Spark in me - Internet, data science, math, deep learning, philosophy. Страница 23
2440
@snakers4
Канал про интересные мне темы - интернет - статистика - наука о данных Без рекламы и буллшита.
-
Spark in me - Internet, data science, math, deep learning, philosophy
nvidia-gpu-accelerator-feeds-speeds.jpg 
-
Spark in me - Internet, data science, math, deep learning, philosophy
Парсим ГАР БД ФИАС в удобный формат в питоне. Бесплатно, без регистрации и СМС https://habr.com/ru/post/595107/ -
Spark in me - Internet, data science, math, deep learning, philosophy
А ты используешь VAD? Что это такое и зачем он нужен Мы сделали уникальный VAD мирового уровня (судите сами по метрикам), который работает на 1 ядре процессора с задержкой в 1 миллисекунду на кусочках аудио от 30 миллисекунд. В этой статье мы расскажем вам, что такое VAD, покажем на примерах как использовать его и наглядно потестировать на своем голосе. - https://habr.com/ru/post/594745/ С вас плюсики на Хабре и помощь в дальнейшей упаковке - github.com/snakers…sues/130 -
Реклама
-
Spark in me - Internet, data science, math, deep learning, philosophy
Even Better V3 Silero VAD - Models with even higher quality (just see the plots with metrics!); - New model ~ large model >> all previous (even large) models; - Now model works properly quality-wise, i.e. 100ms > 60ms > 30ms and16 kHz > 8 kHz; -
Spark in me - Internet, data science, math, deep learning, philosophy
Community Help Wanted We just published a radically simplified and minimalized VAD. Unlike other tasks, VAD seems like a more or less solved task. We were previously reluctant to package it properly, while it was in active development. Now adding a condapackage may actually be a good idea. We have not donecondapackages yet. I believe there may be some people on community with experience in using Github Actions to publishcondapackages. The VAD in Pytorch essentially is just ajitmodel with 50 lines of utils code. If you would like to help us - please just PM me personally. Probably you will help us do it 10x faster. github.com/snakers…sues/130. -
Spark in me - Internet, data science, math, deep learning, philosophy
Восстановление знаков пунктуации и заглавных букв — теперь и на длинных текстах https://habr.com/ru/post/594565/ P. S. Делитесь остроумными примерами работы модели в комментариях и голосуйте за понравившиеся! Лучшие фразы добавим как встроенные примеры в следующем релизе :D -
Spark in me - Internet, data science, math, deep learning, philosophy
Improved Text Recapitalization and Repunctuation Model for 4 Languages - The model now can work with long inputs, 512 tokens or ca. 150 words; - Inputs longer than 150 words are automatically processed in chunks; - The bugs with newer PyTorch versions have been fixed; - Model was trained longer with larger batches; - Model size slightly reduced to 85 MB; - The rest of model optimizations were deemed too high maintenance; Link -
Spark in me - Internet, data science, math, deep learning, philosophy
-
Spark in me - Internet, data science, math, deep learning, philosophy
BirchPunk is Back! Video I thought they became sell-outs doing rushed ads for Russian corporate agencies / military But looks like they are back on track -
Spark in me - Internet, data science, math, deep learning, philosophy
If you have a twitter account, please re-tweet twitter.com/Alexand…08981506 More in depth articles on Habr coming soon -
Spark in me - Internet, data science, math, deep learning, philosophy
New V3 Silero VAD is Already Here Main changes - One VAD to rule them all! - New model includes the functionality of all of the previous ones with improved quality and speed! - As far as we know, our VAD is the best in the world now; - Flexible sampling rate, 8000 Hz and 16000 Hz are supported; - Flexible chunk size, minimum chunk size is just 30 milliseconds! - Only 100k parameters; - GPU inference and batching are supported (the model is small, so we decided not to publish a quantized model); - Radically, drastically simplified examples; We also drastically polished and simplified README, wiki and repo in general. Links: - Silero VAD repo - https://github.com/snakers4/silero-vad - The migration to V3 is quite simple, here are some examples - Quality metrics - Performance metrics - Examples and dependencies - Colab with examples If you like Silero VAD, please give us a ⭐️ and spread the news! -
