def run(input_file, test_file, k):
clf = RandomForestClassifier(n_estimators=k)
df = preprocess_data(input_file)
X, label_dict, dict = extract_features(df)
r, c = X.shape
dft = preprocess_testdata(test_file)
Xt, yt = extract_testfeatures(dft, label_dict, dict)
clf.fit(X[:, 0:c - 1], X[:, c - 1])
z = clf.predict(Xt)
print(accuracy_score(yt, z))
def build_model_input():
print("Import class names")
cols = [
'Label', 'Latin Name', 'Common Name', 'Train Images',
'Validation Images'
]
info = pd.read_csv("monkey_labels.txt", names=cols, skiprows=1)
global LABELS
LABELS = info['Common Name']
util.display_image('training/n0/n0018.jpg')
print("Data augmentation/Preprocessing")
height, width, channels = 299, 299, 3
train_datagen = ImageDataGenerator(rescale=1. / 255)
train_generator = train_datagen.flow_from_directory(
TRAIN_DIR,
target_size=(height, width),
batch_size=BATCH_SIZE,
class_mode='categorical')
test_datagen = ImageDataGenerator(rescale=1. / 255)
test_generator = test_datagen.flow_from_directory(TEST_DIR,
target_size=(height,
width),
batch_size=BATCH_SIZE,
class_mode='categorical')
print("Import pretrained Inception module")
base_model = Xception(weights=INCEPTION_DIR,
include_top=False,
input_shape=(height, width, channels))
print("Extract features")
train_features, train_labels = util.extract_features(
1097, BATCH_SIZE, train_generator, base_model)
test_features, test_labels = util.extract_features(272, BATCH_SIZE,
test_generator,
base_model)
return train_features, train_labels, test_features, test_labels
def configure_df(static, dynamic, params):
# consolidate static/dynamic features, select features, scale values
master_df = consolidate_features(static, dynamic)
# configure clustering dataframe
features = util.extract_features(params['clustering_features'])
cluster_df = util.choose_features(master_df,
['dt_entropy', 'num_unique_words'])
# return master feature df and scaled cluster df
return master_df, scale_features(cluster_df)
def _extract_optic_flow_features(info_df: pd.DataFrame, dir_path: str,
pickle_path: str):
data = None
print('Extracting optic flow features!')
for index, row in info_df.iterrows():
for camera_no in [1, 2]:
file_path = os.path.join(
dir_path, row['Initials'], str(row['Session']),
f"{row['Session']}K{camera_no}play{row['Segment_No']}.txt")
current = pd.read_csv(file_path,
delim_whitespace=True,
header=None)
child, therapist = current.iloc[:, :3].copy(), current.iloc[:, 3:]
for p in range(2):
cur_person = current.iloc[:, p * 3:(p + 1) * 3].copy()
cur_person.columns = ['x', 'y', 'm']
cur_person['Key'] = row['Key']
cur_feats = util.extract_features(
cur_person[cur_person['x'] != -10000000],
columns=['x', 'y', 'm'])
cur_feats = cur_feats.reset_index()
cur_feats['CameraNo'], cur_feats['Person'] = camera_no, p
if data is None:
data = cur_feats.copy()
else:
if len(cur_feats) == 0:
# Fill all zeros
prev_feats.loc[:, 'x_mean':'m_var'] = 0
prev_feats['CameraNo'], prev_feats[
'Person'] = camera_no, p
prev_feats['Key'] = row['Key']
data = data.append(prev_feats, ignore_index=True)
else:
data = data.append(cur_feats, ignore_index=True)
prev_feats = cur_feats.copy()
data.to_pickle(pickle_path)
return data
if reverse_word_dict[word_id_list[idx]] not in kb_dict:
kb_dict[reverse_word_dict[word_id_list[idx]]] = [int(item) for item in
line.strip('\r\n').split(',')]
f.close()
kb_emb_dict = {}
with open(sys.argv[13], 'r') as f:
for idx, line in enumerate(f):
# key is word index in dictionary, value is embedding vector
kb_emb_dict[idx] = [float(item) for item in line.strip('\r\n').split(',')]
f.close()
new_feature_dict = {}
for k,v in word_dict.items():
features = extract_features(use_other_info, k, word_dict, word_emb_dict, pos_dict,
pos_emb_dict, parser_dict, parser_emb_dict, dict_desc_dict,
kb_dict, kb_emb_dict)
if features is not None:
new_feature_dict[k] = features
else:
sys.stderr.write("%s feature extraction failed" % k)
print(len(new_feature_dict))
counter = 0
real_counter = 0
top_1 = 0
top_3 = 0
for com_item in analogy_list:
counter += 1
flag = False
def main(args):
tokenizer = util.Tokenizer(tokenize_type=args.tok, lowercase=True)
train_toks = tokenizer.tokenize(open(args.train_file).read())
num_train_toks = int(args.train_fraction * len(train_toks))
print('-' * 79)
print('Using %d tokens for training (%g%% of %d)' %
(num_train_toks, 100 * args.train_fraction, len(train_toks)))
train_toks = train_toks[:int(args.train_fraction * len(train_toks))]
val_toks = tokenizer.tokenize(open(args.val_file).read())
num_val_toks = int(args.val_fraction * len(val_toks))
print('Using %d tokens for validation (%g%% of %d)' %
(num_val_toks, 100 * args.val_fraction, len(val_toks)))
val_toks = val_toks[:int(args.val_fraction * len(val_toks))]
train_ngram_counts = tokenizer.count_ngrams(train_toks)
