def eval_perplexity(model, corpus, batch_size=10, time_size=35):
print('evaluating perplexity ...')
corpus_size = len(corpus)
total_loss = 0
max_iters = (corpus_size - 1) // (batch_size * time_size)
jump = (corpus_size - 1) // batch_size
for iters in range(max_iters):
xs = np.zeros((batch_size, time_size), dtype=np.int32)
ts = np.zeros((batch_size, time_size), dtype=np.int32)
time_offset = iters * time_size
offsets = [time_offset + (i * jump) for i in range(batch_size)]
for t in range(time_size):
for i, offset in enumerate(offsets):
xs[i, t] = corpus[(offset + t) % corpus_size]
ts[i, t] = corpus[(offset + t + 1) % corpus_size]
try:
loss = model.forward(xs, ts, train_flg=False)
except TypeError:
loss = model.forward(xs, ts)
total_loss += loss
sys.stdout.write('\r%d / %d' % (iters, max_iters))
sys.stdout.flush()
print('')
ppl = np.exp(total_loss / max_iters)
return ppl
def __init__(self, vocab_size=10000, wordvec_size=100, hidden_size=100):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
# 重みの初期化
embed_W = (rn(V, D) / 100).astype('f')
lstm_Wx = (rn(D, 4 * H) / np.sqrt(D)).astype('f')
lstm_Wh = (rn(H, 4 * H) / np.sqrt(H)).astype('f')
lstm_b = np.zeros(4 * H).astype('f')
affine_W = (rn(H, V) / np.sqrt(H)).astype('f')
affine_b = np.zeros(V).astype('f')
# レイヤの生成
self.layers = [
TimeEmbedding(embed_W),
TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True),
TimeAffine(affine_W, affine_b)
]
self.loss_layer = TimeSoftmaxWithLoss()
self.lstm_layer = self.layers[1]
# すべての重みと勾配をリストにまとめる
self.params, self.grads = [], []
for layer in self.layers:
self.params += layer.params
self.grads += layer.grads
def most_similar(query, word_to_id, id_to_word, word_matrix, top=5):
'''類似単語の検索
:param query: クエリ（テキスト）
:param word_to_id: 単語から単語IDへのディクショナリ
:param id_to_word: 単語IDから単語へのディクショナリ
:param word_matrix: 単語ベクトルをまとめた行列。各行に対応する単語のベクトルが格納されていることを想定する
:param top: 上位何位まで表示するか
'''
# クエリを取り出す
if query not in word_to_id:
print('%s is not found' % query)
return
print('\n[query] ' + query)
query_id = word_to_id[query]
query_vec = word_matrix[query_id]
# コサイン類似度の算出
vocab_size = len(id_to_word)
similarity = np.zeros(vocab_size)
for i in range(vocab_size):
similarity[i] = cos_similarity(word_matrix[i], query_vec)
# コサイン類似度の結果から、その値を高い順に出力
count = 0
for i in (-1 * similarity).argsort():
if id_to_word[i] == query:
continue
print(' %s: %s' % (id_to_word[i], similarity[i]))
count += 1
if count >= top:
return
def create_co_matrix(corpus, vocab_size, window_size=1):
'''共起行列の作成
:param corpus: コーパス（単語IDのリスト）
:param vocab_size:語彙数
:param window_size:ウィンドウサイズ（ウィンドウサイズが1のときは、単語の左右1単語がコンテキスト）
:return: 共起行列
'''
corpus_size = len(corpus)
co_matrix = np.zeros((vocab_size, vocab_size), dtype=np.int32)
for idx, word_id in enumerate(corpus):
for i in range(1, window_size + 1):
left_idx = idx - i
right_idx = idx + i
if left_idx >= 0:
left_word_id = corpus[left_idx]
co_matrix[word_id, left_word_id] += 1
if right_idx < corpus_size:
right_word_id = corpus[right_idx]
co_matrix[word_id, right_word_id] += 1
return co_matrix
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
embed_W = (rn(V, D) / 100).astype('f')
lstm_Wx = (rn(D, 4 * H) / np.sqrt(D)).astype('f')
lstm_Wh = (rn(H, 4 * H) / np.sqrt(H)).astype('f')
lstm_b = np.zeros(4 * H).astype('f')
