def eval_model():
if not tf.io.gfile.exists(data_conf.EVAL_DIR):
tf.io.gfile.makedirs(data_conf.EVAL_DIR)
util.save_config_values(data_conf, data_conf.EVAL_DIR + "/data_")
util.save_config_values(model_conf, data_conf.EVAL_DIR + "/model_")
filepath = data_conf.EVAL_RECORD_PATH + '/*'
filenames = glob.glob(filepath)
global_step = tf.compat.v1.train.get_or_create_global_step()
dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(get_single_sample)
batch_size = 1
dataset = dataset.padded_batch(batch_size,
padded_shapes=([None], [ANSWER_COUNT, None],
[None], (), [None,
None], ()))
iterator = tf.compat.v1.data.make_one_shot_iterator(dataset)
next_q, next_a, next_l, next_plot_ids, next_plots, next_q_types = iterator.get_next(
)
logits, word_atts, sent_atts, pl_d = predict_batch(
[next_q, next_a, next_plots], training=False)
next_q_types = tf.reshape(next_q_types, ())
probabs = model.compute_probabilities(logits=logits)
loss_example = model.compute_batch_mean_loss(logits, next_l,
model_conf.LOSS_FUNC)
accuracy_example = tf.reduce_mean(input_tensor=model.compute_accuracies(
logits=logits, labels=next_l, dim=1))
# do not restore embeddings in case the vocabulary size has changed
#to_restore = tf.contrib.slim.get_variables_to_restore(exclude=["embeddings"])
saver = tf.compat.v1.train.Saver()
summary_writer = tf.compat.v1.summary.FileWriter(data_conf.TRAIN_DIR)
step = 0
total_acc = 0.0
total_loss = 0.0
p_counts = 0
last_p = ''
type_counts = np.zeros(6, dtype=np.int32)
type_accs = np.zeros(6)
with tf.compat.v1.Session() as sess:
init_op = tf.group(tf.compat.v1.global_variables_initializer(),
tf.compat.v1.local_variables_initializer())
sess.run(init_op)
ckpt = tf.train.get_checkpoint_state(data_conf.TRAIN_DIR)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found')
coord = tf.train.Coordinator()
threads = tf.compat.v1.train.start_queue_runners(sess=sess,
coord=coord)
sess.run(set_embeddings_op, feed_dict={place: vectors})
try:
while not coord.should_stop():
loss_val, acc_val, probs_val, gs_val, q_type_val, q_val, labels_val, p_val, a_val, p_id_val, atts_val, \
sent_atts_val = sess.run([loss_example, accuracy_example, probabs, global_step, next_q_types, next_q,
next_l, next_plots, next_a, next_plot_ids, word_atts, sent_atts])
total_loss += loss_val
total_acc += acc_val
predicted_probabilities = probs_val[0]
sentence_attentions = sent_atts_val[0]
pred_index = np.argmax(probs_val[0])
labels = labels_val[0]
gold = np.argmax(labels)
filename = ''
q_s = ''
for index in q_val[0]:
word = (vocab[index])
q_s += (word + ' ')
filename += (word + '_')
filename += "?"