Spark in me - Internet, data science, math, deep learning, philosophy
Digest 2021-11 # Hardware Cerebras Completes Series F Funding, Another $250M for $4B Valuation - www.anandtech.com/show/17…aluation Testing А2000 до А6000, А10, А16, А30, А40 и A100 PCIe - habr.com/ru/comp…g/589811 # Datasets ACAV100M: Scaling up self-supervised audio-visual learning with automatically curated internet videos - www.microsoft.com/en-us/r…t-videos - We processed 140 million full-length videos (total duration 1,030 years, consuming 2 petabytes of storage) and reduced it to 100 million clips (10 seconds each, 31 years in total) that exhibit the highest audio-visual correspondence. #digest -
Spark in me - Internet, data science, math, deep learning, philosophy
Digest 2021-11 # Tech Tesla's Trillion Dollar Moment: A Valuation Revisit! - aswathdamodaran.blogspot.com/2021/11…ion.html Как работает модерация на Хабре - https://habr.com/ru/company/habr/blog/589587/ Don’t Let Architecture Astronauts Scare You - www.joelonsoftware.com/2001/04…care-you OpenAI’s API Now Available with No Waitlist - https://openai.com/blog/api-no-waitlist/ News Highlight: Facebook, Citing Societal Concerns, Plans to Shut Down Facial Recognition System - thegradientpub.substack.com/p/gradi…uts-down AI-driven search engine You.com takes on Google with $20M - venturebeat.com/2021/11…with-20m # Blogs Strong AI Requires Autonomous Building of Composable Models - thegradient.pub/strong-…e-models Reflections on Foundation Models - thegradient.pub/reflect…n-models Individuals matter (in workload planning) - https://danluu.com/people-matter/ Brain Computer Confusion - blog.piekniewski.info/2021/11…onfusion When big tech buys small tech - www.ben-evans.com/benedic…all-tech Cascade of doom: JIT, and how a Postgres update led to 70% failure on a critical national service - dev.to/xenatis…ice-3f2a Last week in AI 142 - https://lastweekin.ai/p/142 - this time more signal than noise How Much Should You Lie? - www.strangeloopcanon.com/p/how-m…-you-lie Is it worth the money? When to buy products for your job - pythonspeed.com/article…ng-guide Activating a Conda environment in your Dockerfile - pythonspeed.com/article…ckerfile #digest -
Spark in me - Internet, data science, math, deep learning, philosophy
Digest 2021-11 # Code Multi-stage builds #2: Python specifics - pythonspeed.com/article…r-python Unravelling decorators - https://snarky.ca/unravelling-decorators/ A terrible schema from a clueless programmer - https://rachelbythebay.com/w/2021/11/06/sql/ I'm not feeling the async pressure - lucumr.pocoo.org/2020/1/…pressure Installing system packages in Docker with minimal bloat - pythonspeed.com/article…s-docker Selecting a programming language can be a form of premature optimization - snarky.ca/program…mization Copying data is wasteful, mutating data is dangerous - pythonspeed.com/article…-copying Minimal safe Bash script template - betterdev.blog/minimal…template #digest -
Spark in me - Internet, data science, math, deep learning, philosophy
Digest 2021-11 # ML / Papers Predicting Spreadsheet Formulas from Semi-structured Contexts - ai.googleblog.com/2021/10…rom.html ruDALL-E: генерируем изображения по текстовому описанию, или Самый большой вычислительный проект в России - habr.com/ru/comp…g/586926 AI voice-acting tool xVASynth gets even better at recreating videogame voices - www.pcgamer.com/ai-voic…e-voices GPT-3 is No Longer the Only Game in Town - lastweekin.ai/p/gpt-3…nly-game Accelerating PyTorch with CUDA Graphs - pytorch.org/blog/ac…a-graphs Model Ensembles Are Faster Than You Think - ai.googleblog.com/2021/11…you.html Self-Supervised Learning Advances Medical Image Classification - ai.googleblog.com/2021/10…ces.html MetNet-2: Deep Learning for 12-Hour Precipitation Forecasting - ai.googleblog.com/2021/11…our.html Predicting Text Readability from Scrolling Interactions - ai.googleblog.com/2021/11…rom.html Decisiveness in Imitation Learning for Robots - ai.googleblog.com/2021/11…for.html