def load_all_feature(mode='train', feat_type='db4'):
flist = [] # wav file list
label = [] # wav label list
with open(MODE[mode], 'r') as f:
for line in f:
data = line.split()
flist.append(data[0])
label.append(LABEL[data[1]])
final_flist = []
final_label = []
final_feat = []
for idx in range(len(flist)):
wav_path = os.path.join(WAV[mode], flist[idx])
try:
feat = extract(wav_path, feat_type)
except:
print('CORRUPTED WAV: ', wav_path)
continue
final_feat.append(feat)
final_flist.append(flist[idx])
final_label.append([label[idx]] *
len(feat)) # label expand (wavs, frames)
return final_feat, final_label, final_flist
def upload_predict():
#upload file
file = request.files['file']
if file.filename.find('.') <= 0:
return redirect(url_for('main_page'))
file_ext = file.filename.rsplit('.', 1)[1].strip().lower()
if file_ext not in ALLOWED_EXT:
return redirect(url_for('main_page'))
user_input = request.form.get("name")
file_name = str(uuid.uuid1()).replace('-', '') + '.' + file_ext
basepath = os.path.dirname(__file__)
upload_path = os.path.join(basepath, 'static/images', file_name)
file.save(upload_path)
#get feature
feature = extract(upload_path)
#get recommendation
urls = recommend(feature)
origin = os.path.join('static/images', file_name)
return render_template('display.html', origin=origin, urls=urls)
def display_result():
if request.method == 'POST':
f = request.files['file']
f.save(secure_filename(f.filename))
feature = extract(f.filename)
# print(len(feature[0]), file=sys.stdout)
products = predict(feature[0])
metadata = mongo.db.metadata
output = {}
refer = {}
i = 0
while (i < len(products)):
asin = products[i]
asin = str(asin)
print(asin)
try:
cursor = metadata.find({"asin": asin})
p = cursor[0]
output[p['asin']] = p['imUrl']
i += 1
except:
i += 1
continue
#refer[p['title']]=p['asin']
#refer = jsonify(refer)
return render_template('display.html', result=output)
def load_rnn_data(dataset, label_file, mode="train", feature_type="mfcc"):
mat_pattern = r"D:\experiments\anti\asvspoof2017\baseline\features\{}_{}.mat"
if mode != "final":
labels, wav_lists = load_label(label_file)
else:
labels = {}
wav_lists = []
train_labels, train_wav_lists = load_label(label_file[0])
dev_labels, dev_wav_lists = load_label(label_file[1])
labels.update(train_labels)
labels.update(dev_labels)
wav_lists.extend(train_wav_lists)
wav_lists.extend(dev_wav_lists)
dataset = "final"
final_data = []
final_label = []
final_wav_ids = []
final_lengths = []
for wav_name in tqdm(wav_lists[0:100],
desc="load {} data".format(dataset)):
wav_id = wav_name.replace(".wav", "")
label = labels[wav_id]
if "E" in wav_id:
wav_path = "../data/ASVspoof2017_eval/{}.wav".format(wav_id)
if "D" in wav_id:
wav_path = "../data/ASVspoof2017_dev/{}.wav".format(wav_id)
if "T" in wav_id:
wav_path = "../data/ASVspoof2017_train/{}.wav".format(wav_id)
if feature_type != 'cqcc':
feature = extract_feature.extract(wav_path, feature_type)
if feature_type == "mfcc" or feature_type == "imfcc":
feature_delta = librosa.feature.delta(feature)
feature_delta_delta = librosa.feature.delta(feature_delta)
feature = np.concatenate(
(feature, feature_delta, feature_delta_delta), axis=0)
else:
mat_path = mat_pattern.format(wav_id, feature_type)
tmp_data = scio.loadmat(mat_path)
feature = tmp_data['x']
feature = feature.T
final_lengths.append(feature.shape[0])
feature = np.pad(feature,
[[0, rnn_max_frames - feature.shape[0]], [0, 0]],
mode="constant")
final_data.append(feature)
final_label.append(label)
final_wav_ids.append(wav_id)
return final_data, final_label, final_wav_ids, final_lengths
def test_classificatore(imgpath, classificatore, pose_names):
img = cv.imread(imgpath)
img = cv.resize(img, (500, 500))
glob_fe = extract_feature.extract(img)
pred = classificatore.predict(glob_fe.reshape(1, -1))[0]
cv.putText(img, 'Predizione posizione: {}'.format(pose_names[pred]), (20, 30), cv.FONT_HERSHEY_SIMPLEX, 0.8,(255, 255, 0), 2)
cv.putText(img, 'Classificatore: {}'.format(classificatore), (20, 60), cv.FONT_HERSHEY_SIMPLEX, 0.8,(255, 255, 0), 2)
cv.imshow("Predizione ", img)
print("la predizione è': ", pose_names[pred])
cv.waitKey(0)
def extract_verif_feature():
verif_path_feature_dic = {} # {person:[feature1,feature2,...,]}
