def evaluate_LFW(model,
embedding_size,
use_flipped_images=False,
N_folds=5,
distance_metric=1,
verbose=1):
pairs = read_pairs(os.path.expanduser(LFW_PAIRS_PATH))
paths, actual_issame = get_paths(os.path.expanduser(LFW_DIR), pairs)
ds = tf_dataset_from_paths(paths, flip=False)
embeddings = np.zeros([len(paths), embedding_size])
j = 0
if verbose >= 2:
print("Feed forward all pairs")
for batch in ds:
batch_embeddings = model(batch).numpy()
embeddings[j:j + len(batch)] = batch_embeddings
j += len(batch)
if use_flipped_images:
if verbose >= 2:
print("Feed forward all pairs - flipped")
flip_ds = tf_dataset_from_paths(paths, flip=True)
flip_embeddings = np.zeros([len(paths), embedding_size])
j = 0
for batch in flip_ds:
batch_embeddings = model(batch).numpy()
flip_embeddings[j:j + len(batch)] = batch_embeddings
j += len(batch)
full_embeddings = np.zeros((len(paths), embedding_size * 2))
full_embeddings[:, :embedding_size] = embeddings
full_embeddings[:, embedding_size:] = flip_embeddings
if verbose >= 2:
print("Calculating metrics")
if use_flipped_images:
tpr, fpr, accuracy, val, val_std, far, best_thresholds = evaluate(
(embeddings + flip_embeddings) / 2,
actual_issame,
nrof_folds=N_folds,
distance_metric=distance_metric)
else:
tpr, fpr, accuracy, val, val_std, far, best_thresholds = evaluate(
embeddings,
actual_issame,
nrof_folds=N_folds,
distance_metric=distance_metric)
if verbose:
print('Accuracy: %2.5f+-%2.5f' % (np.mean(accuracy), np.std(accuracy)))
print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' %
(val, val_std, far))
print('threshold : %2.5f+-%2.5f' %
(np.mean(best_thresholds), np.std(best_thresholds)))
auc = metrics.auc(fpr, tpr)
print('Area Under Curve (AUC): %1.3f' % auc)
eer = brentq(lambda x: 1. - x - interpolate.interp1d(fpr, tpr)(x), 0.,
1.)
print('Equal Error Rate (EER): %1.3f' % eer)
return accuracy
def gensim_classifier():
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s',
level=logging.INFO)
label_list = get_labels()
tweet_list = get_labelled_tweets()
# split all sentences to list of words
sentences = []
for tweet in tweet_list:
temp_doc = tweet.split()
sentences.append(temp_doc)
# parameters for model
num_features = 100
min_word_count = 1
num_workers = 4
context = 2
downsampling = 1e-3
# Initialize and train the model
w2v_model = Word2Vec(sentences, workers=num_workers, \
size=num_features, min_count = min_word_count, \
window = context, sample = downsampling, seed=1)
index_value, train_set, test_set = train_test_split(0.80, sentences)
train_vector = getAvgFeatureVecs(train_set, w2v_model, num_features)
test_vector = getAvgFeatureVecs(test_set, w2v_model, num_features)
train_vector = Imputer().fit_transform(train_vector)
test_vector = Imputer().fit_transform(test_vector)
# train model and predict
model = LinearSVC()
classifier_fitted = OneVsRestClassifier(model).fit(
train_vector, label_list[:index_value])
result = classifier_fitted.predict(test_vector)
# output result to csv
create_directory('data')
result.tofile("data/w2v_linsvc.csv", sep=',')
# store the model to mmap-able files
create_directory('model')
joblib.dump(model, 'model/%s.pkl' % 'w2v_linsvc')
# evaluation
label_score = classifier_fitted.decision_function(test_vector)
binarise_result = label_binarize(result, classes=class_list)
binarise_labels = label_binarize(label_list, classes=class_list)
evaluate(binarise_result, binarise_labels[index_value:], label_score,
'w2v_linsvc')
def gensim_classifier():
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
label_list = get_labels()
tweet_list = get_labelled_tweets()
# split all sentences to list of words
sentences = []
for tweet in tweet_list:
temp_doc = tweet.split()
sentences.append(temp_doc)
# parameters for model
num_features = 100
min_word_count = 1
num_workers = 4
context = 2
downsampling = 1e-3
# Initialize and train the model
w2v_model = Word2Vec(sentences, workers=num_workers, \
size=num_features, min_count = min_word_count, \
window = context, sample = downsampling, seed=1)
index_value, train_set, test_set = train_test_split(0.80, sentences)
train_vector = getAvgFeatureVecs(train_set, w2v_model, num_features)
test_vector = getAvgFeatureVecs(test_set, w2v_model, num_features)
train_vector = Imputer().fit_transform(train_vector)
test_vector = Imputer().fit_transform(test_vector)
# train model and predict
model = LinearSVC()
classifier_fitted = OneVsRestClassifier(model).fit(train_vector, label_list[:index_value])
result = classifier_fitted.predict(test_vector)
# output result to csv
create_directory('data')
result.tofile("data/w2v_linsvc.csv", sep=',')
# store the model to mmap-able files
create_directory('model')
joblib.dump(model, 'model/%s.pkl' % 'w2v_linsvc')
# evaluation
label_score = classifier_fitted.decision_function(test_vector)
binarise_result = label_binarize(result, classes=class_list)
binarise_labels = label_binarize(label_list, classes=class_list)
