def test_input_fn(data_dir, params):
"""Test input function for the MNIST dataset.
Args:
data_dir: (string) path to the data directory
params: (Params) contains hyperparameters of the model (ex: `params.num_epochs`)
"""
dataset = mnist_dataset.test(data_dir)
dataset = dataset.batch(params.batch_size)
dataset = dataset.prefetch(
1) # make sure you always have one batch ready to serve
return dataset
def __predict__(args, params):
path = os.path.join(args.model_dir, "centroids_embeddings.npy")
centrois = np.load(path)
n_clusters = centrois.shape[1] # remember to get it before reshape
# print("centrois.shape=", centrois.shape)
centrois = centrois.reshape(-1, centrois.shape[2]) # flatten it for knn
# print("centrois.shape=", centrois.shape)
# Define the model
tf.logging.info("Creating the model...")
config = tf.estimator.RunConfig(
tf_random_seed=230,
model_dir=args.model_dir,
save_summary_steps=params.save_summary_steps)
estimator = tf.estimator.Estimator(model_fn, params=params, config=config)
# Compute embeddings on the test set
tf.logging.info("Predicting")
predictions = estimator.predict(
lambda: test_input_fn(args.data_dir, params))
embeddings = np.zeros((params.eval_size, params.embedding_size))
for i, p in enumerate(
predictions): # i:enumerate_id, p:{'embeddings':array(64)}
embeddings[i] = p['embeddings']
labels = np.zeros(shape=[params.num_labels * n_clusters], dtype=np.int)
for i in range(1, params.num_labels):
for j in range(n_clusters):
labels[n_clusters * i + j] = i
knn = KNeighborsClassifier(n_neighbors=3)
knn.fit(centrois, labels)
y_predicted = knn.predict(embeddings)
with tf.Session() as sess:
dataset = mnist_dataset.test(args.data_dir)
dataset = dataset.map(lambda img, lab: lab)
dataset = dataset.batch(params.eval_size)
labels_tensor = dataset.make_one_shot_iterator().get_next()
y_true = sess.run(labels_tensor)
print("Accuracy: " +
str(metrics.accuracy_score(y_true, y_predicted) * 100) + "%")
config = projector.ProjectorConfig()
embedding = config.embeddings.add()
embedding.tensor_name = embedding_var.name
# Specify where you find the sprite (we will create this later)
# Copy the embedding sprite image to the eval directory
shutil.copy2(args.sprite_filename, eval_dir)
embedding.sprite.image_path = pathlib.Path(args.sprite_filename).name
embedding.sprite.single_image_dim.extend(
[params.image_size, params.image_size])
# embedding.sprite.single_image_dim.extend([28, 28])
with tf.Session() as sess:
# TODO (@omoindrot): remove the hard-coded 10000
# Obtain the test labels
dataset = mnist_dataset.test(args.data_dir)
dataset = dataset.map(lambda img, lab: lab)
# dataset = dataset.batch(10000)
dataset = dataset.batch(params.eval_size)
labels_tensor = dataset.make_one_shot_iterator().get_next()
labels = sess.run(labels_tensor)
# Specify where you find the metadata
# Save the metadata file needed for Tensorboard projector
metadata_filename = "mnist_metadata.tsv"
with open(os.path.join(eval_dir, metadata_filename), 'w') as f:
for i in range(params.eval_size):
c = labels[i]
f.write('{}\n'.format(c))
embedding.metadata_path = metadata_filename