Paper Review: Swin Transformer V2 Scaling Up Capacity and Resolution - andlukyane.com//blog/p…-swin-v2 Explain Yourself - A Primer on ML Interpretability & Explainability - https://thegradient.pub/explain-yourself/ SimVLM: Simple Visual Language Model Pre-training with Weak Supervision - ai.googleblog.com/2021/10…pre.html Predicting Text Selections with Federated Learning - ai.googleblog.com/2021/11…ith.html The Inherent Limitations of GPT-3 - lastweekin.ai/p/the-i…of-gpt-3 Turing Bletchley: A Universal Image Language Representation model by Microsoft - www.microsoft.com/en-us/r…icrosoft - Encoder-decoder - Trained using billions of image-caption pairs drawn from the web - Parallel corpus of 500 million translation pairs - Pairs were created by extracting sentences from document-aligned webpages from common crawl corpus - Transformer-based image and text encoders which are both analogous to the BERT-large architecture #digest -
Реклама
-
Spark in me - Internet, data science, math, deep learning, philosophy
Digest 2021-11 # Speech Towards Building ASR Systems for the Next Billion Users - http://arxiv.org/abs/2111.03945 - 17,000 hours of raw speech data for 40 Indian languages from a wide variety of domains (YouTube + newsonair) - Pretrain several variants of wav2vec style models for 40 Indian languages - Fine-tune this model for downstream ASR for 9 languages - WER reported for Indic ASR is significantly higher and sensitively depends on availability of resources: pretraining corpus, fine-tuning data, and task-specific language information - youtube-dl + py-webrtcvad => 16 8kHz, 1.5 TB in wav format - 24 GPUs to train, 10-24 to tune, not feasible without English pre-train - KenLM models search helps a lot, rescoring helps a bit - Just adding more annotated data helps a lot PSEUDO-LABELING FOR MASSIVELY MULTILINGUAL SPEECH RECOGNITION - http://arxiv.org/abs/2111.00161 - Data - CV with (3.5k hours) and Vox Populi (384k) unlabeled, 60 languages - Base model (275M), we also train larger models (1.06B) with CTC loss and classifier - Base - 16 GPUs with dynamic batching using 200s of audio per batch per GPU - Large - 64 GPUs with 50s of audio per GPU - SlimIPL, an iterative approach - a number of updates on labeled data - Followed by continuous training using labeled data and pseudo-labeled data stored in a dynamic cache which is periodically updated with pseudo-labels (PLs) re-generated by the current model state - First fine-tune the trained multilingual model by training only on CV data for that language for 10k updates, and then run slimIPL using the corresponding VP data - Utterances of VP (average duration of 30s) are much longer than those of CV (average duration of 5.3s) - TLDR - VP pre-training / tuning improves quality for very low resource languages, but the quality on CV becomes worse for high-resource langauges #digest -
Spark in me - Internet, data science, math, deep learning, philosophy
Choosing the best GPU for Training / Inference In my opinion - 3090 is still the best even at 2x the recommended price. And A4000 is a clear winner for inference, with 140W TDP and single-slot. -
Spark in me - Internet, data science, math, deep learning, philosophy
📎 Silero TTS Full Articles
Now we have published the accompanying articles both in English and in Russian on habr.com:
- Russian - https://habr.com/ru/post/549480/
- English - https://habr.com/ru/post/549482/
📎 Speakers
10 voices (each available in 16 kHz and 8 kHz):
- 6 Russian voices;
- 1 English voice;
- 1 German voice, 1 Spanish voice, 1 French voice;
📎 Why is this Different?
- One-line usage;
- A large library of voices;
- A fully end-to-end pipeline;
- Naturally sounding speech;
- No GPU or training required;
- Minimalism and lack of dependencies;
- Faster than real-time on one CPU thread (!!!);
- Support for16kHzand8kHzout of the box;
Please upvote, like, share and try!
📎 Bot
Someone even made a bot with it for Ksenia, but it is still buggy - @silero_tts_bot.
it bugs out when you add stress with+and changes its location in a weird way, but something like this worksя ко+тик ко+тик ко+тик, пуши+стый живо+тик!(stress should be before letters, but it works this way, idk why).