p_id = str(p_id_val[0].decode("utf-8"))
path = data_conf.EVAL_DIR + "/plots/" + p_id + "_" + str(
step) + "/" # + filename
# write attention heat-map
if (p_id != last_p and p_counts < data_conf.PLOT_SAMPLES_NUM):
# if True:
for i, a_att in enumerate(atts_val[0]):
# a_att = np.mean(a_att, 2)
qa_s = q_s + "? (acc: " + str(acc_val) + ")\n "
for index in a_val[0][i]:
word = vocab[index]
qa_s += (word + ' ')
filename += word + "_"
lv = " (label: " + str(int(
labels[i])) + " - prediction: " + (str(
"%.2f" %
(predicted_probabilities[i] * 100))) + "%)"
qa_s += lv
a_sents = []
y_labels = []
for j, att in enumerate(a_att):
a_s = []
y_labels.append(
str("%.2f" % (sent_atts_val[0][i][j] * 100)) +
"%")
for index in p_val[0][j]:
a_s.append(vocab[index])
a_sents.append(a_s)
util.plot_attention(np.array(a_att), np.array(a_sents),
qa_s, y_labels, path, filename)
last_p = p_id
p_counts += 1
print("Sample loss: " + str(loss_val))
print("Sample labels: " + str(labels))
print("Sample probabilities: " + str(predicted_probabilities))
print("Sample acc: " + str(acc_val))
util.print_predictions(data_conf.EVAL_DIR, step, gold,
predicted_probabilities, data_conf.MODE)
util.print_sentence_attentions(data_conf.EVAL_DIR, step,
sentence_attentions)
step += 1
print("Total acc: " + str(total_acc / step))
print("Local_step: " + str(step * batch_size))
print("Global_step: " + str(gs_val))
print("===========================================")
except tf.errors.OutOfRangeError:
summary = tf.compat.v1.Summary()
summary.value.add(tag='validation_loss',
simple_value=total_loss / step)
summary.value.add(tag='validation_accuracy',
simple_value=(total_acc / step))
summary_writer.add_summary(summary, gs_val)
keys = util.get_question_keys()
if data_conf.MODE == "val":
with open(data_conf.EVAL_DIR + "/val_accuracy.txt",
"a") as file:
file.write("global step: " + str(gs_val) +
" - total accuracy: " + str(total_acc / step) +
"- total loss: " + str(total_loss / step) +
"\n")
file.write("Types (name / count / correct / accuracy):\n")
for entry in zip(keys, type_counts, type_accs,
(type_accs / type_counts)):
file.write(str(entry) + "\n")
file.write(
"==================================================================="
+ "\n")
util.save_eval_score("global step: " + str(gs_val) +
" - acc : " + str(total_acc / step) +
" - total loss: " +
str(total_loss / step) + " - " +
data_conf.TRAIN_DIR + "_" +
str(gs_val))
finally:
coord.request_stop()
coord.join(threads)
def eval_model(model_type, attack_level, num_modified_words, percentage_attacked_samples):
print("evaluate")
print("%s white-box adversarial attack modifies %d words of %d%% of the instances: " % (
attack_level, num_modified_words, percentage_attacked_samples))
global model_conf
if model_type == "lstm":
import movieqa.conf_lstm as model_conf
else:
import movieqa.conf_cnn as model_conf
if not tf.io.gfile.exists(data_conf.EVAL_DIR):
tf.io.gfile.makedirs(data_conf.EVAL_DIR)
util.save_config_values(data_conf, data_conf.TRAIN_DIR + "/data")
util.save_config_values(model_conf, data_conf.TRAIN_DIR + "/model")
filepath = data_conf.EVAL_RECORD_PATH + '/*'
filenames = glob.glob(filepath)
print("Evaluating adversarial attack on %s" % filenames)
global_step = tf.contrib.framework.get_or_create_global_step()
dataset = tf.contrib.data.TFRecordDataset(filenames)
dataset = dataset.map(get_single_sample)
batch_size = 1
dataset = dataset.padded_batch(batch_size, padded_shapes=(
[None], [ANSWER_COUNT, None], [None], (), [None, None], ()))
iterator = tf.compat.v1.data.make_one_shot_iterator(dataset)
next_q, next_a, next_l, next_plot_ids, next_plots, next_q_types = iterator.get_next()
_, w_atts, s_atts, _ = predict_batch(model_type, [next_q, next_a, next_plots], training=False)
if attack_level == "sentence":
m_p = tf.compat.v1.py_func(remove_plot_sentence, [next_plots, s_atts, next_l], [tf.int64])[0]
elif attack_level == "word":
m_p = tf.compat.v1.py_func(modify_plot_sentence,
[next_plots, w_atts, s_atts, next_l, num_modified_words, percentage_attacked_samples],
[tf.int64])[0]
