我已经构建了一个基于聊天机器人的 seq2seq。我使用的 coupus 是来自https://github.com/Conchylicultor/DeepQA/tree/master/data/cornell的电影对话 我用来训练我的模型的大约 20000 个语料库。在 300 个 epoch 之后，损失约为 0.02。但最后当我输入一个随机问题时，比如“你要去哪里？” 或“你叫什么名字”或其他什么，我得到了相同的答案“它”。如您所见，无论我输入什么，我总是得到一个单词“It”。我发现当我使用 np.argmax 计算预测的概率分布时，每次我得到相同的索引“4”，这意味着接下来的单词' 指数。
我还发现来自 encoder_model 预测的 state_h 和 state_c 有一些非正规数据。例如。来自状态 c 的最大概率 > 16。
embed_layer = Embedding(input_dim=vocab_size, output_dim=50, trainable=False)
embed_layer.build((None,))
embed_layer.set_weights([embedding_matrix])
LSTM_cell = LSTM(300, return_state=True)
LSTM_decoder = LSTM(300, return_sequences=True, return_state=True)
dense = TimeDistributed(Dense(vocab_size, activation='softmax'))
#encoder输入 与 decoder输入
input_context = Input(shape=(maxLen, ), dtype='int32', name='input_context')
input_target = Input(shape=(maxLen, ), dtype='int32', name='input_target')
input_context_embed = embed_layer(input_context)
input_target_embed = embed_layer(input_target)
_, context_h, context_c = LSTM_cell(input_context_embed)
decoder_lstm, _, _ = LSTM_decoder(input_target_embed,
initial_state=[context_h, context_c])
output = dense(decoder_lstm)
model = Model([input_context, input_target], output)
model.compile(optimizer='adam', loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
model.fit([context_, final_target_], outs, epochs=1, batch_size=128, validation_split=0.2)
input_context = Input(shape=(maxLen,), dtype='int32', name='input_context')
input_target = Input(shape=(maxLen,), dtype='int32', name='input_target')
input_ctx_embed = embed_layer(input_context)
input_tar_embed = embed_layer(input_target)
_, context_h, context_c = LSTM_cell(input_ctx_embed)
decoder_lstm, _, _ = LSTM_decoder(input_tar_embed,
initial_state=[context_h, context_c])
output = dense(decoder_lstm)
context_model = Model(input_context, [context_h, context_c])
target_h = Input(shape=(300,))
target_c = Input(shape=(300,))
target, h, c = LSTM_decoder(input_tar_embed, initial_state=[target_h, target_c])
output = dense(target)
target_model = Model([input_target, target_h, target_c], [output, h, c])
maxlen = 12
with open('reverse_dictionary.pkl', 'rb') as f:
index_to_word = pickle.load(f)
question = "what is your name?"
# question = "where are you going?"
print(question)
a = question.split()
for pos, i in enumerate(a):
a[pos] = re.sub('[^a-zA-Z0-9 .,?!]', '', i)
a[pos]= re.sub(' +', ' ', i)
a[pos] = re.sub('([\w]+)([,;.?!#&\'\"-]+)([\w]+)?', r'\1 \2 \3', i)
if len(i.split()) > maxlen:
a[pos] = (' ').join(a[pos].split()[:maxlen])
if '.' in a[pos]:
ind = a[pos].index('.')
a[pos] = a[pos][:ind+1]
if '?' in a[pos]:
ind = a[pos].index('?')
a[pos] = a[pos][:ind+1]
if '!' in a[pos]:
ind = a[pos].index('!')
a[pos] = a[pos][:ind+1]
question = ' '.join(a).split()
print(question)
question = np.array([word_to_index[w] for w in question])
question = sequence.pad_sequences([question], maxlen=maxLen)
# padding='post', truncating='post')
print(question)
question_h, question_c = context_model.predict(question)
answer = np.zeros([1, maxLen])
answer[0, -1] = word_to_index['BOS']
'''
i keeps track of the length of the generated answer.
This won't allow the model to genrate sequences with more than 20 words.
'''
i=1
answer_ = []
flag = 0
while flag != 1:
prediction, prediction_h, prediction_c = target_model.predict([
answer, question_h, question_c
])
# print(prediction[0,-1,4])
word_arg = np.argmax(prediction[0, -1, :]) #
# print(word_arg)
answer_.append(index_to_word[word_arg])
if word_arg == word_to_index['EOS'] or i > 20:
flag = 1
answer = np.zeros([1, maxLen])
answer[0, -1] = word_arg
question_h = prediction_h
question_c = prediction_c
i += 1
print(' '.join(answer_))
我的输入：你叫什么名字？['什么'，'是'，'你的'，'名字'，'？'] [[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 218 85 20 206 22]]
我得到了什么：它