def test_idf_build():
questions_train, vocabulary, idf = load_questions_from_file("train", q_limit['train'])
questions_validate, vocabulary, idf = load_questions_from_file("validate", q_limit['validate'])
questions_test, vocabulary, idf = load_questions_from_file("test", q_limit['test'])
X_train, y_train = load_data(train_data, globals.nn_features_file, globals.nn_labels_file)
#X_validate, y_validate = load_data(validate_data, globals.nn_features_file, globals.nn_labels_file)
#X_test, y_test = load_data(test_data, globals.nn_features_file, globals.nn_labels_file)
y_train_flatten = list(itertools.chain(*y_train))
build_idf([questions_train, questions_validate, questions_test])
samples = decompose_questions(questions_train)
print("Samples len: %d, y_train_flatten len: %d" % (len(samples), len(y_train_flatten)))
#samples = samples[:3]
#pprint(samples[:40])
e = 0
print("%s" % "wo".ljust(10, " ") + " " + "wo_idf".ljust(30, " ") + " " + "q_len".ljust(10, " ")
+ " " + "wo_answers".ljust(30, " ") + " " + "idf_avg".ljust(30, " ")
+ " " + "idf_norm".ljust(30, " ") + " " + "label")
for idx, sample in enumerate(samples):
print("%s, %s" % (str(sample[0]).ljust(10, " ") + " " + str(sample[1]).ljust(30, " ")
+ " " + str(sample[2]).ljust(10, " ") + " " + str(sample[3]).ljust(30, " ")
+ " " + str(sample[4]).ljust(30, " ") + " " + str(sample[5]).ljust(30, " ")
+ " ", y_train_flatten[e]))
e += 1
def test_idf_build():
questions_train, vocabulary, idf = load_questions_from_file(
"train", q_limit['train'])
questions_validate, vocabulary, idf = load_questions_from_file(
"validate", q_limit['validate'])
questions_test, vocabulary, idf = load_questions_from_file(
"test", q_limit['test'])
X_train, y_train = load_data(train_data, globals.nn_features_file,
globals.nn_labels_file)
#X_validate, y_validate = load_data(validate_data, globals.nn_features_file, globals.nn_labels_file)
#X_test, y_test = load_data(test_data, globals.nn_features_file, globals.nn_labels_file)
y_train_flatten = list(itertools.chain(*y_train))
build_idf([questions_train, questions_validate, questions_test])
samples = decompose_questions(questions_train)
print("Samples len: %d, y_train_flatten len: %d" %
(len(samples), len(y_train_flatten)))
#samples = samples[:3]
#pprint(samples[:40])
e = 0
print("%s" % "wo".ljust(10, " ") + " " + "wo_idf".ljust(30, " ") + " " +
"q_len".ljust(10, " ") + " " + "wo_answers".ljust(30, " ") + " " +
"idf_avg".ljust(30, " ") + " " + "idf_norm".ljust(30, " ") + " " +
"label")
for idx, sample in enumerate(samples):
print("%s, %s" %
(str(sample[0]).ljust(10, " ") + " " + str(sample[1]).ljust(
30, " ") + " " + str(sample[2]).ljust(10, " ") + " " +
str(sample[3]).ljust(30, " ") + " " +
str(sample[4]).ljust(30, " ") + " " +
str(sample[5]).ljust(30, " ") + " ", y_train_flatten[e]))
e += 1
def train_and_test(X_train, y_train, X_validate, y_validate, X_test, y_test):
model = cnn.get_cnn("regular")
y_train_flatted = list(itertools.chain(*y_train))
nb_batch = len(X_train) / batch_size + 1
best_f1 = 0.0
best_f1_index = 0
# Load questions and build idf for them
questions_train, v, idf = load_questions_from_file("train",
q_limit['train'])
questions_validate, v, idf = load_questions_from_file(
"validate", q_limit['validate'])
questions_test, v, idf = load_questions_from_file("test", q_limit['test'])
build_idf([questions_train, questions_validate, questions_test])
# Create samples for loaded questions
samples_train = decompose_questions(questions_train)
samples_validate = decompose_questions(questions_validate)
samples_test = decompose_questions(questions_test)
print("len of X_train: %d and samples_train: %d" %
(len(X_train), len(samples_train)))
for e in range(nb_epoch):
print("Epoch %d" % e)
progress_bar = generic_utils.Progbar(X_train.shape[0])
# For training NN, shuffle the data
X_train_shuffled, y_train_shuffled = shuffle_set(
X_train, y_train_flatted)
# Train for number of batches
for i in range(nb_batch):
train_loss, train_accuracy = model.train_on_batch(
X_train_shuffled[i * batch_size:(i + 1) * batch_size],
y_train_shuffled[i * batch_size:(i + 1) * batch_size],
accuracy=True)
progress_bar.add(batch_size,
values=[("train loss", train_loss),
("train accuracy:", train_accuracy)])
# Check the score on the validation data
#results_val = test_model(model, X_validate, y_validate)
#best_threshold = find_threshold(y_validate, results_val["y_predicted_scores"], results_val["y_predicted_scores"])
#precision_val, recall_val, f1_val = evaluate_with_threshold(y_validate, results_val["y_predicted_scores"],
# results_val["y_predicted_scores"],
# best_threshold)
# Check the score on the test data
#results_test = test_model(model, X_test, y_test)
#precision_test, recall_test, f1_test = evaluate_with_threshold(y_test, results_test["y_predicted_scores"],
