File size: 11,447 Bytes
4822df2
23a1215
4822df2
 
 
 
 
 
 
 
 
 
 
8a55e13
3d1c770
 
94ad7ba
3d1c770
8a55e13
4822df2
 
e68b946
 
4822df2
 
 
 
 
 
 
 
94ad7ba
 
 
df05196
94ad7ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e68b946
 
 
 
 
14ea568
 
 
f9281a4
c415610
14ea568
 
f9281a4
c415610
 
 
 
 
 
 
14ea568
e68b946
85661b3
e68b946
 
85661b3
e68b946
85661b3
e68b946
85661b3
 
e68b946
 
85661b3
e68b946
85661b3
 
e68b946
 
4822df2
004109e
 
 
 
 
4822df2
3d1c770
4822df2
9ce97dc
8a55e13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3d1c770
8a55e13
 
 
 
23a1215
 
3d1c770
 
23a1215
8a55e13
23a1215
 
14ea568
f11d2c2
4822df2
c415610
 
3d1c770
 
c415610
 
3d1c770
4822df2
c415610
3d1c770
 
00f77c2
c415610
4822df2
 
8a55e13
 
 
 
 
 
 
 
4822df2
8a55e13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
921fb1d
8a55e13
 
 
 
 
 
 
 
e926596
 
4822df2
 
e926596
a9123fa
 
4822df2
e926596
e68b946
8a55e13
 
 
 
 
 
 
 
 
4822df2
96dad7a
 
 
 
 
 
 
 
 
 
 
e926596
 
 
 
 
4822df2
 
e926596
a9123fa
4822df2
e926596
4822df2
 
e926596
 
 
 
a9123fa
e926596
 
 
 
 
8a55e13
 
 
 
 
 
 
 
 
 
 
 
 
e926596
8a55e13
e926596
 
 
 
 
 
 
3d1c770
 
e926596
 
 
 
 
 
 
 
 
3d1c770
e926596
 
 
 
 
3d1c770
e926596
 
 
 
 
 
 
 
 
 
 
3d1c770
e926596
 
 
96dad7a
8a55e13
 
8326048
 
8a55e13
 
 
 
 
e926596
8a55e13
 
 
 
 
 
 
4822df2
23a1215
4822df2
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319

from functools import partial
from math import ceil, floor
import streamlit.components.v1 as components
import streamlit as st
import sys
import os
import json
from urllib.parse import quote

# Allow direct execution
sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), 'src'))  # noqa

from preprocess import get_words
from predict import PredictArguments, SegmentationArguments, predict as pred
from shared import GeneralArguments, seconds_to_time, CATGEGORY_OPTIONS
from utils import regex_search
from model import get_model_tokenizer_classifier
from errors import TranscriptError

st.set_page_config(
    page_title='SponsorBlock ML',
    page_icon='🤖',
    #  layout='wide',
    #  initial_sidebar_state="expanded",
    menu_items={
        'Get Help': 'https://github.com/xenova/sponsorblock-ml',
        'Report a bug': 'https://github.com/xenova/sponsorblock-ml/issues/new/choose',
        #  'About': "# This is a header. This is an *extremely* cool app!"
    }
)


YT_VIDEO_REGEX = r'''(?x)^
                (?:
                    # http(s):// or protocol-independent URL
                    (?:https?://|//)
                    (?:(?:(?:(?:\w+\.)?[yY][oO][uU][tT][uU][bB][eE](?:-nocookie|kids)?\.com/|
                    youtube\.googleapis\.com/)                        # the various hostnames, with wildcard subdomains
                    (?:.*?\#/)?                                          # handle anchor (#/) redirect urls
                    (?:                                                  # the various things that can precede the ID:
                        # v/ or embed/ or e/
                        (?:(?:v|embed|e)/(?!videoseries))
                        |(?:                                             # or the v= param in all its forms
                            # preceding watch(_popup|.php) or nothing (like /?v=xxxx)
                            (?:(?:watch|movie)(?:_popup)?(?:\.php)?/?)?
                            (?:\?|\#!?)                                  # the params delimiter ? or # or #!
                            # any other preceding param (like /?s=tuff&v=xxxx or ?s=tuff&v=V36LpHqtcDY)
                            (?:.*?[&;])??
                            v=
                        )
                    ))
                    |(?:
                    youtu\.be                                        # just youtu.be/xxxx
                    )/)
                )?                                                       # all until now is optional -> you can pass the naked ID
                # here is it! the YouTube video ID
                (?P<id>[0-9A-Za-z_-]{11})'''

