def find_ads(self, video_ids):
if not self._fitted:
raise Exception("Train or load model before inference")
if not isinstance(video_ids, list):
video_ids = [video_ids]
result = []
for video_id in video_ids:
info = VideoInfo(video_id)
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
ads = []
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
words = list(
map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split()))
x = self.vectorizer.infer_vector(words)
x.append(scenes[l][0] / info.duration)
y = self.subs_classifier.predict([x])[0]
if y == 1:
ads.append((scenes[l][0], scenes[r][1]))
merged_ads = []
for ad in ads:
if len(merged_ads) == 0 or ad[0] - merged_ads[-1][1] > 10:
merged_ads.append(ad)
else:
merged_ads.append((merged_ads.pop()[0], ad[1]))
result.append(merged_ads)
return result if len(result) > 1 else result[0]
def find_ads(self, video_ids):
if not self._fitted:
raise Exception("Train or load model before inference")
if not isinstance(video_ids, list):
video_ids = [video_ids]
result = []
for video_id in video_ids:
info = VideoInfo(video_id)
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
segments = []
xs = []
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
words = map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split())
words = filter(lambda word: word in self.buzz_words, words)
words = map(lambda word: self.buzz_words[word], words)
subs_part = ' '.join(words)
x = list(self.vectorizer.transform([subs_part]).toarray()[0])
x.append(scenes[l][0] / info.duration)
xs.append(x)
seg = (scenes[l][0], scenes[r][1])
segments.append(seg)
segments = [(0, 0)] * (self.window // 2) + segments + [
(0, 0)
] * (self.window // 2)
features = len(xs[0])
xs = [[0] * features] * (
self.window // 2) + xs + [[0] * features] * (self.window // 2)
X = []
for i in range(self.window // 2, len(xs) - self.window // 2):
X.append(xs[i - self.window // 2:i + self.window // 2 + 1])
X = np.array(X)
Y = self.subs_classifier.predict(X) > 0.5
ads = []
for i in range(self.window // 2, len(X) - self.window // 2):
cnt = 0
for d in range(-(self.window // 2), self.window // 2 + 1):
cnt += Y[i + d][self.window // 2 - d][0]
if cnt > self.window // 2:
ads.append(segments[i])
merged_ads = []
for ad in ads:
if len(merged_ads) == 0 or ad[0] - merged_ads[-1][1] > 10:
merged_ads.append(ad)
else:
merged_ads.append((merged_ads.pop()[0], ad[1]))
result.append(merged_ads)
return result if len(result) > 1 else result[0]
def train(self, markups):
video_ids = list(markups.keys())
X = []
Y = []
print('start model training')
for idx, video_id in enumerate(video_ids):
print('\rprocessing video {} ({}/{})'.format(
video_id, idx + 1, len(video_ids)),
end='')
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
xs = []
ys = []
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
words = map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split())
words = filter(lambda word: word in self.buzz_words, words)
words = map(lambda word: self.buzz_words[word], words)
subs_part = ' '.join(words)
x = list(self.vectorizer.transform([subs_part]).toarray()[0])
info = VideoInfo(video_id)
x.append(scenes[l][0] / info.duration)
seg = (scenes[l][0], scenes[r][1])
xs.append(x)
if self._intersect_ad(seg, markups[video_id]):
ys.append(0.5)
elif self._inside_ad(seg, markups[video_id]):
ys.append(1)
else:
ys.append(0)
features = len(xs[0])
xs = [[0] * features] * (
self.window // 2) + xs + [[0] * features] * (self.window // 2)
ys = [0] * (self.window // 2) + ys + [0] * (self.window // 2)
for i in range(self.window // 2, len(xs) - self.window // 2):
X.append(xs[i - self.window // 2:i + self.window // 2 + 1])
Y.append(ys[i - self.window // 2:i + self.window // 2 + 1])
X = np.array(X)
Y = np.array(Y)
Y = Y.reshape((Y.shape[0], Y.shape[1], 1))
print(X.shape, Y.shape)
model = Sequential()
model.add(
Bidirectional(LSTM(100, return_sequences=True),
input_shape=(self.window, X.shape[2])))
model.add(TimeDistributed(Dense(1, activation='sigmoid')))
model.compile(optimizer='adam', loss='binary_crossentropy')
np.random.seed(0)
model.fit(X, Y, batch_size=len(X), shuffle=True, epochs=300)
self.subs_classifier = model
self._fitted = True
def find_ads(self, video_ids):
if not isinstance(video_ids, list):
video_ids = [video_ids]
result = []
for video_id in video_ids:
info = VideoInfo(video_id)
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
xs = []
segments = []
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
words = list(
map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split()))
x = list(self.vectorizer.infer_vector(words))
x.append((scenes[l][0] + scenes[r][1]) / 2 / info.duration)
xs.append(x)
segments.append((scenes[l][0], scenes[r][1]))
segments = [(0, 0)] * (self.window // 2) + segments + [
(0, 0)
] * (self.window // 2)
features = len(xs[0])
xs = [[0] * features] * (
self.window // 2) + xs + [[0] * features] * (self.window // 2)
X = []
for i in range(self.window // 2, len(xs) - self.window // 2):
X.append(xs[i - self.window // 2:i + self.window // 2 + 1])
X = np.array(X)
