-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
294 lines (259 loc) · 12.4 KB
/
utils.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
import os
import copy
import torch
import numpy as np
import pandas as pd
from collections import OrderedDict
def encode(txt, max_length=200):
"""歌詞の1文字1文字をUnicodeに変換する関数"""
txt_list = []
for l in txt:
txt_line = [ord(x) for x in str(l).strip()]
txt_line = txt_line[:max_length]
txt_len = len(txt_line)
if txt_len < max_length:
txt_line += ([0] * (max_length - txt_len))
txt_list.append((txt_line))
return np.array(txt_list)
def remove_random_symbols(txt, remove_rate=0.5):
"""文字列txtから記号をランダムに除去する関数"""
letters, symbols, non_symbols = [],[],[]
for i,let in enumerate(str(txt).strip()):
letters.append(let)
if let in [' ','*',"'",'"','.', ',', '-','/',';']:
symbols.append(i)
else:
non_symbols.append(i)
letters, symbols, non_symbols = np.array(letters), np.array(symbols,dtype='int'), np.array(non_symbols,dtype='int')
selected = np.random.choice(symbols, size=int(len(symbols)*(1-remove_rate)), replace=False)
selected = np.concatenate([non_symbols, selected])
selected = np.sort(selected)
return ''.join(letters[selected])
class Trainer:
def __init__(self, n_epochs, batch_size, learning_rate, criterion, opt, gkf, groups, pretrain, device):
from sklearn.metrics import f1_score, accuracy_score
self.batch_size = batch_size
self.device = device
self.learning_rate = learning_rate
self.n_epochs = n_epochs
self.criterion = criterion
self.opt = opt
self.gkf = gkf
self.groups = groups
self.pretrain = pretrain
self.n_splits = gkf.get_n_splits()
self.bst_model, self.bst_score = dict(), dict()
def set_model(self, model):
self.network = model
self.state_dict = model.state_dict()
def reset_model(self):
model = copy.deepcopy(self.network)
model.load_state_dict(self.state_dict)
return model
def train(self, X, y, verbose=1):
for fold, (tr_idx, va_idx) in enumerate(self.gkf.split(X, y, groups=self.groups)):
# 学習データと評価用データに分割
print(f'-----Fold{fold+1}/{self.n_splits}-----')
X_tr, X_va = X[tr_idx], X[va_idx]
y_tr, y_va = y[tr_idx], y[va_idx]
n_iter = len(y_tr) // self.batch_size
model = self.reset_model().to(self.device)
optimizer = self.opt(model.parameters(), lr=self.learning_rate)
self.bst_model[fold] = None
self.bst_score[fold] = -np.inf
# 学習
for epoch in range(self.n_epochs):
model.train()
total_loss = 0.0
total_correct = 0
random_idx = np.random.permutation(len(y_tr))
for i in range(n_iter):
X_batch = torch.from_numpy(X_tr[random_idx[self.batch_size*i:self.batch_size*(i+1)]]).to(self.device)
y_batch = torch.from_numpy(y_tr[random_idx[self.batch_size*i:self.batch_size*(i+1)]]).to(self.device)
optimizer.zero_grad()
outputs = model(X_batch)
loss = self.criterion(outputs, y_batch)
loss.backward()
optimizer.step()
total_loss += loss.item()
_, predicted = outputs.max(dim=1)
total_correct += (predicted == y_batch).sum().item()
train_loss = total_loss / n_iter
train_acc = total_correct / (self.batch_size*n_iter)
# 評価
model.eval()
with torch.no_grad():
outputs = model(torch.from_numpy(X_va).to(self.device))
loss = self.criterion(outputs, torch.from_numpy(y_va).to(self.device))
valid_loss = loss.item()
_, predicted = outputs.max(dim=1)