# path_feature_dic = msgpack_numpy.load(open(originalimages_verif_fc7_path_feature, 'rb'))
path_set = set()
for line in open(pair_file):
tmp = line.rstrip().split()
path_set.add(tmp[0])
path_set.add(tmp[1])
model, get_Conv_FeatureMap = load_model(layer_index=-5)
for pic_path in path_set:
fine_tune_feature = extract(pic_path, get_Conv_FeatureMap, (224, 224, 3))
verif_path_feature_dic[pic_path] = fine_tune_feature
msgpack_numpy.dump(verif_path_feature_dic, open(feature_pack_file, 'wb'))
def __getitem__(self, idx):
'''
Return:
`tuple(feature, label, file name)`
`feature: (wav, feats, dim*11)
**WAV CORRUPT EXCEPTION NOT HUNDLE**
'''
if type(idx) is not slice:
if self.buf and self.buffer[idx] is not None:
return self.buffer[idx]
wav_path = os.path.join(WAV[self.mode], self.flist[idx])
feat = extract(wav_path, self.feat_type) # wav corrupt not hundle
item = (feat, [self.label[idx]] * feat.shape[0], self.flist[idx])
if self.buf:
self.buffer[idx] = item
return item
# slice reading
items = []
length = self.__len__()
for i in range(idx.start, idx.stop):
i %= length
if self.buf and self.buffer[i] is not None:
items.append(self.buffer[i])
continue
wav_path = os.path.join(WAV[self.mode], self.flist[i])
feat = extract(wav_path, self.feat_type)
items.append(
(feat, [self.label[i]] * feat.shape[0], self.flist[i]))
if self.buf:
self.buffer[i] = items[-1]
return items
def main():
folder = '/data/hanlin'
person_path_dic = load_one_deep_path(folder)
sample_list, person_num = person_path_dic_trans(person_path_dic)
model, get_Conv_FeatureMap = load_model(output_layer_index=18)
data = []
label = []
start = time()
for pic_path, person_index in sample_list:
feature_vector = extract(pic_path, get_Conv_FeatureMap, pic_shape)[0]
data.append(feature_vector)
label.append(person_index)
end = time()
print(end - start)
msgpack_numpy.dump((data, label), open('hanlin.p', 'wb'))
def main():
folder = '/data/hanlin'
person_path_dic = load_one_deep_path(folder)
sample_list, person_num = person_path_dic_trans(person_path_dic)
model, get_Conv_FeatureMap = load_model(output_layer_index=18)
data = []
label = []
start = time()
for pic_path, person_index in sample_list:
feature_vector = extract(pic_path, get_Conv_FeatureMap, pic_shape)[0]
data.append(feature_vector)
label.append(person_index)
end = time()
print (end - start)
msgpack_numpy.dump((data, label), open('hanlin.p', 'wb'))
def extract_verif_feature():
verif_path_feature_dic = {} # {person:[feature1,feature2,...,]}
# path_feature_dic = msgpack_numpy.load(open(originalimages_verif_fc7_path_feature, 'rb'))
path_set = set()
for line in open(pair_file):
tmp = line.rstrip().split()
path_set.add(tmp[0])
path_set.add(tmp[1])
model, get_Conv_FeatureMap = load_model(layer_index=-5)
print model.summary()
for pic_path in path_set:
fine_tune_feature = extract(pic_path, get_Conv_FeatureMap, (224, 224, 3))
verif_path_feature_dic[pic_path] = fine_tune_feature
msgpack_numpy.dump(verif_path_feature_dic, open(feature_pack_file, 'wb'))
X = []
y = []
x = []
imgsize = 150
for training_name in train_labels:
label = training_name
DIR = os.path.join(TRAIN_DIR, training_name)
for img in tqdm(os.listdir(DIR)):
path = os.path.join(DIR, img)
img = cv2.imread(path)
img1 = cv2.imread(path)
img = cv2.resize(img, (resize, resize))
img1 = cv2.resize(img, (imgsize, imgsize))
glob_features_b = extract_feature.extract(img)
X.append(np.array(glob_features_b))
x.append(np.array(img1))
y.append(str(label))
z = []
z = y
'normalizzazione'
X = np.array(X)
X = X / 255
#Label Encoding
le = LabelEncoder()
y = le.fit_transform(y)
def load_data(dataset, label_file, mode="train", feature_type="mfcc"):
extend_num = 6
mat_pattern = r"D:\experiments\anti\asvspoof2017\baseline\features\{}_{}.mat"
if mode != "final":
labels, wav_lists = load_label(label_file)
else:
labels = {}
wav_lists = []
train_labels, train_wav_lists = load_label(label_file[0])
dev_labels, dev_wav_lists = load_label(label_file[1])
labels.update(train_labels)
labels.update(dev_labels)
wav_lists.extend(train_wav_lists)
wav_lists.extend(dev_wav_lists)
dataset = "final"
if mode == "train" or mode == "final":
final_data = []
final_label = []
for wav_name in tqdm(wav_lists, desc="load {} data".format(dataset)):
