How to solve loss: nan & accuracy: 0.0000e+00 in a LSTM problem? Tensorflow 2.x

user_11558934 0
Viewed 1367

I'm working in a LSTM problem. I'm trying to predict MBTI (Myers-Briggs test) personality type based on text classification (there's 16 personality types).

I have a csv file, which was preprocessed: the stopwords were removed, it was lemmatized, tokenized, sequenced and padded. The file doesn't have any NaN values and the text sequence have only int numbers.

However, the problem is generated when trying to train the model I get:

loss: nan - accuracy: 0.0000e+00 - val_loss: nan - val_accuracy: 0.0000e+00

enter image description here

training model

enter image description here

As requested: how's the x, y data and label looks like with the results

print(validation_label_seq)
[[ 5]
 [10]
 [ 4]
 [ 4]
 [15]
 [12]
 [ 1]...]

print(validation_padded[0])
maxlen = 240
array([  23,  353,  147,  677,    1,    1,  409,   10,  845, 1530,    1,
        103,  107,  998,  117, 1389,   25,    1,   28, 1889,  165,    1,
       1520,   49,  718,   65,   55,   34,    0,    0,    0,    0,    0,
          0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,...], dtype=int32)

print(train_label_seq)
[[ 8]
 [ 9]
 [ 3]
 [ 7]
 [ 4]
 [10]
 [15]
 [11]...]

print(train_data_padded[0])
maxlen = 240
array([ 19, 301, 133, 302, 562, 133,  28, 563, 895, 896, 897, 118,  99,
       564, 397,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0...], dtype=int32)


results = model.evaluate(validation_padded, validation_label_seq)

test = validation_padded[10]
predict = model.predict_classes([test])
print(predict[1])

59/59 [==============================] - 0s 1ms/sample - loss: nan - accuracy: 0.0000e+00
[0]
/tensorflow-2.1.0/python3.6/tensorflow_core/python/keras/engine/sequential.py:342: RuntimeWarning: invalid value encountered in greater
  return (proba > 0.5).astype('int32')

print(predict)

array([[0],
       [0],
       ...
       [0],
       [0]], dtype=int32)

What I tried?

  • I already tried to change to different optimizers
  • Lower the batch size
  • Check for value errors in the dataframe and in the sequences (train and validation data).

Expected output: Maybe I'm building wrong the model, so I will explain which is the main idea. I would like to get one output or sixteen outputs, which determines the accuracy of your personality type.

1 output:
INTP: 89%

16 outputs:
ENTP: 5% | INTP: 81% | INTJ: 1% | ...

If you'll like to check, here is the code:mbti personality

Dataframe:mbti_df

Any suggestions to improve the question will be considered

text-classification tensorflow python
1 Answers

You are using softmax in the code as final output. And it is bunch of probability values and check with what you are comparing within this code. The label encoded targets. They are not matching and that's why it is giving 0 accuracy. I would suggest changing the softmax o/p to correct form so that comparison over accuracy metric give the correct result.

Example:

soft max output [0.2, 0.8] Output for other [0 , 1]

Then it will be mismatch and accuracy will suffer.

user_37961130
Related