# https://github.com/google-research/text-to-text-transfer-transformer#released-model-checkpoints
# https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#experimental-t5-pre-trained-model-checkpoints

# https://huggingface.co/docs/transformers/model_doc/t5
# https://huggingface.co/docs/transformers/model_doc/t5v1.1


# Faster caching system for predictions (No need to hash)
@st.cache(persist=True, allow_output_mutation=True)
def create_prediction_cache():
    return {}


@st.cache(persist=True, allow_output_mutation=True)
def create_function_cache():
    return {}


prediction_cache = create_prediction_cache()
prediction_function_cache = create_function_cache()

MODELS = {
    'Small (293 MB)': {
        'pretrained': 'google/t5-v1_1-small',
        'repo_id': 'Xenova/sponsorblock-small',
        'num_parameters': '77M'
    },
    'Base v1 (850 MB)': {
        'pretrained': 't5-base',
        'repo_id': 'Xenova/sponsorblock-base-v1',
        'num_parameters': '220M'
    },

    'Base v1.1 (944 MB)': {
        'pretrained': 'google/t5-v1_1-base',
        'repo_id': 'Xenova/sponsorblock-base-v1.1',
        'num_parameters': '250M'
    }
}

# Create per-model cache
for m in MODELS:
    if m not in prediction_cache:
        prediction_cache[m] = {}


CLASSIFIER_PATH = 'Xenova/sponsorblock-classifier-v2'


TRANSCRIPT_TYPES = {
    'AUTO_MANUAL': {
        'label': 'Auto-generated (fallback to manual)',
        'type': 'auto',
        'fallback': 'manual'
    },
    'MANUAL_AUTO': {
        'label': 'Manual (fallback to auto-generated)',
        'type': 'manual',
        'fallback': 'auto'
    },
    # 'TRANSLATED': 'Translated to English' # Coming soon
}


def predict_function(model_id, model, tokenizer, segmentation_args, classifier, video_id, words, ts_type_id):
    cache_id = f'{video_id}_{ts_type_id}'

    if cache_id not in prediction_cache[model_id]:
        prediction_cache[model_id][cache_id] = pred(
            video_id, model, tokenizer,
            segmentation_args=segmentation_args,
            words=words,
            classifier=classifier
        )
    return prediction_cache[model_id][cache_id]


def load_predict(model_id):
    model_info = MODELS[model_id]

    if model_id not in prediction_function_cache:
        # Use default segmentation and classification arguments
        predict_args = PredictArguments(model_name_or_path=model_info['repo_id'])
        general_args = GeneralArguments()
        segmentation_args = SegmentationArguments()

        model, tokenizer, classifier = get_model_tokenizer_classifier(predict_args, general_args)

        prediction_function_cache[model_id] = partial(
            predict_function, model_id, model, tokenizer, segmentation_args, classifier)


    return prediction_function_cache[model_id]


def create_button(text, url):
    return f"""<div class="row-widget stButton" style="text-align: center">
        <a href="{url}" target="_blank" rel="noopener noreferrer" class="btn-link">
            <button kind="primary" class="btn">{text}</button>
        </a>
    </div>"""


def main():
    st.markdown("""<style>
    .btn {
        display: inline-flex;
        -webkit-box-align: center;
        align-items: center;
        -webkit-box-pack: center;
        justify-content: center;
        font-weight: 600;
        padding: 0.25rem 0.75rem;
        border-radius: 0.25rem;
        margin: 0px;
        line-height: 1.5;
        color: inherit;
        width: auto;
        user-select: none;
        background-color: inherit;
        border: 1px solid rgba(49, 51, 63, 0.2);
    }
    .btn-link {
        color: inherit;
        text-decoration: none;
    }
    </style>""", unsafe_allow_html=True)

    top = st.container()
    output = st.empty()

    # Display heading and subheading
    top.markdown('# SponsorBlock ML')
    top.markdown(
        '##### Automatically detect in-video YouTube sponsorships, self/unpaid promotions, and interaction reminders.')

