CODEWITHSUNDEEP
pythontensorflow
MaMiFreak
MaMiFreak

Reputation: 789

Tensorflow, multi label accuracy calculation

I am working on a multi label problem and i am trying to determine the accuracy of my model.

My model:

NUM_CLASSES = 361

x = tf.placeholder(tf.float32, [None, IMAGE_PIXELS])
y_ = tf.placeholder(tf.float32, [None, NUM_CLASSES])

# create the network
pred = conv_net( x )

# loss
cost = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits( pred, y_) )

# train step
train_step   = tf.train.AdamOptimizer().minimize( cost )

i want to calculate the accuracy in two different ways
- % of all labels that are predicted correctly - % of images where ALL labels are predicted correctly

unfortunately i am only able to calculate the % of all labels that are predicted correctly.

I thought this code would calculate % of images where ALL labels are predicted correctly

correct_prediction = tf.equal( tf.round( pred ), tf.round( y_ ) )

accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

and this code % of all labels that are predicted correctly

pred_reshape = tf.reshape( pred, [ BATCH_SIZE * NUM_CLASSES, 1 ] )
y_reshape = tf.reshape( y_, [ BATCH_SIZE * NUM_CLASSES, 1 ] )

correct_prediction_all = tf.equal( tf.round( pred_reshape ), tf.round( y_reshape ) )

accuracy_all = tf.reduce_mean( tf.cast(correct_prediction_all, tf.float32 ) )

somehow the coherency of the labels belonging to one image is lost and i am not sure why.

Upvotes: 20

Views: 13167

Answers (2)

Alex-zhai
Alex-zhai

Reputation: 311

# to get the mean accuracy over all labels, prediction_tensor are scaled logits (i.e. with final sigmoid layer)
correct_prediction = tf.equal( tf.round( prediction_tensor ), tf.round( ground_truth_tensor ) )
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

# to get the mean accuracy where all labels need to be correct
all_labels_true = tf.reduce_min(tf.cast(correct_prediction, tf.float32), 1)
accuracy2 = tf.reduce_mean(all_labels_true)

reference: https://gist.github.com/sbrodehl/2120a95d57963a289cc23bcfb24bee1b

Upvotes: 1

Olivier Moindrot
Olivier Moindrot

Reputation: 28198

I believe the bug in your code is in: correct_prediction = tf.equal( tf.round( pred ), tf.round( y_ ) ).

pred should be unscaled logits (i.e. without a final sigmoid).

Here you want to compare the output of sigmoid(pred) and y_ (both in the interval [0, 1]) so you have to write:

correct_prediction = tf.equal(tf.round(tf.nn.sigmoid(pred)), tf.round(y_))

Then to compute:

  • Mean accuracy over all labels:
accuracy1 = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
  • Accuracy where all labels need to be correct:
all_labels_true = tf.reduce_min(tf.cast(correct_prediction), tf.float32), 1)
accuracy2 = tf.reduce_mean(all_labels_true)

Upvotes: 31

Related Questions