Spaces:
Running
Running
Joshua Lochner
commited on
Commit
•
25f1183
1
Parent(s):
320a2ba
Use multiclass classifier to filter predictions
Browse files
out/runs/Jan18_13-34-23_DESKTOP-I39NJG7/1642505668.7632372/events.out.tfevents.1642505668.DESKTOP-I39NJG7.27016.1
ADDED
Binary file (5.12 kB). View file
|
|
out/runs/Jan18_13-34-23_DESKTOP-I39NJG7/events.out.tfevents.1642505668.DESKTOP-I39NJG7.27016.0
ADDED
Binary file (3.51 kB). View file
|
|
src/predict.py
CHANGED
@@ -1,3 +1,4 @@
|
|
|
|
1 |
from utils import re_findall
|
2 |
from shared import OutputArguments
|
3 |
from typing import Optional
|
@@ -25,6 +26,7 @@ import logging
|
|
25 |
|
26 |
import re
|
27 |
|
|
|
28 |
def seconds_to_time(seconds, remove_leading_zeroes=False):
|
29 |
fractional = round(seconds % 1, 3)
|
30 |
fractional = '' if fractional == 0 else str(fractional)[1:]
|
@@ -35,6 +37,7 @@ def seconds_to_time(seconds, remove_leading_zeroes=False):
|
|
35 |
hms = re.sub(r'^0(?:0:0?)?', '', hms)
|
36 |
return f"{'-' if seconds < 0 else ''}{hms}{fractional}"
|
37 |
|
|
|
38 |
@dataclass
|
39 |
class TrainingOutputArguments:
|
40 |
|
@@ -68,13 +71,15 @@ class PredictArguments(TrainingOutputArguments):
|
|
68 |
)
|
69 |
|
70 |
|
71 |
-
|
|
|
|
|
72 |
|
73 |
MATCH_WINDOW = 25 # Increase for accuracy, but takes longer: O(n^3)
|
74 |
MERGE_TIME_WITHIN = 8 # Merge predictions if they are within x seconds
|
75 |
|
76 |
|
77 |
-
@dataclass
|
78 |
class ClassifierArguments:
|
79 |
classifier_dir: Optional[str] = field(
|
80 |
default='classifiers',
|
@@ -101,7 +106,7 @@ class ClassifierArguments:
|
|
101 |
default=0.5, metadata={'help': 'Remove all predictions whose classification probability is below this threshold.'})
|
102 |
|
103 |
|
104 |
-
def
|
105 |
"""Use classifier to filter predictions"""
|
106 |
if not predictions:
|
107 |
return predictions
|
@@ -114,14 +119,34 @@ def filter_predictions(predictions, classifier_args): # classifier, vectorizer,
|
|
114 |
])
|
115 |
probabilities = classifier.predict_proba(transformed_segments)
|
116 |
|
|
|
|
|
117 |
filtered_predictions = []
|
118 |
-
for prediction,
|
119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
|
121 |
-
|
122 |
-
|
123 |
-
#
|
124 |
-
|
|
|
|
|
|
|
|
|
|
|
125 |
|
126 |
return filtered_predictions
|
127 |
|
@@ -140,7 +165,6 @@ def predict(video_id, model, tokenizer, segmentation_args, words=None, classifie
|
|
140 |
)
|
141 |
|
142 |
predictions = segments_to_predictions(segments, model, tokenizer)
|
143 |
-
|
144 |
# Add words back to time_ranges
|
145 |
for prediction in predictions:
|
146 |
# Stores words in the range
|
@@ -148,8 +172,8 @@ def predict(video_id, model, tokenizer, segmentation_args, words=None, classifie
|
|
148 |
words, prediction['start'], prediction['end'])
|
149 |
|
150 |
# TODO add back
|
151 |
-
|
152 |
-
|
153 |
|
154 |
return predictions
|
155 |
|
@@ -171,6 +195,9 @@ def greedy_match(list, sublist):
|
|
171 |
return best_i, best_j, best_k
|
172 |
|
173 |
|
|
|
|
|
|
|
174 |
def predict_sponsor_text(text, model, tokenizer):
|
175 |
"""Given a body of text, predict the words which are part of the sponsor"""
|
176 |
input_ids = tokenizer(
|
@@ -189,7 +216,7 @@ def predict_sponsor_matches(text, model, tokenizer):
|
|
189 |
if CustomTokens.NO_SEGMENT.value in sponsorship_text:
|
190 |
return []
|
191 |
|
192 |
-
return re_findall(
|
193 |
|
194 |
|
195 |
def segments_to_predictions(segments, model, tokenizer):
|
@@ -237,12 +264,11 @@ def segments_to_predictions(segments, model, tokenizer):
|
|
237 |
start_time = range['start']
|
238 |
end_time = range['end']
|
239 |
|
240 |
-
if prev_prediction is not None and
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
#
|
245 |
-
# so we extend last prediction range
|
246 |
final_predicted_time_ranges[-1]['end'] = end_time
|
247 |
|
248 |
else: # No overlap, is a new prediction
|
@@ -279,7 +305,7 @@ def main():
|
|
279 |
|
280 |
predict_args.video_id = predict_args.video_id.strip()
|
281 |
predictions = predict(predict_args.video_id, model, tokenizer,
|
282 |
-
segmentation_args
|
283 |
|
284 |
video_url = f'https://www.youtube.com/watch?v={predict_args.video_id}'
|
285 |
if not predictions:
|
@@ -292,7 +318,7 @@ def main():
|
|
292 |
print('Text: "',
|
293 |
' '.join([w['text'] for w in prediction['words']]), '"', sep='')
|
294 |
print('Time:', seconds_to_time(
|
295 |
-
prediction['start']), '
|
296 |
print('Probability:', prediction.get('probability'))
|
297 |
print('Category:', prediction.get('category'))
|
298 |
print()
|
|
|
1 |
+
from shared import START_SEGMENT_TEMPLATE, END_SEGMENT_TEMPLATE
|
2 |
from utils import re_findall
|
3 |
from shared import OutputArguments
|
4 |
from typing import Optional
|
|
|
26 |
|
27 |
import re
|
28 |
|
29 |
+
|
30 |
def seconds_to_time(seconds, remove_leading_zeroes=False):
|
31 |
fractional = round(seconds % 1, 3)
|
32 |
fractional = '' if fractional == 0 else str(fractional)[1:]
|
|
|
37 |
hms = re.sub(r'^0(?:0:0?)?', '', hms)
|
38 |
return f"{'-' if seconds < 0 else ''}{hms}{fractional}"
|
39 |
|
40 |
+
|
41 |
@dataclass
|
42 |
class TrainingOutputArguments:
|
43 |
|
|
|
71 |
)
|
72 |
|
73 |
|
74 |
+
_SEGMENT_START = START_SEGMENT_TEMPLATE.format(r'(?P<category>\w+)')
|
75 |
+
_SEGMENT_END = END_SEGMENT_TEMPLATE.format(r'\w+')
|
76 |
+
SEGMENT_MATCH_RE = fr'{_SEGMENT_START}\s*(?P<text>.*?)\s*(?:{_SEGMENT_END}|$)'
|
77 |
|
78 |
MATCH_WINDOW = 25 # Increase for accuracy, but takes longer: O(n^3)
|
79 |
MERGE_TIME_WITHIN = 8 # Merge predictions if they are within x seconds
|
80 |
|
81 |
|
82 |
+
@dataclass(frozen=True, eq=True)
|
83 |
class ClassifierArguments:
|
84 |
classifier_dir: Optional[str] = field(
|
85 |
default='classifiers',
|
|
|
106 |
default=0.5, metadata={'help': 'Remove all predictions whose classification probability is below this threshold.'})
|
107 |
|
108 |
|
109 |
+
def add_predictions(predictions, classifier_args): # classifier, vectorizer,
|
110 |
"""Use classifier to filter predictions"""
|
111 |
if not predictions:
|
112 |
return predictions
|
|
|
119 |
])
|
120 |
probabilities = classifier.predict_proba(transformed_segments)
|
121 |
|
122 |
+
# Transformer sometimes says segment is of another category, so we
|
123 |
+
# update category and probabilities if classifier is confident it is another category
|
124 |
filtered_predictions = []
|
125 |
+
for prediction, probabilities in zip(predictions, probabilities):
|
126 |
+
predicted_probabilities = {k: v for k,
|
127 |
+
v in zip(CATEGORIES, probabilities)}
|
128 |
+
|
129 |
+
# Get best category + probability
|
130 |