    # Add controls

    col1, col2 = top.columns(2)

    with col1:
        model_id = st.selectbox(
            'Select model', MODELS.keys(), index=0, on_change=output.empty)

    with col2:
        ts_type_id = st.selectbox(
            'Transcript type', TRANSCRIPT_TYPES.keys(), index=0, format_func=lambda x: TRANSCRIPT_TYPES[x]['label'], on_change=output.empty)

    query_params = st.experimental_get_query_params()

    video_id = None
    
    if 'v' in query_params:
        video_id = query_params['v'][0]

    if video_id is None:
        video_input = top.text_input('Video URL/ID:', on_change=output.empty)
    else :
        video_input = top.text_input('Video URL/ID:', on_change=output.empty,value = video_id)
    categories = top.multiselect('Categories:',
                                 CATGEGORY_OPTIONS.keys(),
                                 CATGEGORY_OPTIONS.keys(),
                                 format_func=CATGEGORY_OPTIONS.get, on_change=output.empty
                                 )

    # Hide segments with a confidence lower than
    confidence_threshold = top.slider(
        'Confidence Threshold (%):', min_value=0, value=50, max_value=100, on_change=output.empty)

    if len(video_input) == 0:  # No input, do not continue
        return

    # Load prediction function
    with st.spinner('Loading model...'):
        predict = load_predict(model_id)

    with output.container():  # Place all content in output container
        video_id = regex_search(video_input, YT_VIDEO_REGEX)
        if video_id is None:
            st.exception(ValueError('Invalid YouTube URL/ID'))
            return

        try:
            with st.spinner('Downloading transcript...'):
                words = get_words(video_id,
                                  transcript_type=TRANSCRIPT_TYPES[ts_type_id]['type'],
                                  fallback=TRANSCRIPT_TYPES[ts_type_id]['fallback']
                                  )
        except TranscriptError:
            pass

        if not words:
            st.error('No transcript found!')
            return

        with st.spinner('Running model...'):
            predictions = predict(video_id, words, ts_type_id)

        if len(predictions) == 0:
            st.success('No segments found!')
            return

        submit_segments = []
        for index, prediction in enumerate(predictions, start=1):
            category_key = prediction['category'].upper()
            if category_key not in categories:
                continue  # Skip

            confidence = prediction['probability'] * 100

            if confidence < confidence_threshold:
                continue

            submit_segments.append({
                'segment': [prediction['start'], prediction['end']],
                'category': prediction['category'],
                'actionType': 'skip'
            })
            start_time = seconds_to_time(prediction['start'])
            end_time = seconds_to_time(prediction['end'])
            with st.expander(
                f"[{category_key}] Prediction #{index} ({start_time} \u2192 {end_time})"
            ):

                url = f"https://www.youtube-nocookie.com/embed/{video_id}?&start={floor(prediction['start'])}&end={ceil(prediction['end'])}"
                # autoplay=1controls=0&&modestbranding=1&fs=0

                # , width=None, height=None, scrolling=False
                components.iframe(url, width=670, height=376)

                text = ' '.join(w['text'] for w in prediction['words'])
                st.write(f"**Times:** {start_time} \u2192 {end_time}")
                st.write(
                    f"**Category:** {CATGEGORY_OPTIONS[category_key]}")
                st.write(f"**Confidence:** {confidence:.2f}%")
                st.write(f'**Text:** "{text}"')

        if not submit_segments:
            st.success(
                f'No segments found! ({len(predictions)} ignored due to filters/settings)')
            return

        num_hidden = len(predictions) - len(submit_segments)
        if num_hidden > 0:
            st.info(
                f'{num_hidden} predictions hidden (adjust the settings and filters to view them all).')

        json_data = quote(json.dumps(submit_segments))
        link = f'https://www.youtube.com/watch?v={video_id}#segments={json_data}'
        st.markdown(create_button('Submit Segments', link),
                    unsafe_allow_html=True)

        st.markdown(f"""<div style="text-align: center;font-size: 16px;margin-top: 6px">
        <a href="https://wiki.sponsor.ajay.app/w/Automating_Submissions" target="_blank" rel="noopener noreferrer">(Review before submitting!)</a>
        </div>""", unsafe_allow_html=True)


if __name__ == '__main__':
    main()