evaluate(binarise_result, binarise_labels[index_value:], label_score, 'w2v_linsvc')
def find_alphas(preprocess=None):
index, categories, categories_words, N_words = analyze_training_data(
'./data/')
results = {}
for alpha in [i / 10 for i in range(1, 51)]:
predictions, labels = naive_bayes_on_validation_set(
'./data/',
index,
categories,
categories_words,
N_words,
alpha=alpha,
preprocess=preprocess)
results[alpha] = evaluate(predictions, labels)
if preprocess is not None:
f = open(
'naive_bayes_with_different_alphas_' + str(preprocess) + '_.txt',
'w+')
else:
f = open('naive_bayes_with_different_alphas.txt', 'w+')
for alpha in results.keys():
print("##############################\nRESULTS FOR ALPHA={}\n".format(
alpha),
results[alpha],
file=f)
f.close()
def perform_on_specific_setting(self, method, k):
predictions, labels = self.knn_on_validation_set(k=k, method=method)
result = evaluate(predictions, labels)
f = open('knn_with_' + str(self.preprocess) + '.txt', 'w+')
print("##############################\nRESULTS FOR M={} K={}\n".format(
method, k),
result,
file=f)
f.close()
def search_best_n_est_and_max_depth(data_path):
results = {}
for num_estimators in [i for i in range(20, 200, 20)]:
for max_depth in [i for i in range(100, 500, 50)]:
predictions, labels = validation(data_path, num_estimators, max_depth)
results[(num_estimators, max_depth)] = evaluate(predictions, labels)
f = open('random_forest_with_different_parameters.txt', 'w+')
for k in results.keys():
print("##############################\nRESULTS FOR (N_EST, MAX_DEPTH)={}\n".format(k),
results[k], file=f)
f.close()
def search_best_c(data_path):
results = {}
for C in [i / 10 for i in range(1, 21)]:
predictions, labels = validation(data_path, C)
results[C] = evaluate(predictions, labels)
f = open('SVM_with_different_Cs.txt', 'w+')
for C in results.keys():
print("##############################\nRESULTS FOR C={}\n".format(C),
results[C],
file=f)
f.close()
def lin_svc():
label_list = get_labels()
tweet_list = get_labelled_tweets()
# vectorise using tf-idf
vectoriser = TfidfVectorizer(
min_df=3,
max_features=None,
strip_accents='unicode',
analyzer='word',
token_pattern=r'\w{1,}',
ngram_range=(1, 2),
use_idf=1,
smooth_idf=1,
sublinear_tf=1,
)
## do transformation into vector
fitted_vectoriser = vectoriser.fit(tweet_list)
vectorised_tweet_list = fitted_vectoriser.transform(tweet_list)
train_vector, test_vector, train_labels, test_labels = train_test_split(
vectorised_tweet_list, label_list, test_size=0.8, random_state=42)
# train model and predict
model = LinearSVC()
ovr_classifier = OneVsRestClassifier(model).fit(train_vector, train_labels)
result = ovr_classifier.predict(test_vector)
# output result to csv
create_directory('data')
save_to_csv("data/testset_labels.csv", test_labels)
result.tofile("data/tfidf_linsvc.csv", sep=',')
save_model(ovr_classifier, 'tfidf_linsvc')
save_vectoriser(fitted_vectoriser, 'tfidf_vectoriser')
# evaluation
label_score = ovr_classifier.decision_function(test_vector)
binarise_result = label_binarize(result, classes=class_list)
binarise_labels = label_binarize(test_labels, classes=class_list)
evaluate(binarise_result, binarise_labels, label_score, 'tfidf_linsvc')
def lin_svc():
label_list = get_labels()
tweet_list = get_labelled_tweets()
# vectorise using tf-idf
vectoriser = TfidfVectorizer(min_df=3,
max_features=None,
strip_accents='unicode',
analyzer='word',
token_pattern=r'\w{1,}',
ngram_range=(1, 2),
use_idf=1,
smooth_idf=1,
sublinear_tf=1,)
## do transformation into vector
fitted_vectoriser = vectoriser.fit(tweet_list)
vectorised_tweet_list = fitted_vectoriser.transform(tweet_list)
train_vector, test_vector, train_labels, test_labels = train_test_split(vectorised_tweet_list,
label_list,
test_size=0.8,
random_state=42)
# train model and predict
model = LinearSVC()
ovr_classifier = OneVsRestClassifier(model).fit(train_vector, train_labels)
result = ovr_classifier.predict(test_vector)
# output result to csv
create_directory('data')
save_to_csv("data/testset_labels.csv", test_labels)
result.tofile("data/tfidf_linsvc.csv", sep=',')
save_model(ovr_classifier, 'tfidf_linsvc')
save_vectoriser(fitted_vectoriser, 'tfidf_vectoriser')
# evaluation
label_score = ovr_classifier.decision_function(test_vector)
binarise_result = label_binarize(result, classes=class_list)
binarise_labels = label_binarize(test_labels, classes=class_list)
evaluate(binarise_result, binarise_labels, label_score, 'tfidf_linsvc')
def find_results_of_ks(self):
results = {m: {} for m in self.methods}
for m in self.methods:
for k in [1, 3, 5]:
t1 = time.time()
predictions, labels = self.knn_on_validation_set(k=k, method=m)
results[m][k] = evaluate(predictions, labels)
t2 = time.time()
print(t2 - t1, 'secs')
f = open('knn_with_different_ks.txt', 'w+')
for m in self.methods:
for k in results[m].keys():
print(
"##############################\nRESULTS FOR M={} K={}\n".
format(m, k),
results[m][k],
file=f)
f.close()