScore(nn.Layer):
# feature last
def __init__(self, mu, sigma):
super().__init__()
self.register_buffer("mu", mu)
self.register_buffer("sigma", sigma)
def forward(self, x):
return (x - self.mu) / self.sigma
def inverse(self, x):
return x * self.sigma + self.mu

81
parakeet/modules/positional_encoding.py
View File

@ -14,47 +14,56 @@
import math import math
import numpy as np import numpy as np
import paddle import paddle
from paddle import Tensor
from paddle.nn import functional as F from paddle.nn import functional as F
__all__ = ["sinusoid_positional_encoding"] __all__ = ["sinusoid_position_encoding", "scaled_position_encoding"]
def sinusoid_positional_encoding(start_index, length, size, dtype=None): def sinusoid_position_encoding(num_positions: int,
r"""Generate standard positional encoding matrix. feature_size: int,
omega: float=1.0,
.. math:: start_pos: int=0,
dtype=None) -> Tensor:
pe(pos, 2i) = sin(\frac{pos}{10000^{\frac{2i}{size}}}) \\ # return tensor shape (num_positions, feature_size)
pe(pos, 2i+1) = cos(\frac{pos}{10000^{\frac{2i}{size}}}) if (feature_size % 2 != 0):
Parameters
----------
start_index : int
The start index.
length : int
The timesteps of the positional encoding to generate.
size : int
Feature size of positional encoding.
Returns
-------
Tensor [shape=(length, size)]
The positional encoding.
Raises
------
ValueError
If ``size`` is not divisible by 2.
"""
if (size % 2 != 0):
raise ValueError("size should be divisible by 2") raise ValueError("size should be divisible by 2")
dtype = dtype or paddle.get_default_dtype() dtype = dtype or paddle.get_default_dtype()
channel = np.arange(0, size, 2)
index = np.arange(start_index, start_index + length, 1) channel = paddle.arange(0, feature_size, 2, dtype=dtype)
p = np.expand_dims(index, -1) / (10000**(channel / float(size))) index = paddle.arange(start_pos, start_pos + num_positions, 1, dtype=dtype)
encodings = np.zeros([length, size]) p = (paddle.unsqueeze(index, -1) *
encodings[:, 0::2] = np.sin(p) omega) / (10000.0**(channel / float(feature_size)))
encodings[:, 1::2] = np.cos(p) encodings = paddle.zeros([num_positions, feature_size], dtype=dtype)
encodings = paddle.to_tensor(encodings) encodings[:, 0::2] = paddle.sin(p)
encodings[:, 1::2] = paddle.cos(p)
return encodings
def scaled_position_encoding(num_positions: int,
feature_size: int,
omega: Tensor,
start_pos: int=0,
dtype=None) -> Tensor:
# omega: Tensor (batch_size, )
# return tensor shape (batch_size, num_positions, feature_size)
# consider renaming this as batched positioning encoding
if (feature_size % 2 != 0):
raise ValueError("size should be divisible by 2")
dtype = dtype or paddle.get_default_dtype()
channel = paddle.arange(0, feature_size, 2, dtype=dtype)
index = paddle.arange(
start_pos, start_pos + num_positions, 1, dtype=omega.dtype)
batch_size = omega.shape[0]
omega = paddle.unsqueeze(omega, [1, 2])
p = (paddle.unsqueeze(index, -1) *
omega) / (10000.0**(channel / float(feature_size)))
encodings = paddle.zeros(
[batch_size, num_positions, feature_size], dtype=dtype)
# it is nice to have fancy indexing and inplace operations
encodings[:, :, 0::2] = paddle.sin(p)
encodings[:, :, 1::2] = paddle.cos(p)
return encodings return encodings