logits, atts, sent_atts, pl_d = predict_batch(model_type, [next_q, next_a, m_p], training=False)
next_q_types = tf.reshape(next_q_types, ())
probabs = model.compute_probabilities(logits=logits)
loss_example = model.compute_batch_mean_loss(logits, next_l, model_conf.LOSS_FUNC)
accuracy_example = tf.reduce_mean(input_tensor=model.compute_accuracies(logits=logits, labels=next_l, dim=1))
to_restore = tf.contrib.slim.get_variables_to_restore(exclude=["embeddings"])
saver = tf.compat.v1.train.Saver(to_restore)
summary_writer = tf.compat.v1.summary.FileWriter(data_conf.TRAIN_DIR)
step = 0
total_acc = 0.0
total_prec = 0.0
total_rank = 0.0
total_loss = 0.0
type_counts = np.zeros(6, dtype=np.int32)
type_accs = np.zeros(6)
max_sent_atts = {}
max_atts = {}
p_counts = 0
last_p = ''
with tf.compat.v1.Session() as sess:
init_op = tf.group(tf.compat.v1.global_variables_initializer(), tf.compat.v1.local_variables_initializer())
sess.run(init_op)
ckpt = tf.train.get_checkpoint_state(data_conf.TRAIN_DIR)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found')
_ = sess.run(set_embeddings_op, feed_dict={place: vectors})
coord = tf.train.Coordinator()
threads = tf.compat.v1.train.start_queue_runners(sess=sess, coord=coord)
try:
while not coord.should_stop():
loss_val, acc_val, probs_val, gs_val, q_type_val, q_val, atts_val, sent_atts_val, labels_val, p_val, a_val, p_id_val = sess.run(
[loss_example, accuracy_example, probabs, global_step, next_q_types, next_q, atts, sent_atts,
next_l,
pl_d, next_a, next_plot_ids])
type_accs[q_type_val + 1] += acc_val
type_counts[q_type_val + 1] += 1
predicted_probabilities = probs_val[0]
sentence_attentions = sent_atts_val[0]
total_loss += loss_val
total_acc += acc_val
pred_index = np.argmax(probs_val[0])
labels = labels_val[0]
gold = np.argmax(labels)
filename = ''
q_s = ''
for index in q_val[0]:
word = (vocab[index])
q_s += (word + ' ')
filename += (word + '_')
p_id = str(p_id_val[0].decode("utf-8"))
path = data_conf.EVAL_DIR + "/plots/" + p_id + "/" + filename
corr_ans = np.argmax(labels_val[0])
max_att_val = np.argmax(sent_atts_val[0][corr_ans])
att_row = np.max(atts_val[0][corr_ans][max_att_val], 1)
red = np.max(atts_val[0][corr_ans][max_att_val], 1)
att_inds = np.argsort(red)[::-1]
if (p_id != last_p and p_counts < 20):
for i, a_att in enumerate(atts_val[0]):
a_att = np.mean(a_att, 2)
qa_s = q_s + "? (acc: " + str(acc_val) + ")\n "
for index in a_val[0][i]:
qa_s += (vocab[index] + ' ')
lv = " (label: " + str(int(labels_val[0][i])) + " - prediction: " + (
str("%.2f" % (probs_val[0][i] * 100))) + "%)"
qa_s += lv
a_sents = []
y_labels = []
for j, att in enumerate(a_att):
a_s = []
y_labels.append(str("%.2f" % (sent_atts_val[0][i][j] * 100)) + "%")
for index in p_val[0][j]:
a_s.append(vocab[index])
a_sents.append(a_s)
# util.plot_attention(np.array(a_att), np.array(a_sents),qa_s,y_labels,path,filename)
last_p = p_id
p_counts += 1
m_ap = util.example_precision(probs_val[0], labels_val[0], 5)
rank = util.example_rank(probs_val[0], labels_val[0], 5)
total_prec += m_ap
total_rank += rank
print("Sample loss: " + str(loss_val))
print("Sample acc: " + str(acc_val))
print("Sample prec: " + str(m_ap))
print("Sample rank: " + str(rank))
util.print_predictions(data_conf.EVAL_DIR, step, gold, predicted_probabilities, data_conf.MODE)
util.print_sentence_attentions(data_conf.EVAL_DIR, step, sentence_attentions)
step += 1
print("Total acc: " + str(total_acc / step))
print("Total prec: " + str(total_prec / step))
print("Total rank: " + str(total_rank / step))
print("Local_step: " + str(step * batch_size))
print("Global_step: " + str(gs_val))
if attack_level == "word":
print("%d modified word(s)" % num_modified_words)
print("===========================================")
except tf.errors.OutOfRangeError:
summary = tf.compat.v1.Summary()
summary.value.add(tag='validation_loss', simple_value=total_loss / step)
summary.value.add(tag='validation_accuracy', simple_value=(total_acc / step))
summary_writer.add_summary(summary, gs_val)
keys = util.get_question_keys()
with open(data_conf.EVAL_DIR + "/accuracy.txt", "a") as file:
file.write("global step: " + str(gs_val) + " - total accuracy: " + str(
total_acc / step) + "- total loss: " + str(total_loss / step) + str(num_modified_words) + "" "\n")
file.write("Types (name / count / correct / accuracy):\n")
for entry in zip(keys, type_counts, type_accs, (type_accs / type_counts)):
file.write(str(entry) + "\n")
file.write("===================================================================" + "\n")
util.save_eval_score(
"global step: " + str(gs_val) + " - acc : " + str(
total_acc / step) + " - total loss: " + str(
total_loss / step) + " - " + data_conf.TRAIN_DIR + "_" + str(gs_val))
finally:
coord.request_stop()
coord.join(threads)
def eval_model():
if not tf.gfile.Exists(data_conf.EVAL_DIR):
tf.gfile.MakeDirs(data_conf.EVAL_DIR)
util.save_config_values(data_conf, data_conf.EVAL_DIR + "/data")
util.save_config_values(model_conf, data_conf.EVAL_DIR + "/model")
filepath = data_conf.EVAL_RECORD_PATH + '/*'
filenames = glob.glob(filepath)
print("Evaluate model on %s" % str(filenames))
global_step = tf.contrib.framework.get_or_create_global_step()
dataset = tf.contrib.data.TFRecordDataset(filenames)
dataset = dataset.map(get_single_sample)
batch_size = 1
dataset = dataset.padded_batch(batch_size,
padded_shapes=([None], [ANSWER_COUNT, None],
[None], (), [None], ()))
iterator = dataset.make_one_shot_iterator()
next_q, next_a, next_l, next_plot_ids, next_plots, next_q_types = iterator.get_next(
)
logits = predict_batch([next_q, next_a, next_plots], training=False)
next_q_types = tf.reshape(next_q_types, ())
probabs = model.compute_probabilities(logits=logits)
loss_example = model.compute_batch_mean_loss(logits, next_l,
model_conf.LOSS_FUNC)
accuracy_example = tf.reduce_mean(
model.compute_accuracies(logits=logits, labels=next_l, dim=1))
saver = tf.train.Saver()
summary_writer = tf.summary.FileWriter(data_conf.TRAIN_DIR)
step = 0
total_acc = 0.0
total_loss = 0.0
type_counts = np.zeros(6, dtype=np.int32)
type_accs = np.zeros(6)
with tf.Session() as sess:
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
ckpt = tf.train.get_checkpoint_state(data_conf.TRAIN_DIR)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found')
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
while not coord.should_stop():
loss_val, acc_val, probs_val, gs_val, q_type_val, labels_val = sess.run(
[
loss_example, accuracy_example, probabs, global_step,
next_q_types, next_l
])
predicted_probabilities = probs_val[0]
pred_index = np.argmax(probs_val[0])
labels = labels_val[0]
gold = np.argmax(labels)
type_accs[q_type_val + 1] += acc_val
type_counts[q_type_val + 1] += 1
total_loss += loss_val
total_acc += acc_val
print("Sample loss: " + str(loss_val))
print("Sample acc: " + str(acc_val))
util.print_predictions(data_conf.EVAL_DIR, step, gold,
predicted_probabilities, data_conf.MODE)
step += 1
print("Total acc: " + str(total_acc / step))
print("Local_step: " + str(step * batch_size))
print("Global_step: " + str(gs_val))
print("===========================================")
except tf.errors.OutOfRangeError:
summary = tf.Summary()
summary.value.add(tag='validation_loss',
simple_value=total_loss / step)
summary.value.add(tag='validation_accuracy',
simple_value=(total_acc / step))
summary_writer.add_summary(summary, gs_val)
keys = util.get_question_keys()
if data_conf.MODE == "val":
with open(data_conf.EVAL_DIR + "/val_accuracy.txt",
"a") as file:
file.write("global step: " + str(gs_val) +
" - total accuracy: " + str(total_acc / step) +
"- total loss: " + str(total_loss / step) +
"\n")
file.write("Types (name / count / correct / accuracy):\n")
for entry in zip(keys, type_counts, type_accs,
(type_accs / type_counts)):
file.write(str(entry) + "\n")
file.write(
"==================================================================="
+ "\n")
util.save_eval_score("global step: " + str(gs_val) +
" - acc : " + str(total_acc / step) +
" - total loss: " +
str(total_loss / step) + " - " +
data_conf.TRAIN_DIR + "_" +
str(gs_val))
finally:
coord.request_stop()
coord.join(threads)