tensorfboard embeddings hangs with "Computing PCA"

Question

I'm trying to display my embeddings in tensorboard. When I open embeddings tab of tensorboard I get: "Computing PCA..." and tensorboard hangs infinitely.

Before that it does load my tensor of shape 200x128. It does find the metadata file too.

I tried that on TF versions 0.12 and 1.1 with the same result.

features = np.zeros(shape=(num_batches*batch_size, 128), dtype=float)
embedding_var = tf.Variable(features, name='feature_embedding')
config = projector.ProjectorConfig()
embedding = config.embeddings.add()
embedding.tensor_name = 'feature_embedding'
metadata_path = os.path.join(self.log_dir, 'metadata.tsv')
embedding.metadata_path = metadata_path

with tf.Session(config=self.config) as sess:
  tf.global_variables_initializer().run()
  restorer = tf.train.Saver()
  restorer.restore(sess, self.pretrained_model_path)

  with open(metadata_path, 'w') as f:

    for step in range(num_batches):
      batch_images, batch_labels = data.next()

        for label in batch_labels:
          f.write('%s\n' % label)

        feed_dict = {model.images: batch_images}
        features[step*batch_size : (step+1)*batch_size, :] = \ 
                    sess.run(model.features, feed_dict)

  sess.run(embedding_var.initializer)
  projector.visualize_embeddings(tf.summary.FileWriter(self.log_dir), config)

Answer 1

I don't know what was wrong in the code above, but I rewrote it in a different way (below), and it works. The difference is when and how the embedding_var is initialized.

I also made a gist to copy-paste code from out of this.

# a numpy array for embeddings and a list for labels
features = np.zeros(shape=(num_batches*self.batch_size, 128), dtype=float)
labels   = []   


# compute embeddings batch by batch
with tf.Session(config=self.config) as sess:
  tf.global_variables_initializer().run()
  restorer = tf.train.Saver()
  restorer.restore(sess, self.pretrained_model)

  for step in range(num_batches):
    batch_images, batch_labels = data.next()

    labels += batch_labels

    feed_dict = {model.images: batch_images}                     
    features[step*self.batch_size : (step+1)*self.batch_size, :] = \
                sess.run(model.features, feed_dict)


# write labels
metadata_path = os.path.join(self.log_dir, 'metadata.tsv')
with open(metadata_path, 'w') as f:
  for label in labels:
    f.write('%s\n' % label)


# write embeddings
with tf.Session(config=self.config) as sess:

  config = projector.ProjectorConfig()
  embedding = config.embeddings.add()
  embedding.tensor_name = 'feature_embedding'
  embedding.metadata_path = metadata_path

  embedding_var = tf.Variable(features, name='feature_embedding')
  sess.run(embedding_var.initializer)
  projector.visualize_embeddings(tf.summary.FileWriter(self.log_dir), config)                  

  saver = tf.train.Saver({"feature_embedding": embedding_var})
  saver.save(sess, os.path.join(self.log_dir, 'model_features'))