Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    josh 🐸
    @josh-coqui:matrix.org
    [m]
    welcome to the community room Lakshman 👋
    I'll invite you to the STT room:)
    josh 🐸
    @josh-coqui:matrix.org
    [m]

    👋 Hello everyone!

    We are hosting a STT model training competition, for tons of languages 🥳

    ... and you can get free GPU access for a week!

    the competition lasts one week, and starts very soon

    Check it out here: https://coqui.ai/

    josh 🐸
    @josh-coqui:matrix.org
    [m]
    kdavis-coqui
    @kdavis-coqui

    @/all Don't forget the Long Tail Language Challenge just launched!

    You can get 1 week of free V100 GPU time to train a STT model for one of many languages. Checkout the details here
    Long Tail Language Challenge

    erogol 🐸
    @golero:matrix.org
    [m]
    Our talk from the GTC https://youtu.be/7S0u9eSR664
    spectie
    @spectie:matrix.org
    [m]
    Sorry, this post has been removed by the moderators of r/MachineLearning.
    O_o
    jmukiibi
    @jmukiibi:mozilla.org
    [m]
    1 reply
    weberjulian 🐸
    @weberjulian:matrix.org
    [m]
    That's interesting indeed but some things aren't clear to me. erogol 🐸 your idea is to use the Q-network for STT or as input for TTS (unsupervised dataset) ? If so, how do we choose/compute the 1s segments since there is no CTC loss or similar mechanism ? Also, it's at the lexical item level (word?) so how do we deal with OOV ? Will the combinations of probability of codes suffise to encode unseen words ?
    I've only skimmed the article so it's highly probable I'm missing something haha
    erogol 🐸
    @golero:matrix.org
    [m]
    weberjulian 🐸: I also need to read in more detail. I'll get back to you
    limasse
    @limasse:matrix.org
    [m]
    Hello ! I am a student and I want to use a coqui-stt for speech recognition. Is there any article that shows clear results after training a model, especially with audios of people with strong accents when they speak english ? Thanks !
    josh 🐸
    @josh-coqui:matrix.org
    [m]
    limasse: welcome to the community room 👋
    I'll invite you to the STT room at gitter.im/coqui-ai/STT
    reuben
    @reuben_m:matrix.org
    [m]
    erogol 🐸: for Coqpit what do you think about a way to do unscoped composition, like mixins?
    erogol 🐸: basically I have a base/shared Coqpit with common arguments, and I want to extend it with some use-case specific options, but without putting the entire base Coqpit under a scope/name by making it an attribute of my use-case specific Coqpit class
    oh, maybe I can just use inheritance for my simple use case
    let me see if that works
    reuben
    @reuben_m:matrix.org
    [m]
    looks like inheritance works as long as there are no required fields
    from coqpit import Coqpit
    from dataclasses import dataclass, field
    from typing import Optional
    
    @dataclass
    class BaseConfig(Coqpit):
        checkpoint_dir: Optional[str] = None
    
    @dataclass
    class AppConfig(BaseConfig):
        src_file: Optional[str] = None
    
    print(AppConfig.new_from_dict(dict(
         checkpoint_dir = "/bar",
         src_file = "foo",
    )))
    erogol 🐸
    @golero:matrix.org
    [m]
    reuben: inheritance is the way I use for TTS
    reuben
    @reuben_m:matrix.org
    [m]
    erogol 🐸: yeah, it seems to work for my use case as well. it's unfortunate that the subclasses fields come last in the argparse definition tho...
    erogol 🐸
    @golero:matrix.org
    [m]
    reuben: what is the impact of it ?
    reuben
    @reuben_m:matrix.org
    [m]
    erogol 🐸: just readability. I filed an issue for an improvement, will try to tackle it if I have some spare cycles
    josh 🐸
    @josh-coqui:matrix.org
    [m]
    Aya-AlJafari (Aya-AlJafari): welcome to the community room 👋😁
    @jreus: you should ping the TTS room, theyd have more to say there:)
    1 reply
    erogol 🐸
    @golero:matrix.org
    [m]

    reuben: does it make sense to do this? (creating Coqpit classes dynamically from generic classes so you don't need to create a separate static Coqpit)

    Below I intend to create a Coqpit object from BaseCharacters class dynamically by to_coqpit().