# results_test["y_predicted_scores"],
# best_threshold)
#nn_string = "NN tests:\n" + "Threshold".ljust(40, ".") + " %.4f" + "\nOver validation set\n" \
# + "validation loss, validation acc".ljust(40, ".") + " %.4f %.4f\n" \
# + "precision, recall, f1".ljust(40, ".") + " %.4f %.4f %.4f\n" \
# + "Over test set\n" \
# + "test loss, test acc".ljust(40, ".") + " %.4f %.4f\n" \
# + "precision, recall, f1".ljust(40, ".") + " %.4f %.4f %.4f\n" \
#globals.logger.info(nn_string % (best_threshold, results_val['test_loss'], results_val['test_accuracy'],
# precision_val, recall_val, f1_val,
# results_test['test_loss'], results_test['test_accuracy'],
# precision_test, recall_test, f1_test))
# Now evaluate with logistic regression
# Get predictions from NN
predictions_train = model.predict(X_train)
predictions_validate = model.predict(X_validate)
predictions_test = model.predict(X_test)
# Evaluate on logistic regression
precision, recall, f1 = validate_on_lr(samples_train, samples_validate,
samples_test, predictions_train,
predictions_validate,
predictions_test, y_train,
y_validate, y_test)
lr_string = "LR tests:\n" + "Over test set\n" \
+ "precision, recall, f1".ljust(40, ".") + " %.4f %.4f %.4f\n"
globals.logger.info(lr_string % (precision, recall, f1))
if f1 >= best_f1:
best_f1 = f1
best_f1_index = e
globals.logger.info(
"Training done, best f1 on logistic regression is: %.4f for epoch nr: %d"
% (best_f1, best_f1_index))
def train_and_test(X_train, y_train, X_validate, y_validate, X_test, y_test):
model = cnn.get_cnn("regular")
y_train_flatted = list(itertools.chain(*y_train))
nb_batch = len(X_train)/batch_size + 1
best_f1 = 0.0
best_f1_index = 0
# Load questions and build idf for them
questions_train, v, idf = load_questions_from_file("train", q_limit['train'])
questions_validate, v, idf = load_questions_from_file("validate", q_limit['validate'])
questions_test, v, idf = load_questions_from_file("test", q_limit['test'])
build_idf([questions_train, questions_validate, questions_test])
# Create samples for loaded questions
samples_train = decompose_questions(questions_train)
samples_validate = decompose_questions(questions_validate)
samples_test = decompose_questions(questions_test)
print("len of X_train: %d and samples_train: %d" % (len(X_train), len(samples_train)))
for e in range(nb_epoch):
print("Epoch %d" % e)
progress_bar = generic_utils.Progbar(X_train.shape[0])
# For training NN, shuffle the data
X_train_shuffled, y_train_shuffled = shuffle_set(X_train, y_train_flatted)
# Train for number of batches
for i in range(nb_batch):
train_loss,train_accuracy = model.train_on_batch(X_train_shuffled[i*batch_size:(i+1)*batch_size],
y_train_shuffled[i*batch_size:(i+1)*batch_size],
accuracy=True)
progress_bar.add(batch_size, values=[("train loss", train_loss),("train accuracy:", train_accuracy)])
# Check the score on the validation data
#results_val = test_model(model, X_validate, y_validate)
#best_threshold = find_threshold(y_validate, results_val["y_predicted_scores"], results_val["y_predicted_scores"])
#precision_val, recall_val, f1_val = evaluate_with_threshold(y_validate, results_val["y_predicted_scores"],
# results_val["y_predicted_scores"],
# best_threshold)
# Check the score on the test data
#results_test = test_model(model, X_test, y_test)
#precision_test, recall_test, f1_test = evaluate_with_threshold(y_test, results_test["y_predicted_scores"],
# results_test["y_predicted_scores"],
# best_threshold)
#nn_string = "NN tests:\n" + "Threshold".ljust(40, ".") + " %.4f" + "\nOver validation set\n" \
# + "validation loss, validation acc".ljust(40, ".") + " %.4f %.4f\n" \
# + "precision, recall, f1".ljust(40, ".") + " %.4f %.4f %.4f\n" \
# + "Over test set\n" \
# + "test loss, test acc".ljust(40, ".") + " %.4f %.4f\n" \
# + "precision, recall, f1".ljust(40, ".") + " %.4f %.4f %.4f\n" \
#globals.logger.info(nn_string % (best_threshold, results_val['test_loss'], results_val['test_accuracy'],
# precision_val, recall_val, f1_val,
# results_test['test_loss'], results_test['test_accuracy'],
# precision_test, recall_test, f1_test))
# Now evaluate with logistic regression
# Get predictions from NN
predictions_train = model.predict(X_train)
predictions_validate = model.predict(X_validate)
predictions_test = model.predict(X_test)
# Evaluate on logistic regression
precision, recall, f1 = validate_on_lr(samples_train, samples_validate, samples_test,
predictions_train, predictions_validate, predictions_test,
y_train, y_validate, y_test)
lr_string = "LR tests:\n" + "Over test set\n" \
+ "precision, recall, f1".ljust(40, ".") + " %.4f %.4f %.4f\n"
globals.logger.info(lr_string % (precision, recall, f1))
if f1 >= best_f1:
best_f1 = f1
best_f1_index = e
globals.logger.info("Training done, best f1 on logistic regression is: %.4f for epoch nr: %d" %
(best_f1, best_f1_index))