valid_acc = accuracy_score(y_va, predicted.cpu().numpy())
valid_f1 = f1_score(y_va, predicted.cpu().numpy(), average='macro')
if (epoch+1) % verbose == 0:
print(f'Epoch[{epoch+1}/{self.n_epochs}], TrainLoss: {train_loss:.4f}, ValidLoss: {valid_loss:.4f}, TrainAcc: {train_acc*100:.4f}%, ValidAcc: {valid_acc*100:.4f}%, ValidF1: {valid_f1:.4f}')
# best更新処理
if valid_f1 > self.bst_score[fold]:
self.bst_model[fold] = copy.deepcopy(model)
self.bst_score[fold] = valid_f1
if self.pretrain:
break
def save_all(self, dirname):
os.makedirs(dirname, exist_ok=True)
for fold, model in self.bst_model.items():
filepath = os.path.join(dirname, f'model_fold{fold+1}.pth')
torch.save(model.state_dict(), filepath)
print(f'[Saved] score:{self.bst_score[fold]:.4f} @ {filepath}')
def save_clf(self, dirname):
os.makedirs(dirname, exist_ok=True)
for fold, model in self.bst_model.items():
filepath = os.path.join(dirname, f'classifier_fold{fold+1}.pth')
clf_s = OrderedDict(list(model.state_dict().items())[1:])
torch.save(clf_s, filepath)
print(f'[Saved] score:{self.bst_score[fold]:.4f} @ {filepath}')
def predict_one_block(txt, embed, classifier, device):
"""歌詞1ブロックに対する予測を行う関数
txt[string] : 歌詞1ブロック分
--> 各アーティストに対する確率値の配列(1次元)
"""
txt = txt.strip().replace('\n','*').replace('\u3000','*')
txt_enc = torch.from_numpy(encode([txt]))
txt_emb = embed(txt_enc.to(device))
probs = []
for artist,model_list in classifier.items():
probs_fold = []
for model in model_list:
with torch.no_grad():
model.eval()
output = model(txt_emb)
probs_fold.append(output.softmax(dim=1).cpu().numpy()[0])
probs.append(np.array(probs_fold).mean(axis=0))
prob = np.array(probs)[:,1]
return prob
def predict_some_block(txt_list, embed, classifier, device):
"""歌詞ブロックのリストに対する予測を行う関数
txt_list[list] : 歌詞1ブロックを1要素とするリスト
--> 各アーティストに対する確率値の配列。(txt_listの長さ, 全アーティスト数)の2次元配列
"""
txt_arr = []
for txt in txt_list:
txt = txt.replace(' ','')
txt = txt.strip().replace('\n','*').replace('\u3000','*')
txt_arr.append(txt)
txt_enc = torch.from_numpy(encode(txt_arr))
txt_emb = embed(txt_enc.to(device))
probs = []
for artist,model_list in classifier.items():
probs_fold = []
for model in model_list:
with torch.no_grad():
model.eval()
output = model(txt_emb)
probs_fold.append(output.softmax(dim=1).cpu().numpy())
probs.append(np.array(probs_fold).mean(axis=0))
probs = np.array(probs)[:,:,1].transpose(1,0)
return probs
def predict_whole_song(txt_list, embed, classifier, device):
"""1曲のリストに対する予測を行う関数
txt_list[list] : 1曲の歌詞を1要素とするリスト。ただし1曲の歌詞は'\n\n'によりブロックで分割されていること。
--> 各アーティストに対する確率値の配列。(txt_listの長さ, 全アーティスト数)の2次元配列
"""
prob_arr = []
for txt in txt_list:
prob_arr.append(predict_some_block(txt.split('\n\n'), embed, classifier, device).mean(axis=0))
return np.array(prob_arr)
def show_predict_one_block(prob, artists, sort, figsize=(10,3)):
"""予測結果を可視化する関数(歌詞ブロック版)"""
import matplotlib.pyplot as plt
if sort:
order = np.argsort(prob)[::-1]
prob = np.array(prob)[order]
artists = np.array(artists)[order]
plt.figure(figsize=figsize)
plt.bar(artists, prob, color='green', zorder=100)
for i,p in enumerate(prob):
plt.text(i, p+0.05, f'{int(p*100)}%', horizontalalignment='center', zorder=100)
plt.xticks(rotation=90)
plt.ylim(0,1)