# Get vocab and threshold.
print('Using vocab size %d (excluding UNK) (original %d)' % (min(
args.vocab, len(train_ngram_counts[0])), len(train_ngram_counts[0])))
vocab = [
tup[0] for tup, _ in train_ngram_counts[0].most_common(args.vocab)
]
train_toks = tokenizer.threshold(train_toks, vocab, args.unk)
val_toks = tokenizer.threshold(val_toks, vocab, args.unk)
if args.features == 'basic1':
feature_extractor = util.basic_features1
elif args.features == 'basic1suffix3':
feature_extractor = util.basic_features1_suffix3 # TODO: Implement
elif args.features == 'basic2':
feature_extractor = util.basic_features2
else:
raise ValueError('Unknown feature extractor type.')
# for feature_extractor in [util.basic_features1, util.basic_features2, util.basic_features1_suffix3, util.basic_features2_prefix3, util.basic_features2_suffix3]:
# We'll cheat and cache features for validation data to make things faster
# for this assignment. The correct thing to do here would be
#
#f2i, fcache, num_feats_cached, x2ys \
# util.extract_features(train_toks, feature_extractor)
#
f2i, fcache, num_feats_cached, x2ys \
= util.extract_features(train_toks + val_toks, feature_extractor)
print('%d feature types extracted' % len(f2i))
print('%d feature values cached for %d window types' %
(num_feats_cached, len(fcache)))
for seed in [82, 95, 11, 29, 49, 8, 42, 36, 71, 65]:
# The language model assumes a trucated vocab and a feature definition.
lm = util.LogLinearLanguageModel(args.model,
vocab,
args.unk,
feature_extractor,
f2i,
fcache,
x2ys,
init=args.init,
lr=args.lr,
check_interval=args.check_interval,
seed=seed)
# lm = util.LogLinearLanguageModel(args.model, vocab, args.unk,
# feature_extractor, f2i, fcache, x2ys,
# init=args.init, lr=args.lr,
# check_interval=args.check_interval,
# seed=args.seed)
if args.test:
# Load trained parameters
lm.load()
else:
# Estimate parameters.
lm.train(train_toks, val_toks, args.epochs)
val_ppl = lm.test(val_toks)
print('Optimized Perplexity: %f' % (val_ppl))
sample = open('results/1.8.txt', 'a')
print('%f,%f' % (seed, val_ppl), file=sample)
sample.close()
print('-' * 79)
for (i, f, w) in lm.topK_feats(args.K):
print('{:10d}: {:40s} ({:8.4f})'.format(i, f, w))
from sklearn.cross_validation import StratifiedKFold, cross_val_score
data_dir = '../../data'
cleaned_file = 'trauma_los_cleaned.csv'
extract_file = 'trauma_los_cleaned_extract.csv'
desired_headings = [
'sex', 'normalvitals', 'gcs1', 'iss8', 'age65', 'transfer', 'penetrating',
'mechcode', 'bodyregions', 'headany', 'faceany', 'neckany', 'chestany',
'abdoany', 'spineany', 'upperlimbany', 'lowerlimbany', 'head3', 'face3',
'neck3', 'chest3', 'abdo3', 'spine3', 'upper3', 'lower3', 'operation',
'neurosurgery', 'laparotomy', 'thoracotomy', 'married', 'english',
'mentalhealth', 'comorbidity', 'ssa'
]
# trim the input file into only the features we want to use
extract_features(data_dir, cleaned_file, extract_file, desired_headings)
X, y, headings = separate_data(True, data_dir, extract_file)
# convert all strings to ints
X, y = map(lambda x: list(map(int, x)), X), map(int, y)
X, y = np.array(list(X)), np.array(list(y))
n_samples, n_features = X.shape
# create a stratified 10-fold cross-validation iterator that generates the
# indices for us
cv = StratifiedKFold(y=y, n_folds=10, shuffle=True)
# create a support vector machine classifier
clf_svm = svm.SVC(kernel='linear', probability=True)
# create a Gaussian naive Bayes classifier
clf_gauss_nb = GaussianNB()
def queryaudio():
result_song = ''
begin = 0
end = 0
artists = []
song_artist = []
if request.method == "POST":
print("FORM DATA RECEIVED")
if "file" not in request.files:
return redirect(request.url)
file = request.files["file"]
# file1 =file.filename.replace(".mp3",".wav")
# sound = AudioSegment.from_mp3(file.filename)
# sound.export(file1, format="wav")
if file.filename == "":
return redirect(request.url)
if file:
def first_feature(feat, index_song, temp, u):
l = []
for i in range(int(temp[index_song]),
int(temp[index_song + 1]), 1):
l.append((u.get_item_vector(i)))
z = AnnoyIndex(100, metric='angular')
for i in range(len(l)):
v = l[i]
z.add_item(i, v)
z.build(100)
x = []
for i in range(0, int(feat.shape[0] / 50), 1):
crop_feat = util.crop_feature(feat, i, nb_step=10)
result1 = z.get_nns_by_vector(crop_feat, n=1)
x.append(result1[0])
return x
def most_common(List):
return (mode(List))
def begin_second(count_s_begin):
count = 0
for i in range(0, len(count_s_begin), 5):
k = count_s_begin[i:i + 5]
try:
kq = (most_common(k) - count)
return kq
except:
pass
count += 1
ef, sr = librosa.load(file.filename, sr=16000)
exact_feature = util.extract_features(ef)
results, count_s_begin = util.predict(exact_feature, songs, model1)
result_song = []
top = 5
for q in range(top):
k = results[q][0]
song = k.split(
"E:/Computer Science/HK5/CS336/Audio_query/data/")[1]
song_artist.append(song)
s_a = song.split(" -")
artists.append(s_a[1].split(".mp3")[0])
result_song.append(s_a[0])
index_song = list_song.index(result_song[0] + " -" + artists[0] +
".mp3")
second_begin = first_feature(exact_feature, index_song, time,
model1)
begin = begin_second(second_begin)
begin = int(begin) * 0.1
end = begin + (len(ef) / 16000)
return render_template("queryaudio.html",
songs=result_song,
song_artist=song_artist,
artists=artists,
begin=begin,
end=round(end, 1))
from sklearn.metrics import roc_auc_score
from sklearn.cross_validation import StratifiedKFold, cross_val_score
data_dir = '../../data'
cleaned_file = 'trauma_los_cleaned.csv'
extract_file = 'trauma_los_cleaned_extract.csv'
desired_headings = ['sex', 'normalvitals', 'gcs1', 'iss8', 'age65', 'transfer',
'penetrating', 'mechcode', 'bodyregions', 'headany', 'faceany', 'neckany',
'chestany', 'abdoany', 'spineany', 'upperlimbany', 'lowerlimbany', 'head3',
'face3', 'neck3', 'chest3', 'abdo3', 'spine3', 'upper3', 'lower3',
'operation', 'neurosurgery', 'laparotomy', 'thoracotomy', 'married',
'english', 'mentalhealth', 'comorbidity', 'ssa']
# trim the input file into only the features we want to use
extract_features(data_dir, cleaned_file, extract_file, desired_headings)