    For now I've no idea how but looks useful

    class MyModelConfig(Coqpit):
        field_1: int = 0
        field_2: str = ""
        characters: Coqpit = BaseCharacters().to_coqpit()
        ...
    reuben
    @reuben_m:matrix.org
    [m]
    erogol 🐸: I think having an explicit annotation on config classes is quite useful
    erogol 🐸: but I guess even with this one you could still grep for to_coqpit. and yeah I have no idea how either
    erogol 🐸
    @golero:matrix.org
    [m]

    PyTorch just started M1 compat work

    pytorch/pytorch#68811

    spectie
    @spectie:matrix.org
    [m]
    Hopefully they're more competent than the AMD engineers :D
    reuben
    @reuben_m:matrix.org
    [m]
    but expect 4 months to ship it 😭
    weberjulian 🐸
    @weberjulian:matrix.org
    [m]
    Hey guys, what do you think about using CMP 170HX for DL inference ? It's 7nm so it's more power efficient than a 3090 while having significantly more punch in CUDA applications (based on a A100 die). Although it has only 8Gb of VRAM, it's much faster HBM2 memory. The big downside is the lack of support from NVIDIA I guess...
    OM3GA SOLUTIONS d.o.o
    @om3gasolutions_twitter
    I was unlucky with the CMP 170HX. I failed to start training, each attempt started process on 3090 that was GPU with monitor connected. If I force to use just CMP 170HX process freeze on Epoch 0 | Training | Elapsed Time: 0:00:00 | Steps: 0 . $kill -9 -1 was only option to stop it. GPU activity 0, but it didn't track it on NVIDIA-smi, usage 0 ram 0
    weberjulian 🐸
    @weberjulian:matrix.org
    [m]
    Oh too bad, it's weird that it works for mining, how do they know which CUDA application can run ? Limiting the GPU cuda capability maybe?
    OM3GA SOLUTIONS d.o.o
    @om3gasolutions_twitter
    Maybe it's broken one, as I find out it is A100 die that failed, maybe this one had some issues. It was borrowed from mining farm for test, mining ETC without problem. I tried with that new LHR GPU and ETC mining was limited but training was at same speed as non LHR one. So it's not just CUDA, something else is going on.
    For the record, all this testing was running under Windows 11 WSL2 on Ubuntu 20.04
    1 reply
    thorsten.mueller
    @thorsten.mueller:matrix.org
    [m]

    Good morning 👋
    I've created a Wiki for ALL OPEN VOICE Enthusiasts (STT, TTS, Voice Assistants, Paper Stuff).

    https://OpenVoice-Tech.net

    Maybe we can share/collect our knowledge (lessons learned, best practices, ...) to make them publically available.

    Your feedback is highly appreciated - useful or useless?

    Best
    Thorsten

    reuben
    @reuben_m:matrix.org
    [m]

    interesting paper:
    Sparse is Enough in Scaling Transformers - https://arxiv.org/abs/2111.12763

    Large Transformer models yield impressive results on many tasks, but are expensive to train, or even fine-tune, and so slow at decoding that their use and study becomes out of reach. We address this problem by leveraging sparsity. We study sparse variants for all layers in the Transformer and propose Scaling Transformers, a family of next generation Transformer models that use sparse layers to scale efficiently and perform unbatched decoding much faster than the standard Transformer as we scale up the model size. Surprisingly, the sparse layers are enough to obtain the same perplexity as the standard Transformer with the same number of parameters. We also integrate with prior sparsity approaches to attention and enable fast inference on long sequences even with limited memory. This results in performance competitive to the state-of-the-art on long text summarization.

    1 reply
    wiedymi
    @wiedymi:matrix.org
    [m]
    Hello! How to run bin/import_cv2.py I have errors about imports
    wiedymi
    @wiedymi:matrix.org
    [m]

    No matching distribution found for tensorflow==1.15

    Anyone knows how to solve this?

    1 reply
    reuben
    @reuben_m:matrix.org
    [m]
    or downgrade Python if you're on 3.8 or newer
    max supported by TF 1.15 is Python 3.7
    also, we have a channel for STT discussion: https://gitter.im/coqui-ai/STT
    josh 🐸
    @josh-coqui:matrix.org
    [m]

    ⏰ Reminder

    Just a reminder. Tomorrow we're gonna have our first 🐸TTS community meeting v0.5

    We plan to answer live questions then jump to the link below in the order of votes. Feel free to post your questions starting from today.

    We'll post the recording somewhere in case you miss the call

    So let's see how it's gonna turn out 😄

    👉 Meeting link

    https://meet.google.com/jys-yfpv-wyd

    👉 Meeting time

    Dec 2, 2021 Thursday 17:30 - 18:15 CET
    (Add to your calendar from here))

    👉 Ask or upvote questions here

    https://github.com/coqui-ai/TTS/discussions/categories/qs-for-live-q-a

    this is happening NOW 🎉
    seunbayo
    @seunbayo:matrix.org
    [m]

    Hello everyone,i am seun and i am a frontend developer turned community manager..

    it is awesome to join a fantastic commuity like this.

    i think more awareness about this project needs to be made..i am excited to also start making meaningful contributions to the community