Title

import pandas as pd

---

This article is also a Jupyter Notebook available to be run from the top down. There will be code snippets that you can then run in any environment.

Below are the versions of fastai, fastcore, wwf, fastaudio, and torchaudio currently running at the time of writing this:

• fastai : 2.1.5
• fastcore : 1.3.4
• wwf : 0.0.16
• fastaudio : 0.1.3
• torchaudio : 0.9.0

---

import pkg_resources
def placeholder(x):
    raise pkg_resources.DistributionNotFound
pkg_resources.get_distribution = placeholder
import fastai
import fastaudio
from fastai.vision.all import *
from fastaudio.core.all import *
from fastaudio.augment.all import *

data = pd.read_csv("comma labels.csv", index_col='path')

data

	pattern	kana	syl	drop
path
ある.yomi000142BB_0596.mp3	頭高	アル	2	1
ある.yomi000142BB_0596.mp3	頭高	アル	2	1
思う.yomi0006C617_043A.mp3	中高	オモウ	3	2
など.yomi000240B7_0028.mp3	頭高	ナド	2	1
私.yomi00092F63_0072.mp3	平板	ワタくシ	4	0
...	...	...	...	...
くも膜下出血_蜘蛛膜下出血.yomi0001AAD1_0622.mp3	中高	クモマッカしュッケツ	10	6
捜す.yomi00072507_0088.mp3	平板	サカ゚ス	4	0
捜し物.yomi000724FD_0424.mp3	平板	サカ゚シモノ	6	0
あこや貝_阿古屋貝.yomi00013767_0114.mp3	中高	アコヤカ゚イ	6	3
あこや貝_阿古屋貝.yomi00013767_0114.mp3	中高	アコヤカ゚イ	6	3

74191 rows × 4 columns

data.loc['１すくい_一掬い.yomi000120BB_01C8.mp3']

	pattern	kana	syl	drop
path
１すくい_一掬い.yomi000120BB_01C8.mp3	中高	ひトすクイ	5	2
１すくい_一掬い.yomi000120BB_01C8.mp3	中高	ひトすクイ	5	2

data.loc['出し抜く.yomi0004BDF8_06F2.mp3']

pattern      平板
kana       ダシヌク
syl           4
drop          0
Name: 出し抜く.yomi0004BDF8_06F2.mp3, dtype: object

data.loc['ある.yomi000142BB_0596.mp3']

	pattern	kana	syl	drop
path
ある.yomi000142BB_0596.mp3	頭高	アル	2	1
ある.yomi000142BB_0596.mp3	頭高	アル	2	1

from google.colab import drive
drive.mount('/content/drive')

Mounted at /content/drive

path = Path('/content/drive/MyDrive')

path = path / '1000sample'

at = AudioTensor.create(path.ls()[0])

at.show()

<matplotlib.axes._subplots.AxesSubplot at 0x7f7a5b38aed0>

cfg = AudioConfig.Voice()

aud2spec = AudioToSpec.from_cfg(cfg)

aud2spec(at).show()

<matplotlib.axes._subplots.AxesSubplot at 0x7f7a5b50f950>

crop2s = ResizeSignal(2000)

crop2s(at).show()

<matplotlib.axes._subplots.AxesSubplot at 0x7f7a5b50f5d0>

pipe = Pipeline([AudioTensor.create, crop2s, aud2spec])

for i in range(5):
  pipe(path.ls()[i]).show()

data.loc['ある.yomi000142BB_0596.mp3']['pattern']

'頭高'

data.reset_index()[data.reset_index().duplicated(keep=False)].to_csv('duplicates.csv')

data[data.index.duplicated(keep=False)].to_csv('path_dup.csv')

data.duplicated().sum()

