Spaces:
Build error
Build error
File size: 4,793 Bytes
f3772cc |
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 |
import os
import gradio as gr
import torch
from utils.unifiedmodel import RRUMDataset
from utils.huggingface_model_wrapper import YoutubeVideoSimilarityModel
from torch.utils.data import DataLoader
from helpers import get_example_videos, update_youtube_embedded_html, get_input_data_df
RR_EXAMPLES_URL = os.environ.get(
'RR_EXAMPLES_URL', 'https://public-data.telemetry.mozilla.org/api/v1/tables/telemetry_derived/regrets_reporter_study/v1/files/000000000000.json')
NUM_RR_EXAMPLES = 5
example_videos, example_videos_rr = get_example_videos(
RR_EXAMPLES_URL, NUM_RR_EXAMPLES)
demo_title = 'Mozilla RegretsReporter YouTube video similarity'
demo_description = f'''
# {demo_title}
This demo showcases the YouTube video semantic similarity model developed as part of the RegretsReporter research project at Mozilla Foundation. You can read more about the project [here](https://foundation.mozilla.org/en/youtube/user-controls/) and about the semantic similarity model [here](https://foundation.mozilla.org/en/blog/the-regretsreporter-user-controls-study-machine-learning-to-measure-semantic-similarity-of-youtube-videos/). Note: the model is multilingual so you can try it with non-English videos too while it probably works the best with English videos.
This demo works by inserting two YouTube video URLs below and clicking the Run button. After a few seconds, you will see model's predicted probability of how similar those two videos are. You can copy URLs from YouTube or also try out a few predefined examples by clicking them on the examples table.
'''
placeholder_youtube_embedded_html = '''
<p>Insert video URL first</p>
'''
model_wt = YoutubeVideoSimilarityModel.from_pretrained(
'mozilla-foundation/youtube_video_similarity_model_wt', use_auth_token=True)
model_nt = YoutubeVideoSimilarityModel.from_pretrained(
'mozilla-foundation/youtube_video_similarity_model_nt', use_auth_token=True)
cross_encoder_model_name_or_path = model_wt.cross_encoder_model_name_or_path
def get_video_similarity(video1_url, video2_url):
df = get_input_data_df(video1_url, video2_url)
if df['regret_transcript'].isna().any() or df['recommendation_transcript'].isna().any():
with_transcript = False
else:
with_transcript = True
dataset = RRUMDataset(df, with_transcript=with_transcript, label_col=None,
cross_encoder_model_name_or_path=cross_encoder_model_name_or_path)
data_loader = DataLoader(dataset.test_dataset, shuffle=False,
batch_size=1, num_workers=0, pin_memory=False)
with torch.inference_mode():
if with_transcript:
pred = model_wt(next(iter(data_loader)))
else:
pred = model_nt(next(iter(data_loader)))
pred = torch.special.expit(pred).squeeze().tolist()
return f'YouTube videos are {pred:.0%} similar'
with gr.Blocks(title=demo_title) as demo:
gr.Markdown(demo_description)
with gr.Row():
with gr.Column():
input_text1 = gr.Textbox(
label='Video 1', placeholder='Insert first YouTube video URL')
input_text2 = gr.Textbox(
label='Video 2', placeholder='Insert second YouTube video URL')
inputs = [input_text1, input_text2]
with gr.Row():
clear_btn = gr.Button('Clear', variant='secondary')
run_btn = gr.Button('Run', variant='primary')
with gr.Column():
outputs = [gr.Label(label='Model prediction')]
with gr.Accordion('See video details', open=False):
with gr.Row():
with gr.Column():
video_embedded = gr.HTML(
value=placeholder_youtube_embedded_html)
with gr.Column():
video_embedded2 = gr.HTML(
value=placeholder_youtube_embedded_html)
with gr.Column():
if example_videos:
examples = gr.Examples(examples=example_videos, inputs=inputs)
if example_videos_rr:
examples_rr = gr.Examples(examples=example_videos_rr, inputs=inputs,
label='Example bad becommendations from the RegretsReporter report')
run_btn.click(fn=get_video_similarity, inputs=inputs, outputs=outputs)
clear_btn.click(lambda value_1, value_2, value_3: (
None, None, None), inputs=inputs + outputs, outputs=inputs + outputs)
input_text1.change(lambda input: update_youtube_embedded_html(
input, 1) if input else placeholder_youtube_embedded_html, inputs=input_text1, outputs=video_embedded)
input_text2.change(lambda input: update_youtube_embedded_html(
input, 2) if input else placeholder_youtube_embedded_html, inputs=input_text2, outputs=video_embedded2)
demo.launch()
|