[ new blog ]

new blog

https://jihyo-jeon.github.io/

Some TF version errors in [Text generation using a RNN with eager execution]

Some TF version errors in [Text generation using a RNN with eager execution]

Tensorflow RNN tutorials are basically based on the TF version 1.13 or higher.

I was following this one below.

https://www.tensorflow.org/tutorials/sequences/text_generation

• Error 1

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-70-7589cf238f4d> in <module>
----> 1 sampled_indices = tf.random.categorical(example_batch_predictions[0], num_samples=1)
      2 #sampled_indices = tf.random.multinomial(example_batch_predictions[0], num_samples=1)
      3 # use tf.random.multinomial if using TF 1.12 or lower.
      4 sampled_indices = tf.squeeze(sampled_indices,axis=-1).numpy()

AttributeError: module 'tensorflow._api.v1.random' has no attribute 'categorical'

Solution

sampled_indices = tf.random.multinomial(example_batch_predictions[0], num_samples=1)
# use tf.random.multinomial if using TF 1.12 or lower.

Error 2



---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-72-9ea97acf159a> in <module>
      4 # if using TF 1.12 or lower, use 'tf.keras.backend.sparse_categorical_crossentropy'
      5 
----> 6 example_batch_loss  = loss(target_example_batch, example_batch_predictions)
      7 print("Prediction shape: ", example_batch_predictions.shape, " # (batch_size, sequence_length, vocab_size)")
      8 print("scalar_loss:      ", example_batch_loss.numpy().mean())

<ipython-input-72-9ea97acf159a> in loss(labels, logits)
      1 def loss(labels, logits):
----> 2   return tf.keras.losses.sparse_categorical_crossentropy(labels, logits, from_logits=True)
      3   #return tf.keras.backend.sparse_categorical_crossentropy(labels, logits, from_logits=True)
      4 # if using TF 1.12 or lower, use 'tf.keras.backend.sparse_categorical_crossentropy'
      5 

TypeError: sparse_categorical_crossentropy() got an unexpected keyword argument 'from_logits'


Solution
return tf.keras.backend.sparse_categorical_crossentropy(labels, logits, from_logits=True)
# if using TF 1.12 or lower, use 'tf.keras.backend.sparse_categorical_crossentropy'