15695

data.reset_index()

	path	pattern	kana	syl	drop
0	ある.yomi000142BB_0596.mp3	頭高	アル	2	1
1	ある.yomi000142BB_0596.mp3	頭高	アル	2	1
2	思う.yomi0006C617_043A.mp3	中高	オモウ	3	2
3	など.yomi000240B7_0028.mp3	頭高	ナド	2	1
4	私.yomi00092F63_0072.mp3	平板	ワタくシ	4	0
...	...	...	...	...	...
74186	くも膜下出血_蜘蛛膜下出血.yomi0001AAD1_0622.mp3	中高	クモマッカしュッケツ	10	6
74187	捜す.yomi00072507_0088.mp3	平板	サカ゚ス	4	0
74188	捜し物.yomi000724FD_0424.mp3	平板	サカ゚シモノ	6	0
74189	あこや貝_阿古屋貝.yomi00013767_0114.mp3	中高	アコヤカ゚イ	6	3
74190	あこや貝_阿古屋貝.yomi00013767_0114.mp3	中高	アコヤカ゚イ	6	3

74191 rows × 5 columns

data = data.reset_index()

removing complete duplicates

data = data[-data.duplicated()]

data.to_csv('labels complete duplicates removed.csv')

duplicate labels

data.duplicated(subset=['path'], keep=False).sum()

687

data = data[-data.duplicated(subset=['path'], keep=False)]

data.to_csv('labels_no_duplicates.csv')

sr = data[data.path == '捜す.yomi00072507_0088.mp3']['pattern'].values[0]

sr.values[0]

'平板'

item_tfms = [ResizeSignal(1000), aud2spec]

def get_label(fn):
  return data[data.path == fn]['pattern'].values[0]

audpitch = DataBlock(blocks=(AudioBlock, CategoryBlock),  
                 get_items=get_audio_files, 
                 splitter=RandomSplitter(),
                 item_tfms = [crop2s, aud2spec],
                 get_y=using_attr(get_label, 'name'))

for fn in path.ls():
  print(fn)
  try:
    get_label(fn.name)
  except: fn.unlink()

for fn in path.ls():
  print(fn)
  print(get_label(fn.name))

learn = Learner(dls, xresnet18(), CrossEntropyLossFlat(), metrics=accuracy)

n_c = dls.one_batch()[0].shape[1]; n_c

1

def alter_learner(learn, n_channels=1):
  "Adjust a `Learner`'s model to accept `1` channel"
  layer = learn.model[0][0]
  layer.in_channels=n_channels
  layer.weight = nn.Parameter(layer.weight[:,1,:,:].unsqueeze(1))
  learn.model[0][0] = layer

alter_learner(learn, n_c)

learn.lr_find()

/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at  /pytorch/c10/core/TensorImpl.h:1156.)
  return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

SuggestedLRs(lr_min=0.03019951581954956, lr_steep=0.00363078061491251)

learn.fit_one_cycle(5, 1e-3)

epoch	train_loss	valid_loss	accuracy	time
0	0.250854	0.228218	0.937931	00:12
1	0.246093	1.324811	0.544828	00:12
2	0.241305	0.378529	0.882759	00:12
3	0.232387	0.211727	0.951724	00:12
4	0.225047	0.219429	0.965517	00:12

learn.fit_one_cycle(5, 1e-4)

epoch	train_loss	valid_loss	accuracy	time
0	0.204818	0.208217	0.958621	00:12
1	0.204138	0.206059	0.951724	00:12
2	0.202360	0.217017	0.958621	00:12
3	0.198590	0.214045	0.958621	00:12
4	0.198357	0.220390	0.944828	00:12

train = dls.train

m = learn.model

y = dls[0]

names = [fn.name for fn in path.ls()]

datasample =  data[data.path.isin(names)]

datasample.groupby('pattern').count()

	path	kana	syl	drop
pattern
中高	229	229	229	229
尾高	18	18	18	18
平板	403	403	403	403
頭高	78	78	78	78

data.groupby('pattern').count()

	path	kana	syl	drop
pattern
中高	24444	24444	24444	24444
尾高	1461	1461	1461	1461
平板	30551	30551	30551	30551
頭高	10038	10038	10038	10038