wav_id = wav_name.replace(".wav", "")
label = labels[wav_id]
if "T" in wav_id:
wav_path = "../data/ASVspoof2017_train/{}.wav".format(wav_id)
if "D" in wav_id:
wav_path = "../data/ASVspoof2017_dev/{}.wav".format(wav_id)
if feature_type != 'cqcc':
feature = extract_feature.extract(wav_path, feature_type)
else:
mat_path = mat_pattern.format(wav_id, feature_type)
tmp_data = scio.loadmat(mat_path)
feature = tmp_data['x']
# feature = feature[0:60, :]
# feature = feature.T
# feature = scale(feature.T, with_mean=True, with_std=True)
feature = np.pad(feature,
[[0, 0], [extend_num - 1, extend_num - 1]],
mode="edge")
for i in range(extend_num - 1, feature.shape[1] - extend_num):
tmp_feature = feature[:, i - extend_num + 1:i +
extend_num].reshape(-1)
final_data.append(tmp_feature)
final_label.append(label)
return final_data, final_label
elif mode == "test":
final_data = []
final_label = []
final_wav_ids = []
for wav_name in tqdm(wav_lists, desc="load {} data".format(dataset)):
wav_id = wav_name.replace(".wav", "")
label = labels[wav_id]
if "E" in wav_id:
wav_path = "../data/ASVspoof2017_eval/{}.wav".format(wav_id)
if "D" in wav_id:
wav_path = "../data/ASVspoof2017_dev/{}.wav".format(wav_id)
if feature_type != 'cqcc':
feature = extract_feature.extract(wav_path, feature_type)
else:
mat_path = mat_pattern.format(wav_id, feature_type)
tmp_data = scio.loadmat(mat_path)
feature = tmp_data['x']
feature = np.pad(feature,
[[0, 0], [extend_num - 1, extend_num - 1]],
mode="edge")
final_feature = []
for i in range(extend_num - 1, feature.shape[1] - extend_num):
tmp_feature = feature[:, i - extend_num + 1:i +
extend_num].reshape(-1)
final_feature.append(tmp_feature)
final_feature = np.array(final_feature).astype(np.float32)
final_data.append(final_feature)
final_label.append(label)
final_wav_ids.append(wav_id)
return final_data, final_label, final_wav_ids
else:
raise ValueError("the mode doesn't exist")
def load_cnn_data(dataset, label_file, mode="train", feature_type="mfcc"):
mat_pattern = r"D:\experiments\anti\asvspoof2017\baseline\features\{}_{}.mat"
if mode != "final":
labels, wav_lists = load_label(label_file)
else:
labels = {}
wav_lists = []
train_labels, train_wav_lists = load_label(label_file[0])
dev_labels, dev_wav_lists = load_label(label_file[1])
labels.update(train_labels)
labels.update(dev_labels)
wav_lists.extend(train_wav_lists)
wav_lists.extend(dev_wav_lists)
dataset = "final"
final_data = []
final_label = []
final_wav_ids = []
for wav_name in tqdm(wav_lists[:], desc="load {} data".format(dataset)):
wav_id = wav_name.replace(".wav", "")
label = labels[wav_id]
if "T" in wav_id:
wav_path = "../data/ASVspoof2017_train/{}.wav".format(wav_id)
if "D" in wav_id:
wav_path = "../data/ASVspoof2017_dev/{}.wav".format(wav_id)
if "E" in wav_id:
wav_path = "../data/ASVspoof2017_eval/{}.wav".format(wav_id)
if feature_type != 'cqcc':
feature = extract_feature.extract(wav_path, feature_type)
else:
mat_path = mat_pattern.format(wav_id, feature_type)
tmp_data = scio.loadmat(mat_path)
feature = tmp_data['x']
feature = feature.T
feature = scale(feature, with_mean=True, with_std=False, copy=True)
# if feature_type == 'cqcc':
# if feature.shape[0] < 400:
# feature = np.pad(feature, [[0, 400 - feature.shape[0]], [0, 0]], mode='edge')
# feature = feature[0:400, :]
# final_data.append(feature)
if feature_type == 'cqcc':
if feature.shape[0] < 20:
feature = np.pad(feature, [[0, 20 - feature.shape[0]], [0, 0]],
mode='edge')
feature = feature[0:20, :]
final_data.append(feature)
# elif feature_type == "fft":
# if feature.shape[0] < 400:
# feature = np.pad(feature, [[0, 400 - feature.shape[0]], [0, 0]], mode='edge')
# feature = feature[0:400, 0:864]
# final_data.append(feature)
else:
if feature.shape[0] < 500:
feature = np.pad(feature,
[[0, 500 - feature.shape[0]], [0, 0]],
mode='edge')
feature = feature[0:500, 0:501]
final_data.append(feature)
final_label.append(label)
final_wav_ids.append(wav_id)
final_data = np.array(final_data)
if mode != "test":
return final_data, final_label
else:
return final_data, final_label, final_wav_ids
import extract_feature as EF
import xgb_useing as model
print('提取特征')
EF.extract('windows')
print('模型训练')
model.run_model('windows')