84
parakeet/modules/ssim.py Normal file
View File

@ -0,0 +1,84 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from math import exp
import numpy as np
import paddle
from paddle import nn
import paddle.nn.functional as F
def gaussian(window_size, sigma):
gauss = paddle.to_tensor([
exp(-(x - window_size // 2)**2 / float(2 * sigma**2))
for x in range(window_size)
])
return gauss / gauss.sum()
def create_window(window_size, channel):
_1D_window = gaussian(window_size, 1.5).unsqueeze(1)
_2D_window = paddle.matmul(_1D_window,
paddle.transpose(_1D_window,
[1, 0])).unsqueeze([0, 1])
window = paddle.expand(_2D_window, [channel, 1, window_size, window_size])
return window
def _ssim(img1, img2, window, window_size, channel, size_average=True):
mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)
mu1_sq = mu1.pow(2)
mu2_sq = mu2.pow(2)
mu1_mu2 = mu1 * mu2
sigma1_sq = F.conv2d(
img1 * img1, window, padding=window_size // 2, groups=channel) - mu1_sq
sigma2_sq = F.conv2d(
img2 * img2, window, padding=window_size // 2, groups=channel) - mu2_sq
sigma12 = F.conv2d(
img1 * img2, window, padding=window_size // 2,
groups=channel) - mu1_mu2
C1 = 0.01**2
C2 = 0.03**2
ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) \
/ ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
if size_average:
return ssim_map.mean()
else:
return ssim_map.mean(1).mean(1).mean(1)
class SSIM(nn.Layer):
def __init__(self, window_size=11, size_average=True):
super().__init__()
self.window_size = window_size
self.size_average = size_average
self.channel = 1
self.window = create_window(window_size, self.channel)
def forward(self, img1, img2):
return _ssim(img1, img2, self.window, self.window_size, self.channel,
self.size_average)
def ssim(img1, img2, window_size=11, size_average=True):
(_, channel, _, _) = img1.shape
window = create_window(window_size, channel)
return _ssim(img1, img2, window, window_size, channel, size_average)

2
parakeet/training/trainer.py
View File

@ -123,8 +123,6 @@ class Trainer(object):
update = self.updater.update # training step update = self.updater.update # training step
stop_trigger = self.stop_trigger stop_trigger = self.stop_trigger
print(self.updater.state)
# display only one progress bar # display only one progress bar
max_iteration = None max_iteration = None
if isinstance(stop_trigger, LimitTrigger): if isinstance(stop_trigger, LimitTrigger):

15
parakeet/training/triggers/interval_trigger.py
View File

@ -12,6 +12,8 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
from copy import deepcopy
class IntervalTrigger(object): class IntervalTrigger(object):
"""A Predicate to do something every N cycle.""" """A Predicate to do something every N cycle."""
@ -23,9 +25,16 @@ class IntervalTrigger(object):
raise ValueError("period should be a positive integer.") raise ValueError("period should be a positive integer.")
self.period = period self.period = period
self.unit = unit self.unit = unit
self.last_index = None
def __call__(self, trainer): def __call__(self, trainer):
state = trainer.updater.state if self.last_index is None:
index = getattr(state, self.unit) last_index = getattr(trainer.updater.state, self.unit)
fire = index % self.period == 0 self.last_index = last_index
last_index = self.last_index
index = getattr(trainer.updater.state, self.unit)
fire = index // self.period != last_index // self.period
self.last_index = index
return fire return fire

6
parakeet/training/updaters/standard_updater.py
View File

@ -106,8 +106,8 @@ class StandardUpdater(UpdaterBase):
self.update_core(batch) self.update_core(batch)
self.state.iteration += 1 self.state.iteration += 1
if self.updaters_per_epoch is not None: if self.updates_per_epoch is not None:
if self.state.iteration % self.updaters_per_epoch == 0: if self.state.iteration % self.updates_per_epoch == 0:
self.state.epoch += 1 self.state.epoch += 1
def update_core(self, batch): def update_core(self, batch):
@ -139,7 +139,7 @@ class StandardUpdater(UpdaterBase):
self.optimizer.update() self.optimizer.update()
@property @property
def updaters_per_epoch(self): def updates_per_epoch(self):
"""Number of updater per epoch, determined by the length of the """Number of updater per epoch, determined by the length of the
dataloader.""" dataloader."""
length_of_dataloader = None length_of_dataloader = None

29
tests/test_expansion.py Normal file
View File

@ -0,0 +1,29 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from parakeet.modules import expansion
def test_expand():
x = paddle.randn([2, 4, 3]) # (B, T, C)
lengths = paddle.to_tensor([[1, 2, 2, 1], [3, 1, 4, 0]])
y = expansion.expand(x, lengths)
assert y.shape == [2, 8, 3]
print("the first sequence")
print(y[0])
print("the second sequence")
print(y[1])

34
tests/test_to_static.py Normal file
View File

@ -0,0 +1,34 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import paddle
from paddle.jit import to_static
from paddle.static import InputSpec
def test_applicative_evaluation():
def m_sqrt2(x):
return paddle.scale(x, math.sqrt(2))
subgraph = to_static(m_sqrt2, input_spec=[InputSpec([-1])])
paddle.jit.save(subgraph, './temp_test_to_static')
fn = paddle.jit.load('./temp_test_to_static')
x = paddle.arange(10, dtype=paddle.float32)
y = fn(x)
print(x)
print(y)