def main():
batch_size = 10
wordvec_size = 100
hidden_size = 100
time_size = 5
lr = 0.1
max_epoch = 100
corpus, word_to_id, id_to_word = ptb.load_data('train')
corpus_size = 1000
corpus = corpus[:1000]
vocab_size = int(max(corpus) + 1)
xs = corpus[:-1]
ts = corpus[1:]
data_size = len(xs)
print(f'corpus size: {corpus_size}, vocabulary size: {vocab_size}')
max_iters = data_size // (batch_size + time_size)
time_idx = 0
total_loss = 0
loss_count = 0
ppl_list = []
model = SimpleRnnlm(vocab_size, wordvec_size, hidden_size)
optimizer = SGD(lr)
jump = (corpus_size - 1) // batch_size
offsets = [i * jump for i in range(batch_size)]
for epoch in range(1, max_epoch + 1):
for iter_ in range(max_iters):
batch_x = np.empty((batch_size, time_size), dtype=int)
batch_t = np.empty((batch_size, time_size), dtype=int)
for t in range(time_size):
for i, offset in enumerate(offsets):
batch_x[i, t] = xs[(offset + time_idx) % data_size]
batch_t[i, t] = xs[(offset + time_idx) % data_size]
time_idx += 1
loss = model.forward(batch_x, batch_t)
model.backward()
optimizer.update(model.params, model.grads)
total_loss += loss
loss_count += 1
ppl = np.exp(total_loss / loss_count)
print(f'| epoch {epoch} | perplexity {ppl}')
ppl_list.append(float(ppl))
total_loss, loss_count = 0, 0
print('DONE')
vocab_size = int(max(corpus) + 1)
xs = corpus[:-1] # input
ts = corpus[1:] # output(answer)
data_size = len(xs)
print(f'Corpus Size: {corpus_size}, Number of Vocab: {vocab_size}')
# variables for learning
max_iters = data_size // (batch_size * time_size)
time_idx = 0
total_loss = 0
loss_count = 0
ppl_list = []
# init model
model = SimpleRnnlm(vocab_size, wordvec_size, hidden_size)
optimizer = SGD(lr)
# calculate start reading point of each sample of mini batch
jump = (corpus_size - 1) // batch_size
offsets = [i * jump for i in range(batch_size)]
for epoch in range(max_epoch):
for iter in range(max_iters):
# get mini batch
batch_x = np.empty((batch_size, time_size), dtype='i')
batch_t = np.empty((batch_size, time_size), dtype='i')
for t in range(time_size):
for i, offset in enumerate(offsets):
batch_x[i, t] = xs[(offset + time_idx) % data_size]
batch_t[i, t] = ts[(offset + time_idx) % data_size]
import sys, os
sys.path.append(os.pardir)
from common.optimizer import SGD
from common.trainer import RnnlmTrainer
from dataset import ptb
from simple_rnnlm import SimpleRnnlm
# ハイパーパラメータの設定
batch_size = 10
wordvec_size = 100
hidden_size = 100 # RNNの隠れ状態ベクトルの要素数
time_size = 5 # RNNを展開するサイズ
lr = 0.1
max_epoch = 100
# 学習データの読み込み
corpus, word_to_id, id_to_word = ptb.load_data('train')
corpus_size = 1000 # テスト用にデータセットを小さくする
corpus = corpus[:corpus_size]
vocab_size = int(max(corpus) + 1)
xs = corpus[:-1] # 入力
ts = corpus[1:] # 出力（教師ラベル）
# モデルの生成
model = SimpleRnnlm(vocab_size, wordvec_size, hidden_size)
optimizer = SGD(lr)
trainer = RnnlmTrainer(model, optimizer)
trainer.fit(xs, ts, max_epoch, batch_size, time_size)
trainer.plot()
def main():
# ハイパーパラメータの設定
batch_size = 10
wordvec_size = 100
hidden_size = 500
time_size = 5 # Truncated BPTTの展開する時間サイズ
lr = 0.1
max_epoch = 100
# 学習データセットの読み込み
corpus, word_to_id, id_to_word = ptb.load_data('train')
corpus_size =1000
corpus = corpus[:corpus_size]
vocab_size = int(max(corpus) + 1)
print("corpus: ", corpus)
print("corpus size: ", len(corpus))
# 入力
xs = corpus[:-1]
ts = corpus[1:] # 教師ラベル = 入力の次の単語ID
print("corpus size: %d, vocabulary size: %d" % (corpus_size, vocab_size))
data_size = len(xs)
print("data_size: ", data_size)
# 学習時に使用する変数
max_iters = data_size // (batch_size * time_size)
time_idx = 0
total_loss = 0
loss_count =0
ppl_list = []
# モデルの生成
model = SimpleRnnlm(vocab_size, wordvec_size, hidden_size)
optimizer = SGD(lr)
# (1) ミニバッチの各サンプルの読み込み開始位置を計算
jump = (corpus_size - 1) // batch_size # e.g. corpus_size=1000, batch_size=20 -> (1000-1)//20 -> 49
print("jump: ", jump)
offsets = [i * jump for i in range(batch_size)] # batch_size=20, jump=49 -> [0, 49, 98, ..., 931]
print("offsets: ", offsets)
# 学習
for epoch in range(max_epoch):
for iter in range(max_iters):
# (2) ミニバッチの取得
batch_x = np.empty((batch_size, time_size), dtype=np.int32)
batch_t = np.empty((batch_size, time_size), dtype=np.int32)
for t in range(time_size):
for i, offset in enumerate(offsets):
print(f"offset: {offset}, time_idx: {time_idx} -> {(offset + time_idx) % data_size}")
batch_x[i, t] = xs[(offset + time_idx) % data_size]
batch_t[i, t] = ts[(offset + time_idx) % data_size]
time_idx += 1 # 0リセットされないので、iterループ毎にcorpusのoffsetからのズレを変更している.
# 勾配を求め、パラメータを更新
loss = model.forward(batch_x, batch_t)
model.backward()
optimizer.update(model.params, model.grads)
total_loss += loss
loss_count += 1
# (3) エポック毎にパープレキシティの評価
ppl = np.exp(total_loss / loss_count)
print('| epoch %d | perplexity %.2f' % (epoch+1, ppl))
ppl_list.append(float(ppl))
total_loss, loss_count = 0, 0
# グラフの描画
x = np.arange(len(ppl_list))
plt.plot(x, ppl_list, label='train')
plt.xlabel('epochs')
plt.ylabel('perplexity')
plt.show()