+
classifier_category = max(
|
131 |
+
predicted_probabilities, key=predicted_probabilities.get)
|
132 |
+
classifier_probability = predicted_probabilities[classifier_category]
|
133 |
+
|
134 |
+
if classifier_category is None and classifier_probability > classifier_args.min_probability:
|
135 |
+
continue # Ignore
|
136 |
+
|
137 |
+
if classifier_category is not None and classifier_probability > 0.5: # TODO make param
|
138 |
+
# Confident enough to overrule, so we update category
|
139 |
+
prediction['category'] = classifier_category
|
140 |
|
141 |
+
prediction['probability'] = predicted_probabilities[prediction['category']]
|
142 |
+
|
143 |
+
# TODO add probabilities, but remove None and normalise rest
|
144 |
+
prediction['probabilities'] = predicted_probabilities
|
145 |
+
|
146 |
+
# if prediction['probability'] < classifier_args.min_probability:
|
147 |
+
# continue
|
148 |
+
|
149 |
+
filtered_predictions.append(prediction)
|
150 |
|
151 |
return filtered_predictions
|
152 |
|
|
|
165 |
)
|
166 |
|
167 |
predictions = segments_to_predictions(segments, model, tokenizer)
|
|
|
168 |
# Add words back to time_ranges
|
169 |
for prediction in predictions:
|
170 |
# Stores words in the range
|
|
|
172 |
words, prediction['start'], prediction['end'])
|
173 |
|
174 |
# TODO add back
|
175 |
+
if classifier_args is not None:
|
176 |
+
predictions = add_predictions(predictions, classifier_args)
|
177 |
|
178 |
return predictions
|
179 |
|
|
|
195 |
return best_i, best_j, best_k
|
196 |
|
197 |
|
198 |
+
CATEGORIES = [None, 'SPONSOR', 'SELFPROMO', 'INTERACTION']
|
199 |
+
|
200 |
+
|
201 |
def predict_sponsor_text(text, model, tokenizer):
|
202 |
"""Given a body of text, predict the words which are part of the sponsor"""
|
203 |
input_ids = tokenizer(
|
|
|
216 |
if CustomTokens.NO_SEGMENT.value in sponsorship_text:
|
217 |
return []
|
218 |
|
219 |
+
return re_findall(SEGMENT_MATCH_RE, sponsorship_text)
|
220 |
|
221 |
|
222 |
def segments_to_predictions(segments, model, tokenizer):
|
|
|
264 |
start_time = range['start']
|
265 |
end_time = range['end']
|
266 |
|
267 |
+
if prev_prediction is not None and \
|
268 |
+
(start_time <= prev_prediction['end'] <= end_time or # Merge overlapping segments
|
269 |
+
(range['category'] == prev_prediction['category'] # Merge disconnected segments if same category and within threshold
|
270 |
+
and start_time - prev_prediction['end'] <= MERGE_TIME_WITHIN)):
|
271 |
+
# Extend last prediction range
|
|
|
272 |
final_predicted_time_ranges[-1]['end'] = end_time
|
273 |
|
274 |
else: # No overlap, is a new prediction
|
|
|
305 |
|
306 |
predict_args.video_id = predict_args.video_id.strip()
|
307 |
predictions = predict(predict_args.video_id, model, tokenizer,
|
308 |
+
segmentation_args, classifier_args=classifier_args)
|
309 |
|
310 |
video_url = f'https://www.youtube.com/watch?v={predict_args.video_id}'
|
311 |
if not predictions:
|
|
|
318 |
print('Text: "',
|
319 |
' '.join([w['text'] for w in prediction['words']]), '"', sep='')
|
320 |
print('Time:', seconds_to_time(
|
321 |
+
prediction['start']), '\u2192', seconds_to_time(prediction['end']))
|
322 |
print('Probability:', prediction.get('probability'))
|
323 |
print('Category:', prediction.get('category'))
|
324 |
print()
|