Question

我已经训练了一个用于文本分类的模型（在这种情况下，每个输入实际上都是使用形状为（300，）的预训练单词嵌入将一个单词转换为数字）。问题是模型训练和评估都很好（或者正如我可能看到的那样），但是当使用预测函数时，无论我提供多少输入，所有预测都是相同的，换句话说，第一个预测元素的预测是正确的，但其他所有元素的返回值都相同。

    def define_model(vocab_size, max_length):
        model = Sequential()
        model.add(Embedding(vocab_size, 300, input_length=X.shape[1], weights=[embedding_matrix]))
        model.add(Conv1D(filters=32, kernel_size=8, activation='selu'))
        model.add(MaxPooling1D(pool_size=2))
        model.add(Flatten())
        model.add(Dense(10, activation='relu' ))
        model.add(Dense(1, activation='sigmoid' ))
        model. compile(loss='binary_crossentropy' , optimizer='adam' , metrics=['accuracy' ])
        model.summary()
        return model

    model = define_model(vocab_size, max_length)
    checkpoint = ModelCheckpoint("object.h5", monitor='val_acc', save_best_only=True, mode='max')
    model.fit(X_train, y_train, epochs=20, validation_data = (X_val, y_val), batch_size=256, callbacks= 
    [checkpoint])
    model.save('saved models/Keras/object.h5' )

    _________________________________________________________________
    Layer (type)                 Output Shape              Param #   
    =================================================================
    embedding_9 (Embedding)      (None, 300, 300)          3361200   
    _________________________________________________________________
    conv1d_6 (Conv1D)            (None, 293, 32)           76832     
    _________________________________________________________________
    max_pooling1d_6 (MaxPooling1 (None, 146, 32)           0         
    _________________________________________________________________
    flatten_6 (Flatten)          (None, 4672)              0         
    _________________________________________________________________
    dense_11 (Dense)             (None, 10)                46730     
    _________________________________________________________________
    dense_12 (Dense)             (None, 1)                 11        
    =================================================================
    Total params: 3,484,773
    Trainable params: 3,484,773
    Non-trainable params: 0
    _________________________________________________________________
    Train on 10102 samples, validate on 2526 samples
    Epoch 1/20
    10102/10102 [==============================] - 36s 4ms/step - loss: 0.6943 - acc: 0.5282 - val_loss: 
    0.6922 - val_acc: 0.5313
    Epoch 2/20
    10102/10102 [==============================] - 33s 3ms/step - loss: 0.6899 - acc: 0.5301 - val_loss: 
    0.6952 - val_acc: 0.5313
    Epoch 3/20
    10102/10102 [==============================] - 34s 3ms/step - loss: 0.6823 - acc: 0.5383 - val_loss: 
    0.7346 - val_acc: 0.3915
    Epoch 4/20
    10102/10102 [==============================] - 33s 3ms/step - loss: 0.6441 - acc: 0.6352 - val_loss: 
    0.7665 - val_acc: 0.3670
    Epoch 5/20
    10102/10102 [==============================] - 33s 3ms/step - loss: 0.5570 - acc: 0.6949 - val_loss: 
    1.3648 - val_acc: 0.2989
    Epoch 6/20
    10102/10102 [==============================] - 33s 3ms/step - loss: 0.4524 - acc: 0.7762 - val_loss: 
    1.1479 - val_acc: 0.3187
    Epoch 7/20
    10102/10102 [==============================] - 32s 3ms/step - loss: 0.3823 - acc: 0.7928 - val_loss: 
    1.5121 - val_acc: 0.3032
    Epoch 8/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.3224 - acc: 0.8039 - val_loss: 
    1.6736 - val_acc: 0.2977
    Epoch 9/20
    10102/10102 [==============================] - 36s 4ms/step - loss: 0.2839 - acc: 0.8130 - val_loss: 
    2.1001 - val_acc: 0.3064
    Epoch 10/20
    10102/10102 [==============================] - 33s 3ms/step - loss: 0.2705 - acc: 0.8132 - val_loss: 
    2.3334 - val_acc: 0.2977
    Epoch 11/20
    10102/10102 [==============================] - 34s 3ms/step - loss: 0.2618 - acc: 0.8159 - val_loss: 
    2.6393 - val_acc: 0.2985
    Epoch 12/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.2579 - acc: 0.8137 - val_loss: 
    2.7506 - val_acc: 0.2997
    Epoch 13/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.2573 - acc: 0.8164 - val_loss: 
    2.8572 - val_acc: 0.2989
    Epoch 14/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.2559 - acc: 0.8173 - val_loss: 
    2.9541 - val_acc: 0.3021
    Epoch 15/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.2544 - acc: 0.8170 - val_loss: 
    3.0950 - val_acc: 0.3040
    Epoch 16/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.2546 - acc: 0.8166 - val_loss: 
    3.0565 - val_acc: 0.3013
    Epoch 17/20
    10102/10102 [==============================] - 36s 4ms/step - loss: 0.2547 - acc: 0.8177 - val_loss: 
    3.1455 - val_acc: 0.3040
    Epoch 18/20
    10102/10102 [==============================] - 35s 3ms/step - loss: 0.2572 - acc: 0.8179 - val_loss: 
    2.9103 - val_acc: 0.3044
    Epoch 19/20
    10102/10102 [==============================] - 37s 4ms/step - loss: 0.2533 - acc: 0.8177 - val_loss: 
    3.3200 - val_acc: 0.3040 
    Epoch 20/20
    10102/10102 [==============================] - 36s 4ms/step - loss: 0.2553 - acc: 0.8176 - val_loss: 
    3.1079 - val_acc: 0.3100

    model = load_model('saved models/Keras/object.h5' )
    _, acc = model.evaluate(X_train, y_train)
    print('Train Accuracy: %.2f' % (acc*100))
    _, acc = model.evaluate(X_test, y_test)
    print('Test Accuracy: %.2f' % (acc*100))

    10102/10102 [==============================] - 13s 1ms/step
    Train Accuracy: 81.83
    3158/3158 [==============================] - 4s 1ms/step
    Test Accuracy: 31.19

预测：

predictions

该模型中的同一问题发生在我的其他模型中，该模型使用keras标记器对句子进行分类。

keras模型针对不同的输入预测相同的输出

0 个答案: