import tensorflow as tf

import numpy as np

import utils

# Data loading parameters

tf.flags.DEFINE_string("testing_embeddings_dir", "./data/embeddings_test_eth/", "Path to the embeddings used for testing")

tf.flags.DEFINE_string("testing_stories", "./data/test_nlu18.csv", "Path to the file with the stories.")

# Model parameters

tf.flags.DEFINE_integer("embedding_dim", 4800, "The dimension of the embeddings")

# Testing parameters

tf.flags.DEFINE_string("checkpoint_dir", "./runs/1528468039/checkpoints", "Checkpoint directory from training run")

tf.flags.DEFINE_string("output_file", "./output.csv", "Csv file containing the results")

tf.flags.DEFINE_boolean("has_labels", False, "if has_labels => compute accuracy, if not dump output in file")

FLAGS = tf.flags.FLAGS

# load testing embeddings

all_testing_embeddings = utils.load_embeddings(FLAGS.testing_embeddings_dir,

FLAGS.embedding_dim)

# generate data

test_stories, test_true_endings, test_wrong_endings = utils.generate_data(all_testing_embeddings)

test_stories = np.concatenate((test_stories, test_stories), axis=0)

test_endings = np.concatenate((test_true_endings, test_wrong_endings), axis=0)

# construct test input

test_labels = [1] * len(test_true_endings) + [0] * len(test_wrong_endings)

## EVALUATION ##

checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)

graph = tf.Graph()

with graph.as_default():

session_conf = tf.ConfigProto(

allow_soft_placement=True,

log_device_placement=False)

sess = tf.Session(config=session_conf)

with sess.as_default():

# Load the saved meta graph and restore variables

saver = tf.train.import_meta_graph("{}.meta".format(checkpoint_file))

saver.restore(sess, checkpoint_file)

# Get the placeholders from the graph by name

stories_ph = graph.get_operation_by_name("stories").outputs[0]

endings_ph = graph.get_operation_by_name("endings").outputs[0]

# Tensor we want to evaluate

predictions_ph = graph.get_operation_by_name("accuracy/predictions").outputs[0]

probabilities_ph = graph.get_operation_by_name("softmax/probabilities").outputs[0]

if FLAGS.has_labels:

labels_ph = graph.get_operation_by_name("labels").outputs[0]

accuracy_ph = graph.get_operation_by_name("accuracy/accuracy").outputs[0]

predictions, accuracy, probabilities = sess.run([

predictions_ph, accuracy_ph, probabilities_ph

], {

stories_ph: test_stories,

endings_ph: test_endings,

labels_ph: test_labels

})

slice_index = int(len(probabilities)/2)

prob = np.concatenate((probabilities[slice_index:,:], probabilities[:slice_index,:]), axis = 1)

res = [int(prob[i][0] > prob[i][2]) for i in range(len(prob))]

accuracy = sum(res)/len(res)

print('Accuracy: ', accuracy)

else:

predictions, probabilities = sess.run([

predictions_ph, probabilities_ph

], {

stories_ph: test_stories,

endings_ph: test_endings

})

slice_index = int(len(probabilities) / 2)

prob = np.concatenate((probabilities[slice_index:, :], probabilities[:slice_index, :]), axis=1)

res = [int(prob[i][0] > prob[i][2]) for i in range(len(prob))]

res = [r + 1 for r in res]

all_testing_stories = utils.load_and_process_text_data(

FLAGS.testing_stories, for_testing=True,

is_labeled=FLAGS.has_labels)

utils.write_results_to_csv(FLAGS.output_file, all_testing_stories, res)