def generate_data(corpus, window_size, V):
    maxlen = window_size*2
    for words in corpus:
        L = len(words)
        for index, word in enumerate(words):
            contexts = []
            labels   = []            
            s = index - window_size
            e = index + window_size + 1
            
            contexts.append([words[i] for i in range(s, e) if 0 <= i < L and i != index])
            labels.append(word)
            x = sequence.pad_sequences(contexts, maxlen=maxlen) # List[List[int]]
            y = np_utils.to_categorical(labels, V) # カテゴリカルデータにする
            yield (x, y)

model = Sequential()
model.add(Embedding(input_dim=V, output_dim=dim, input_length=window_size*2))
model.add(Lambda(lambda x: K.mean(x, axis=1), output_shape=(dim,))) # windowサイズを1に減らすために平均を取る
model.add(Dense(V, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adadelta')

for ite in range(10):
    loss = 0.
    for x, y in generate_data(corpus, window_size, V):
        loss += model.train_on_batch(x, y)
    print(ite, loss)

f = open('vectors.txt' ,'w')
f.write('{} {}\n'.format(V-1, dim))

vectors = cbow.get_weights()[0]
for word, i in tokenizer.word_index.items():
    str_vec = ' '.join(map(str, list(vectors[i, :])))
    f.write('{} {}\n'.format(word, str_vec))
f.close()