X, y, headings = separate_data(True, data_dir, extract_file)
# convert all strings to ints
X, y = map(lambda x:list(map(int, x)), X), map(int, y)
X, y = np.array(list(X)), np.array(list(y))
n_samples, n_features = X.shape
# create a stratified 10-fold cross-validation iterator that generates the
# indices for us
cv = StratifiedKFold(y=y, n_folds=10, shuffle=True)
# create a support vector machine classifier
clf_svm = svm.SVC(kernel='linear', probability=True)
# create a Gaussian naive Bayes classifier
clf_gauss_nb = GaussianNB()
for idx, line in enumerate(f):
real_counter += 1
elements = line.strip('\r\n').split('\t')
word_1 = elements[0]
word_2 = elements[1]
score = float(elements[2])
if use_other_info == 'false':
if word_1 not in word_dict or word_2 not in word_dict:
continue
pred_score_dict[counter] = cosin_distance(word_emb_dict[word_dict[word_1]],
word_emb_dict[word_dict[word_2]])
index_dict[counter] = score
counter += 1
else:
sim_vec1 = extract_features(use_other_info, word_1, word_dict, word_emb_dict,
pos_dict, pos_emb_dict, parser_dict, parser_emb_dict,
dict_desc_dict, kb_dict, kb_emb_dict)
if sim_vec1 == None:
continue
sim_vec2 = extract_features(use_other_info, word_2, word_dict, word_emb_dict,
pos_dict, pos_emb_dict, parser_dict, parser_emb_dict,
dict_desc_dict, kb_dict, kb_emb_dict)
if sim_vec2 == None:
continue
pred_score_dict[counter] = cosin_distance(sim_vec1, sim_vec2)
index_dict[counter] = score
counter += 1
f.close()
print("counter is %d, real_counter is %d" % (counter, real_counter))
sorted_x = sorted(index_dict.items(), key=operator.itemgetter(1))
real_value_list = []
def features(): # display a page with the plots of the zcr and spectral centroid (possibly other spectral features) variations here f = extract_features(signal = session['samples'])
def classify(model, wav):
x = extract_features(wav)
prob = model.predict_proba(x)[0, 1]
return prob
def load(path):
wav, sr = librosa.load(path)
x1 = extract_features(wav)
return x1
def load(path):
wav, _ = librosa.load(path)
x = extract_features(wav)
return x
def main():
util.set_api_key()
messenger.print_task('Fetching song list')
genres_songs = collector.fetch_songs_from_artists(genres_artists.MAP)
messenger.print_task('Fetching songs analyses')
try:
with open(settings.REGISTRY) as fd:
downloaded_songs = simplejson.load(fd)
except JSONDecodeError:
downloaded_songs = {}
for genre, songs in genres_songs.iteritems():
for ith_song, song in enumerate(songs):
if song.id in downloaded_songs:
continue
try:
analysis = collector.fetch_song_analysis(song)
except Exception as exc:
messenger.print_subtask_error(exc)
continue
analysis_file = '%s_%d.json' % (genre, ith_song)
analysis_file = os.path.join(settings.ANALYSES_DIR, analysis_file)
with open(analysis_file, 'w') as analysis_fd:
simplejson.dump(analysis, analysis_fd)
downloaded_songs[song.id] = analysis_file
# save which songs have are already downloaded to avoid
# re-downloading their analysis in the future.
# Ugly and bad as hell. I could append lines to a file or use
# sqlite but it does not worth it.
with open(settings.REGISTRY, 'w') as fd:
simplejson.dump(downloaded_songs, fd)
messenger.print_task('Constructing dataset')
analysis_files = os.listdir(settings.ANALYSES_DIR)
# construct a dict behaving like an enum struct: {'genre': number}
genres = set([filename.split('_')[0] for filename in analysis_files])
genres -= settings.IGNORE_GENRES
genres = dict([(genre, i) for i, genre in enumerate(genres)])
with open(settings.DATASET_FILE, 'w') as dataset_fd:
dataset_writer = csv.writer(dataset_fd)
headers = util.song_features_names()
dataset_writer.writerow(headers + ['genre'])
for analysis_file in analysis_files:
# dirty hack: get genre from filename
song_genre = analysis_file.split('_')[0]
if song_genre not in genres:
continue
file_path = os.path.join(settings.ANALYSES_DIR, analysis_file)
with open(file_path) as fd:
analysis = simplejson.load(fd)
song_features = util.extract_features(analysis)
# Construct a csv row containing all the features of a song, sorted
# by feature name. The last column is a number representing a
# specific genre.
csv_row = [song_features[f] for f in headers]
csv_row.append(genres[song_genre])
dataset_writer.writerow(csv_row)
print 'Done.'
