validation accuracy not improving

No matter how many epochs I use or change learning rate, my validation accuracy only remains in 50’s. Im using 1 dropout layer right now and if I use 2 dropout layers, my max train accuracy is 40% with 59% validation accuracy. And currently with 1 dropout layer, here’s my results:

2527/2527 [==============================] - 26s 10ms/step - loss: 1.2076 - accuracy: 0.7944 - val_loss: 3.0905 - val_accuracy: 0.5822
Epoch 10/20
2527/2527 [==============================] - 26s 10ms/step - loss: 1.1592 - accuracy: 0.7991 - val_loss: 3.0318 - val_accuracy: 0.5864
Epoch 11/20
2527/2527 [==============================] - 26s 10ms/step - loss: 1.1143 - accuracy: 0.8034 - val_loss: 3.0511 - val_accuracy: 0.5866
Epoch 12/20
2527/2527 [==============================] - 26s 10ms/step - loss: 1.0686 - accuracy: 0.8079 - val_loss: 3.0169 - val_accuracy: 0.5872
Epoch 13/20
2527/2527 [==============================] - 31s 12ms/step - loss: 1.0251 - accuracy: 0.8126 - val_loss: 3.0173 - val_accuracy: 0.5895
Epoch 14/20
2527/2527 [==============================] - 26s 10ms/step - loss: 0.9824 - accuracy: 0.8165 - val_loss: 3.0013 - val_accuracy: 0.5917
Epoch 15/20
2527/2527 [==============================] - 26s 10ms/step - loss: 0.9417 - accuracy: 0.8216 - val_loss: 2.9909 - val_accuracy: 0.5938
Epoch 16/20
2527/2527 [==============================] - 26s 10ms/step - loss: 0.9000 - accuracy: 0.8264 - val_loss: 3.0269 - val_accuracy: 0.5943
Epoch 17/20
2527/2527 [==============================] - 26s 10ms/step - loss: 0.8584 - accuracy: 0.8332 - val_loss: 3.0011 - val_accuracy: 0.5934
Epoch 18/20
2527/2527 [==============================] - 26s 10ms/step - loss: 0.8172 - accuracy: 0.8378 - val_loss: 2.9918 - val_accuracy: 0.5949
Epoch 19/20
2527/2527 [==============================] - 26s 10ms/step - loss: 0.7796 - accuracy: 0.8445 - val_loss: 2.9974 - val_accuracy: 0.5929
Epoch 20/20
2527/2527 [==============================] - 25s 10ms/step - loss: 0.7407 - accuracy: 0.8502 - val_loss: 3.0005 - val_accuracy: 0.5907

Again max, it can reach is 59%. Here’s the graph obtained:

No matter how much changes I make, the validation accuracy reaches max 59%. Here’s my code:

BATCH_SIZE = 64
EPOCHS = 20
LSTM_NODES = 256
NUM_SENTENCES = 3000
MAX_SENTENCE_LENGTH = 50
MAX_NUM_WORDS = 5000
EMBEDDING_SIZE = 100

encoder_inputs_placeholder = Input(shape=(max_input_len,))
x = embedding_layer(encoder_inputs_placeholder)
encoder = LSTM(LSTM_NODES, return_state=True)

encoder_outputs, h, c = encoder(x)
encoder_states = [h, c]

decoder_inputs_placeholder = Input(shape=(max_out_len,))

decoder_embedding = Embedding(num_words_output, LSTM_NODES)
decoder_inputs_x = decoder_embedding(decoder_inputs_placeholder)

decoder_lstm = LSTM(LSTM_NODES, return_sequences=True, return_state=True)
decoder_outputs, _, _ = decoder_lstm(decoder_inputs_x, initial_state=encoder_states)

decoder_dropout1 = Dropout(0.2)
decoder_outputs = decoder_dropout1(decoder_outputs)

decoder_dense1 = Dense(num_words_output, activation='softmax')
decoder_outputs = decoder_dense1(decoder_outputs)

opt = tf.keras.optimizers.RMSprop()

model = Model([encoder_inputs_placeholder,
  decoder_inputs_placeholder],
  decoder_outputs)
model.compile(
    optimizer=opt,
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

history = model.fit(
    [encoder_input_sequences, decoder_input_sequences],
    decoder_targets_one_hot,
    batch_size=BATCH_SIZE,
    epochs=EPOCHS,
    validation_split=0.1,
)

Im very confused why only my training accuracy is updating, not the validation accuracy.

Here’s the model summary:

Model: "model_1"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 25)           0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            (None, 23)           0                                            
__________________________________________________________________________________________________
embedding_1 (Embedding)         (None, 25, 100)      299100      input_1[0][0]                    
__________________________________________________________________________________________________
embedding_2 (Embedding)         (None, 23, 256)      838144      input_2[0][0]                    
__________________________________________________________________________________________________
lstm_1 (LSTM)                   [(None, 256), (None, 365568      embedding_1[0][0]                
__________________________________________________________________________________________________
lstm_2 (LSTM)                   [(None, 23, 256), (N 525312      embedding_2[0][0]                
                                                                 lstm_1[0][1]                     
                                                                 lstm_1[0][2]                     
__________________________________________________________________________________________________
dropout_1 (Dropout)             (None, 23, 256)      0           lstm_2[0][0]                     
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 23, 3274)     841418      dropout_1[0][0]                  
==================================================================================================
Total params: 2,869,542
Trainable params: 2,869,542
Non-trainable params: 0
__________________________________________________________________________________________________
None

Answer

The size of the training dataset is less than 3K. While the amount of the trainable parameters is around 3 million. The answer to your question is classical overfitting – the model is so huge, that just remember the training subset instead of a generalization.

How to improve the current situation:

• try to generate or find more data;
• reduce the complexity of the model:
  • use pre-trained embedding (glove, fasttext, etc.)
  • reduce the number of the LSTM nodes;