def __init__(self, rate, data, filename):
self.filename = filename
shp = data.shape
if len(shp) > 1 and shp[0] > shp[1]:
data = data.T
data = wnl.lib_to_mono(data)
data = data[: rate * 30]
self.rate = rate
self.data = data
self.bvratio = wnl.beat_variance_ratio(rate, data)
self.silence_ratio = wnl.silence_ratio(rate, data)
self.mfcc, self.mfcc_delta = wnl.librosa_mfcc_delta(rate, data)
self.chromagram = wnl.librosa_chromagram(rate, data)
self.spectroid = wnl.spec_centroid(rate, data)
def get_feature_choice_cmd(ftr=None, ftr_sel=None, path=None, cls=None, win_len=None):
if ftr is None:
print("Please select the feature (Separated by | for multiple):")
print("1. Beat Variance Ratio")
print("2. Silence Ratio")
print("3. MFCC")
print("4. MFCC Delta")
print("5. Chromagram")
print("6. Spectroid")
print("7. FFT average over 1s window")
print("8. ZCR over window")
print("9. MFCC over window")
ftr = input()
# ftr = [int(x) for x in ftr.split('|')]
ftr = [opts.split('$')[0] for opts in ftr.split('|')]
ftr_sel = {opts.split('$')[0]: opts.split('$')[1] for opts in ftr.split('|')}
if path is None:
path = input("Enter path to features: \n")
if cls is None:
cls = ["Entertainment", "Music", "Comedy", "Film & Animation", "News & Politics", "Sports", "People & Blogs",
"Howto & Style", "Pets & Animals"]
if win_len is None:
win_len = 0.04
start = True
# features = np.empty(shape=(0, 0))
features = {}
f_max_len = {}
classes = {}
# Path is an array containing potentially multiple features files that can be used to load Video objects from disk.
for p in path:
with open(p + SHORT_FEATS, "rb") as inp:
unpickle = pkl.Unpickler(inp)
count = 0
# Create the UnPickler object and loop until there are no objects left in the file. Break from loop then.
while True:
try:
cur_feature = {}
vid = unpickle.load()
# If video is in the approved class list add all selected features to a dictionary cur_feature
if vid.get_category_from_name() in cls:
count += 1
if 1 in ftr:
cur_feature[1] = vid.bvratio
if 2 in ftr:
cur_feature[2] = vid.silence_ratio
if 3 in ftr:
cur_feature[3] = np.array(vid.mfcc).reshape((1, -1))[0]
if 4 in ftr:
cur_feature[4] = np.array(vid.mfcc_delta).reshape((1, -1))[0]
if 5 in ftr:
cur_feature[5] = np.array(vid.chromagram).reshape((1, -1))[0]
if 6 in ftr:
cur_feature[6] = vid.spectroid[0]
if 7 in ftr:
cur_feature[7] = vid.get_windowed_fft(int(np.ceil(vid.rate * win_len)))
if 8 in ftr:
cur_feature[8] = vid.get_windowed_zcr(int(np.ceil(vid.rate * win_len)))
if 9 in ftr:
cur_feature[9] = np.array(wnl.get_window_mfcc(vid.mfcc, int(np.ceil(vid.rate * win_len)))) \
.reshape((1, -1))[0]
# This section was designed under the assumption that the features could be returned in various
# 2d layouts. It essentially checks the size of the current feature against the largest
# number of columns so far. It then pads the smaller one with 0's
# This can definitely be refactored into simpler, more readable code.
if start:
for i in ftr:
features[i] = np.array([cur_feature[i]])
f_shape = features[i].shape
if hasattr(cur_feature[i], "__len__"):
if len(f_shape) > 1:
f_max_len[i] = f_shape[1]
else:
f_max_len[i] = len(f_shape)
start = False
# classes = np.array(vid.get_category_from_name())
classes[i] = [vid.get_category_from_name()]
else:
for i in ftr:
if hasattr(cur_feature[i], "__len__"):
if len(cur_feature[i].shape) > 1:
if cur_feature[i].shape[1] > f_max_len[i]:
if len(features[i].shape) > 1:
features[i] = np.pad(features[i],
((0, 0), (0, cur_feature[i].shape[1] - f_max_len[i])),
mode="constant")
f_max_len[i] = cur_feature[i].shape[1]
else:
features[i] = np.pad(features[i],
(0, cur_feature[i].shape[1] - f_max_len[i]),
mode="constant")
f_max_len[i] = cur_feature[i].shape[1]
elif cur_feature[i].shape[1] < f_max_len[i]:
cur_feature[i] = np.pad(cur_feature[i],
((0, 0), (0, f_max_len[i] - cur_feature[i].shape[1])),
mode="constant")
elif len(cur_feature[i].shape) == 1:
if cur_feature[i].shape[0] > f_max_len[i]:
if len(features[i].shape) > 1:
features[i] = np.pad(features[i],
((0, 0), (0, cur_feature[i].shape[0] - f_max_len[i])),
mode="constant")
f_max_len[i] = cur_feature[i].shape[0]
else:
features[i] = np.pad(features[i],
(0, cur_feature[i].shape[0] - f_max_len[i]),
mode="constant")
f_max_len[i] = cur_feature[i].shape[0]
elif cur_feature[i].shape[0] < f_max_len[i]:
cur_feature[i] = np.pad(cur_feature[i],
(0, f_max_len[i] - cur_feature[i].shape[0]),
mode="constant")
features[i] = np.vstack((features[i], [cur_feature[i]]))
# classes = np.append(classes, [vid.get_category_from_name()])
classes[i].append(vid.get_category_from_name())
except EOFError:
print("EOF")
break
except TypeError:
print("Unable to load object")
except pkl.UnpicklingError:
print("Unable to load object2")
gc.collect()
# Join each feature into one large array.
# Keep track of the indices for each feature that needs reduction applied later
select_ind = []
print("Count = ", count)
total_feature = features[ftr[0]]
if ftr_sel[ftr[0]] > 0:
select_ind = [(0, len(total_feature[0]), ftr_sel[ftr[0]])]
for i in range(1, len(ftr)):
if ftr_sel[ftr[i]] > 0:
start = len(total_feature[0])
select_ind.append((start, start + len(features[ftr[i]][0]), ftr_sel[ftr[i]]))
total_feature = np.hstack((total_feature, features[ftr[i]]))
# targets = np.array(classes)
targets = [value for key, value in classes.items()]
return total_feature, targets[0], select_ind
def get_windowed_zcr(self, block_length):
return wnl.get_windowed_zcr(self.data, block_length)