tf_embedding_projector.Example.py

#!/usr/bin/env python3

import tensorflow as tf   # tf.VERSION == 1.4.0
from tensorflow.contrib.tensorboard.plugins import projector

# ... write something here ...

dx, x_embedding = discriminator(x, is_training=True)
dz, z_embedding = discriminator(z, is_training=True)

embedding_tensor = tf.concat([x_embedding, z_embedding], axis=0)

embedding_var = tf.Variable(tf.zeors_like(embedding_tensor), name='embedding')
assign_embedding = tf.assign(embedding_var, embedding_tensor)

writer = tf.summary.FileWriter(save_folder)
# ===[ Create tf.summary
# ...

# ===[ Create projector
vis_config = projector.ProjectorConfig() 
for e in [embedding_var]:
    embedding = vis_config.embeddings.add()                     
    embedding.tensor_name = e.name
    # Link this tensor to its metadata file (e.g. labels).
    embedding.metadata_path = e.op.name+'.tsv'
    embedding.sprite.image_path = e.op.name+'.png'  
    embedding.sprite.single_image_dim.extend([img_len, img_len])
projector.visualize_embeddings(writer, vis_config)



# ===[ Create label tsv (metadata)
with open(os.path.join(save_folder, embedding_var.op.name+'.tsv'), 'w') as f:
    f.write("Index\tLabel\n")
    idx = 0
    for _ in range(x_embedding.shape.as_list()[0]):
        f.write("%d\t%d\n" % (idx,1))
        idx += 1
    for _ in range(z_embedding.shape.as_list()[0]):
        f.write("%d\t%d\n" % (idx,0))
        idx += 1


# ===[ Session
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver([embedding_var])
# ... processing ...
sess.run(assign_embedding)
_embedding_map_to_img = np.concatenate(sess.run([x,z]), axis=0)
save_visualization(_embedding_map_to_img, save_path=os.path.join(save_folder, embedding_var.op.name+'.png'))
saver.save(sess, save_path=os.path.join(save_folder, 'model.ckpt'), global_step=global_step)