Question

尝试通过LSTM模型使用新闻和股票数据预测股票。型号代码：

filter_length1 = 3
filter_length2 = 5
dropout = 0.5
learning_rate = 0.0001
weights = initializers.TruncatedNormal(mean=0.0, stddev=0.1, seed=2)
nb_filter = 16
rnn_output_size = 128
hidden_dims = 128
wider = True
deeper = True

if wider == True:
    nb_filter *= 2
    rnn_output_size *= 2
    hidden_dims *= 2


def build_model():

    model1 = Sequential()

    model1.add(Embedding(nb_words, 
                         embedding_dim,
                         weights=[word_embedding_matrix], 
                         input_length=max_daily_length))
    model1.add(Dropout(dropout))

    model1.add(Convolution1D(filters = nb_filter, 
                             kernel_size = filter_length1, 
                             padding = 'same',
                            activation = 'relu'))
    model1.add(Dropout(dropout))

    if deeper == True:
        model1.add(Convolution1D(filters = nb_filter, 
                                 kernel_size = filter_length1, 
                                 padding = 'same',
                                activation = 'relu'))
        model1.add(Dropout(dropout))

    model1.add(LSTM(rnn_output_size, 
                   activation=None,
                   kernel_initializer=weights,
                   dropout = dropout))

    ####

    model2 = Sequential()

    model2.add(Embedding(nb_words, 
                         embedding_dim,
                         weights=[word_embedding_matrix], 
                         input_length=max_daily_length))
    model2.add(Dropout(dropout))


    model2.add(Convolution1D(filters = nb_filter, 
                             kernel_size = filter_length2, 
                             padding = 'same',
                             activation = 'relu'))
    model2.add(Dropout(dropout))

    if deeper == True:
        model2.add(Convolution1D(filters = nb_filter, 
                                 kernel_size = filter_length2, 
                                 padding = 'same',
                                 activation = 'relu'))
        model2.add(Dropout(dropout))

    model2.add(LSTM(rnn_output_size, 
                    activation=None,
                    kernel_initializer=weights,
                    dropout = dropout))

    ####

    model = Sequential()

    model = Add()([model1.output, model2.output])
    model = Dense(hidden_dims, kernel_initializer=weights)(model)
    model = Dropout(dropout)(model)
    if deeper == True:
        model = Dense(hidden_dims//2, kernel_initializer=weights)(model)
        model = Dropout(dropout)(model)
    model = Dense(1, kernel_initializer = weights, name='output')(model)
    new_model = Model([model1.input, model2.input], model)
    new_model.compile(loss='mean_squared_error', optimizer=Adam(lr=learning_rate, clipvalue=1.0))
    return new_model

使用模型的代码：

# Use grid search to help find a better model
for deeper in [False]:
    for wider in [True,False]:
        for learning_rate in [0.0001]:
            for dropout in [0.3, 0.5]:
                model = build_model()
                print()
                print("Current model: Deeper={}, Wider={}, LR={}, Dropout={}".format(
                    deeper,wider,learning_rate,dropout))
                print()
                save_best_weights = 'question_pairs_weights_deeper={}_wider={}_lr={}_dropout={}.h5'.format(
                    deeper,wider,learning_rate,dropout)

                callbacks = [ModelCheckpoint(save_best_weights, monitor='val_loss', save_best_only=True),
                             EarlyStopping(monitor='val_loss', patience=5, verbose=1, mode='auto'),
                             ReduceLROnPlateau(monitor='val_loss', factor=0.2, verbose=1, patience=3)]

                history = model.fit([x_train,x_train],
                                    y_train,
                                    batch_size=64,
                                    epochs=10,
                                    validation_split=0.15,
                                    verbose=True,
                                    shuffle=True,
                                    callbacks = callbacks)

Numerical Loss

Loss in Nan

尝试降低lr，调整clipvalue ..................................................... ................. ................................................... ......................................

一段时间后损失变得微弱

0 个答案: