Save feature vectors for projection with Tensorboard

save_feature_vectors_for_tensorboard.py

import tensorflow as tf
from tensorflow.contrib.tensorboard.plugins import projector

LOG_DIR = "your-tensorboard-log-dir-here"
FEATURE_VECTORS = "your-feature-vectors-as-npy"
MAX_NUMBER_SAMPLES = 8191
METADATA_FILE = os.path.join(LOG_DIR, 'metadata.tsv')
CHECKPOINT_FILE = os.path.join(LOG_DIR, 'features.ckpt')

# Create metadata
# Can include class data in here if interested / have available
with open(METADATA_FILE, 'w+') as wrf:
    wrf.write("\n".join([str(a) for a,i in enumerate(image_files[:MAX_NUMBER_SAMPLES])]))

feature_vectors = np.load(FEATURE_VECTORS)

features = tf.Variable(feature_vectors[:MAX_NUMBER_SAMPLES], name='features')

# Write summaries for tensorboard
with tf.Session() as sess:
    saver = tf.train.Saver([features])

    sess.run(features.initializer)
    saver.save(sess, CHECKPOINT_FILE)

    config = projector.ProjectorConfig()
    embedding = config.embeddings.add()
    embedding.tensor_name = features.name
    embedding.metadata_path = METADATA_FILE

    # This adds the sprite images
    embedding.sprite.image_path = SPRITES_FILE
    embedding.sprite.single_image_dim.extend(IMAGE_SIZE)
    projector.visualize_embeddings(tf.summary.FileWriter(LOG_DIR), config)