affine_W = (rn(H, V) / np.sqrt(H)).astype('f')
affine_b = np.zeros(V).astype('f')
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True)
self.affine = TimeAffine(affine_W, affine_b)
self.params, self.grads = [], []
for layer in (self.embed, self.lstm, self.affine):
self.params += layer.params
self.grads += layer.grads
def forward(self, xs):
Wx, Wh, b = self.params
N, T, _ = xs.shape
H = Wh.shape[0]
self.layers = []
hs = np.empty((N, T, H), dtype='f')
if (not self.stateful) or (self.h is None):
self.h = np.zeros((N, H), dtype='f')
if (not self.stateful) or (self.c is None):
self.c = np.zeros((N, H), dtype='f')
for t in range(T):
layer = LSTM(Wx, Wh, b)
self.h, self.c = layer.forward(xs[:, t, :], self.h, self.c)
hs[:, t, :] = self.h
self.layers.append(layer)
return hs
def convert_one_hot(corpus, vocab_size):
'''one-hot表現への変換
:param corpus: 単語IDのリスト（1次元もしくは2次元のNumPy配列）
:param vocab_size: 語彙数
:return: one-hot表現（2次元もしくは3次元のNumPy配列）
'''
N = corpus.shape[0]
if corpus.ndim == 1:
one_hot = np.zeros((N, vocab_size), dtype=np.int32)
for idx, word_id in enumerate(corpus):
one_hot[idx, word_id] = 1
elif corpus.ndim == 2:
C = corpus.shape[1]
one_hot = np.zeros((N, C, vocab_size), dtype=np.int32)
for idx_0, word_ids in enumerate(corpus):
for idx_1, word_id in enumerate(word_ids):
one_hot[idx_0, idx_1, word_id] = 1
return one_hot
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
embed_W = (rn(V, D) / 100).astype('f')
lstm_Wx = (rn(D, 4 * H) / np.sqrt(D)).astype('f')
lstm_Wh = (rn(H, 4 * H) / np.sqrt(H)).astype('f')
lstm_b = np.zeros(4 * H).astype('f')
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=False)
self.params = self.embed.params + self.lstm.params
self.grads = self.embed.grads + self.lstm.grads
self.hs = None
def forward(self, h, target):
batch_size = target.shape[0]
negative_sample = self.sampler.get_negative_sample(target)
# 正例のフォワード
score = self.embed_dot_layers[0].forward(h, target)
correct_label = np.ones(batch_size, dtype=np.int32)
loss = self.loss_layers[0].forward(score, correct_label)
# 負例のフォワード
negative_label = np.zeros(batch_size, dtype=np.int32)
for i in range(self.sample_size):
negative_target = negative_sample[:, i]
score = self.embed_dot_layers[1+i].forward(h, negative_target)
loss += self.loss_layers[1+i].forward(score, negative_label)
return loss
def __init__(self, corpus, power, sample_size):
self.sample_size = sample_size
self.vocab_size = None
self.word_p = None
counts = collections.Counter()
for word_id in corpus:
counts[word_id] += 1
vocab_size = len(counts)
self.vocab_size = vocab_size
self.word_p = np.zeros(vocab_size)
for i in range(vocab_size):
self.word_p[i] = counts[i]
self.word_p = np.power(self.word_p, power)
self.word_p /= np.sum(self.word_p)
def get_negative_sample(self, target):
batch_size = target.shape[0]
if not GPU:
negative_sample = np.zeros((batch_size, self.sample_size), dtype=np.int32)
for i in range(batch_size):
p = self.word_p.copy()
target_idx = target[i]
p[target_idx] = 0
p /= p.sum()
negative_sample[i, :] = np.random.choice(self.vocab_size, size=self.sample_size, replace=False, p=p)
else:
# GPU(cupy）で計算するときは、速度を優先
# 負例にターゲットが含まれるケースがある
negative_sample = np.random.choice(self.vocab_size, size=(batch_size, self.sample_size),
replace=True, p=self.word_p)
return negative_sample