plt.grid()
plt.show()
def show_predict_whole_song(prob_arr, artists, sort, raw_txt_arr, figsize=(10,7)):
"""予測結果を可視化する関数(歌詞全体版)"""
import matplotlib.pyplot as plt
fig,axs = plt.subplots(2,1, figsize=figsize)
# 予測値の平均
prob_mean = prob_arr.mean(axis=0)
if sort:
order = np.argsort(prob_mean)[::-1]
prob_mean = np.array(prob_mean)[order]
artists_sort = np.array(artists)[order]
axs[0].bar(artists_sort, prob_mean, color='green', zorder=100)
for i,p in enumerate(prob_mean):
axs[0].text(i, p+0.05, f'{int(p*100)}%', horizontalalignment='center', zorder=100)
axs[0].set_xticks(range(len(artists_sort)), labels=artists_sort, rotation=90)
axs[0].set_ylim(0,1)
axs[0].grid()
axs[0].set_title('平均値')
# 各ブロックの予測値
prob_norm = prob_arr / np.repeat(prob_arr.sum(axis=1).reshape(-1,1), prob_arr.shape[1], axis=1)
raw_txt_arr = [raw_txt[:5] for raw_txt in raw_txt_arr]
df = pd.DataFrame(prob_norm, columns=artists, index=raw_txt_arr)
df.plot.bar(stacked=True, cmap='tab20c', ax=axs[1], zorder=100)
axs[1].legend(bbox_to_anchor=(1.05, 1), loc='upper left')
axs[1].set_xticks(range(len(df.index)), labels=df.index, rotation=90)
axs[1].set_xlabel('')
axs[1].grid()
axs[1].legend(ncol=2, bbox_to_anchor=(1, 0.5), loc='center left', fontsize=10)
axs[1].set_title('各ブロックの予測値')
plt.tight_layout()
plt.show()
def wakachi_one_block(txt):
"""日本語テキストを分かち書きする(歌詞ブロック版)"""
from janome.tokenizer import Tokenizer
txt = txt.replace('\u3000','')
tokenizer = Tokenizer()
tokens = tokenizer.tokenize(txt)
words = [token.surface for token in tokens]
txt_split = []
line = []
for i,word in enumerate(words):
if word == '\n':
txt_split.append(' '.join(line))
line = []
else:
line.append(word)
txt_split.append(' '.join(line))
return '\n'.join(txt_split)
def wakachi_some_block(txt):
"""日本語テキストを分かち書きする(歌詞全体版)"""
wakachi_txt = ''
for block in txt.split('\n\n'):
wakachi_txt += wakachi_one_block(block) + '\n\n'
return wakachi_txt
def highlight(exp, wakachi_txt, artists, sort_by=False):
"""LIMEの結果をハイライト表示"""
highlighted_text = '<h2><span style="color: black; background-color: rgba(255,128,0,1);">ぽい</span> / <span style="color: black; background-color: rgba(0,128,255,1);">ぽくない</span>判定理由</h2>'
if sort_by is not False:
order = np.argsort(sort_by)[::-1]
artist_labels = np.arange(len(artists))[order]
else:
artist_labels = range(len(artists))
for label in artist_labels:
words = [word for word, weight in exp.as_list(label)]
weights = np.array([weight for word, weight in exp.as_list(label)])
weights_alpha = weights / np.abs(weights).max()
alpha_dict = {word:alpha for word,alpha in zip(words,weights_alpha)}
highlighted_text += f'<h3>「{artists[label]}」っぽさ</h3><div style="overflow:auto; max-height:300px; max-width:600px;border: 2px solid black; padding: 20px; box-sizing: border-box"><p style="line-height: 2;">'
for line in wakachi_txt.split('\n'):
for word in line.split(' '):
try:
alpha = alpha_dict[word]
if alpha > 0:
bg_color = (255,128,0,alpha)
else:
bg_color = (0,128,255,-alpha)
highlighted_text += f'<span style="color: black; background-color: rgba{bg_color};">{word}</span> '
except KeyError:
highlighted_text += word
highlighted_text += '<br>'
highlighted_text += '</p></div><br>'
return highlighted_text