CODEWITHSUNDEEP
pythontensorflowkerasneural-network
Alex Zaitsev
Alex Zaitsev

Reputation: 2766

Tensorflow accuracy is 0.5

I'm working on an application that should predict interesting moments in audio files with a length of 10 seconds. I converted each 50ms of audio to a note, so each my record has 1 label (1,0 - interesting or not) and 200 note features. Then I created 200 train examples:

from __future__ import absolute_import, division, print_function, unicode_literals
from google.colab import drive
import functools
import tensorflow as tf
import tensorflow_datasets as tfds

from google.colab import drive
drive.mount('/content/gdrive')

def get_dataset(file_path):
  dataset = tf.data.experimental.make_csv_dataset(
      file_path,
      batch_size=12
      label_name='label',
      na_value='?',
      num_epochs=1,
      ignore_errors=False)
  return dataset

train = get_dataset('/content/gdrive/My Drive/myProject/train.csv')
test = get_dataset('/content/gdrive/My Drive/myProject/test.csv')

feature_columns = []

for number in range(200):
  feature_columns.append(tf.feature_column.numeric_column('note' + str(number + 1) ))

preprocessing_layer = tf.keras.layers.DenseFeatures(feature_columns)

model = tf.keras.Sequential([
    preprocessing_layer,
    tf.keras.layers.Dense(50, activation=tf.nn.relu),
    tf.keras.layers.Dense(2, activation=tf.nn.softmax)
])

model.compile(
    loss='binary_crossentropy',
    optimizer='adam',
    metrics=['accuracy'])

model.fit(train, epochs=20)

Then my model returns such output on 20 epoch:

17/17 [==============================] - 0s 7ms/step - loss: 0.6959 - acc: 0.5000

What am I doing wrong?

Upvotes: 0

Views: 2105

Answers (1)

KrisR89
KrisR89

Reputation: 1541

You are using tf.nn.softmax activation hence a single class is correct for each prediction, therefor you should be using categorical_crossentropy instead of binary_crossentropy as loss function. I don't know if this is your only problem, but that should at least solve one.

You could also use a single output class, with sigmoid activation, and binary_crossentropyloss. Instead of having two classes 1. "interesting" or 2. "not-interesting" (which is always the inverse of the other). Then you would be training with a single boolean "interesting" [1/0]

You can read more about softmax, sigmoid, binary_crossentropy and categorical_crossentropy here

Upvotes: 3

Related Questions