跨时期的准确性和val_acc停滞

Question

我已经为二进制文本分类构建了CNN模型，我的模型如下。作为参考，embedding_matrix包含预训练的词向量权重。

model = Sequential()
model.add(Embedding(num_words, EMBEDDING_DIM, weights = 
[embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable = False)) #numbers match up to tokenizer
model.add(Dropout(0.3))
model.add(Conv1D(64, 5, activation='relu')) 
model.add(MaxPooling1D(pool_size=4))
model.add(Flatten())
model.add(Dense(1, activation='softmax'))

编译并评估模型。

print(model.summary())
model.compile(loss='binary_crossentropy', optimizer="adam", metrics=['acc'])
model.fit(data, np.array(HIT), validation_split = VALIDATION_SPLIT, epochs = 3, batch_size=20, verbose = 2)
score = model.evaluate(data2, np.array(HIT2), verbose = 2)
print("Test Loss = ", score[0])
print("Test Accuracy = ", score[1])

在各个时期都没有发生变化。模型摘要和培训日志发布在下面。

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_1 (Embedding)      (None, 300, 100)          1813200   
_________________________________________________________________
dropout_1 (Dropout)          (None, 300, 100)          0         
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 296, 64)           32064     
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 74, 64)            0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 4736)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 4737      
=================================================================
Total params: 1,850,001
Trainable params: 36,801
Non-trainable params: 1,813,200
_________________________________________________________________


Train on 3919 samples, validate on 560 samples
Epoch 1/5
 - 4s - loss: 9.5353 - acc: 0.4019 - val_loss: 10.1633 - val_acc: 0.3625
Epoch 2/5
 - 3s - loss: 9.5353 - acc: 0.4019 - val_loss: 10.1633 - val_acc: 0.3625
Epoch 3/5
 - 3s - loss: 9.5353 - acc: 0.4019 - val_loss: 10.1633 - val_acc: 0.3625
Epoch 4/5
 - 3s - loss: 9.5353 - acc: 0.4019 - val_loss: 10.1633 - val_acc: 0.3625
Epoch 5/5
 - 3s - loss: 9.5353 - acc: 0.4019 - val_loss: 10.1633 - val_acc: 0.3625

我尝试使用不同的优化器，更改批处理大小，并更改CNN超参数，尽管在各个时期损耗，acc，val_loss和val_acc仍然保持不变。但是，当在最后一层使用S型函数时，跨历元的准确性和损失都会提高。