Y = self.subs_classifier.predict(X) > 0.5
ads = []
for i in range(self.window // 2, len(X) - self.window // 2):
cnt = 0
for d in range(-(self.window // 2), self.window // 2 + 1):
cnt += Y[i + d][self.window // 2 - d][0]
if cnt > self.window // 2:
ads.append(segments[i])
merged_ads = []
for ad in ads:
if len(merged_ads) == 0 or ad[0] - merged_ads[-1][1] > 10:
merged_ads.append(ad)
else:
merged_ads.append((merged_ads.pop()[0], ad[1]))
result.append(merged_ads)
return result if len(result) > 1 else result[0]
def train(self, markups):
video_ids = list(markups.keys())
X = []
Y = []
print('start model training')
tagged_data = []
subs_parts = []
for idx, video_id in enumerate(video_ids):
print('\rprocessing video {} ({}/{})'.format(
video_id, idx + 1, len(video_ids)),
end='')
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
seg = (scenes[l][0], scenes[r][1])
if self._intersect_ad(seg, markups[video_id]):
continue
words = list(
map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split()))
subs_parts.append(words)
tagged_data.append(
TaggedDocument(words=words, tags=[str(len(tagged_data))]))
if self._inside_ad(seg, markups[video_id]):
Y.append(1)
else:
Y.append(0)
self.vectorizer = Doc2Vec(size=50, alpha=0.025, dm=0)
self.vectorizer.build_vocab(tagged_data)
self.vectorizer.train(tagged_data,
total_examples=self.vectorizer.corpus_count,
epochs=500)
X = np.array(
[self.vectorizer.infer_vector(words) for words in subs_parts])
Y = np.array(Y)
print('X shape: ', X.shape)
print('Y shape: ', Y.shape)
self.subs_classifier = MLPClassifier(hidden_layer_sizes=(50, 10),
solver='adam',
random_state=0,
learning_rate='adaptive',
max_iter=1000)
self.subs_classifier.fit(X, Y)
self._fitted = True
def train(self, markups):
video_ids = list(markups.keys())
X = []
Y = []
print('start model training')
for idx, video_id in enumerate(video_ids):
print('\rprocessing video {} ({}/{})'.format(
video_id, idx + 1, len(video_ids)),
end='')
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
words = map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split())
words = filter(lambda word: word in self.buzz_words, words)
words = map(lambda word: self.buzz_words[word], words)
subs_part = ' '.join(words)
x = list(self.vectorizer.transform([subs_part]).toarray()[0])
info = VideoInfo(video_id)
x.append(scenes[l][0] / info.duration)
seg = (scenes[l][0], scenes[r][1])
if self._intersect_ad(seg, markups[video_id]):
continue
X.append(x)
if self._inside_ad(seg, markups[video_id]):
Y.append(1)
else:
Y.append(0)
X = np.array(X)
Y = np.array(Y)
self.subs_classifier = MLPClassifier(hidden_layer_sizes=(15, 5),
solver='adam',
random_state=0,
learning_rate='adaptive',
max_iter=1000)
self.subs_classifier.fit(X, Y)
self._fitted = True
def train(self, markups):
video_ids = list(markups.keys())
print('start model training')
tagged_data = []
subs_parts = []
Ys = []
ps = []
for idx, video_id in enumerate(video_ids):
print('\rprocessing video {} ({}/{})'.format(
video_id, idx + 1, len(video_ids)),
end='')
subs = Subtitles(video_id, preprocess_russian_text_with_morph)
scenes = self.sdm.detect_scenes(video_id)
r = 0
sp = []
ys = []
p = []
info = VideoInfo(video_id)
for l in range(len(scenes)):
while r + 1 < len(scenes) and scenes[r][1] - scenes[l][
0] < self.subs_part_len:
r += 1
seg = (scenes[l][0], scenes[r][1])
words = list(
map(str.strip,
subs.fulltext(scenes[l][0], scenes[r][1]).split()))
sp.append(words)
tagged_data.append(
TaggedDocument(words=words, tags=[str(len(tagged_data))]))
p.append((seg[0] + seg[1]) / 2 / info.duration)
if self._inside_ad(seg, markups[video_id]):
ys.append(1)
elif self._intersect_ad(seg, markups[video_id]):
ys.append(0.5)
else:
ys.append(0)
subs_parts.append(sp)
Ys.append(ys)
ps.append(p)
self.vectorizer = Doc2Vec(size=50, alpha=0.025, dm=0)
self.vectorizer.build_vocab(tagged_data)
self.vectorizer.train(tagged_data,
total_examples=self.vectorizer.corpus_count,
epochs=500)
X = []
Y = []
for idx, video_id in enumerate(video_ids):
print('\rprocessing video {} ({}/{})'.format(
video_id, idx + 1, len(video_ids)),
end='')
xs = [
self.vectorizer.infer_vector(words)
for words in subs_parts[idx]
]
xs = list(
np.hstack((xs, np.reshape(ps[idx],
(len(ps[idx]), 1)))).tolist())
ys = Ys[idx]
features = len(xs[0])
xs = [[0] * features] * (
self.window // 2) + xs + [[0] * features] * (self.window // 2)
ys = [0] * (self.window // 2) + ys + [0] * (self.window // 2)
for i in range(self.window // 2, len(xs) - self.window // 2):
X.append(xs[i - self.window // 2:i + self.window // 2 + 1])
Y.append(ys[i - self.window // 2:i + self.window // 2 + 1])
X = np.array(X)
Y = np.array(Y)
Y = Y.reshape((Y.shape[0], Y.shape[1], 1))
print(X.shape, Y.shape)
model = Sequential()
model.add(
Bidirectional(LSTM(100, return_sequences=True),
input_shape=(self.window, X.shape[2])))
model.add(TimeDistributed(Dense(1, activation='sigmoid')))
model.compile(optimizer='adam', loss='binary_crossentropy')
np.random.seed(0)
model.fit(X, Y, batch_size=len(X), shuffle=True, epochs=500)
self.subs_classifier = model
self._fitted = True