Pranav (@0xprv) 's Twitter Profile
Pranav

@0xprv

ML • Low level • Blockchain • defence tech • Winner of DIF'24 hackathon @DecentralizedID | Sharing the learnings, i love to learn

ID: 1773621435199586304

linkhttp://pranavvishnoi.com calendar_today29-03-2024 08:01:27

504 Tweet

92 Takipçi

32 Takip Edilen

Pranav (@0xprv) 's Twitter Profile Photo

Just a small note, The default weights of nn.Linear is uniform distribution, sampled from -sqrt(k) to sqrt(k) where k = 1/n_input_features

Just a small note,

The default weights of nn.Linear is uniform distribution, sampled from -sqrt(k) to sqrt(k) where k = 1/n_input_features
Pranav (@0xprv) 's Twitter Profile Photo

Some Fundamentals: Diff b/w Parallelism & Concurrency: Parallelism: Doing 2 or more tasks at the exact same time. Concurrency: Looks like doing many tasks at same time, but actually doing one task a little, then switching fast to another, then another, and so on.

Some Fundamentals:

Diff b/w Parallelism & Concurrency:

Parallelism: Doing 2 or more tasks at the exact same time. 

Concurrency: Looks like doing many tasks at same time, but actually doing one task a little, then switching fast to another, then another, and so on.
Pranav (@0xprv) 's Twitter Profile Photo

The PJSUA2 SIP library is great, but it's also very difficult to work with. I'm trying to integrate a large language model (LLM) that can have conversations with people over VoIP, but managing SIP is quite challenging.

Pranav (@0xprv) 's Twitter Profile Photo

I think the voice quality in VoIP should be improved. Currently, AI TTS systems that send audio chunks through VoIP calls use 8000 Hz sampling rate and mono channel, which degrades the quality of the AI-generated voice.

Pranav (@0xprv) 's Twitter Profile Photo

I think if someone wants to build a pipeline that converts audio to text and then back from text to audio, while maintaining both quality and speed,they should use the following : Fast Whisper for speech-to-text MeloTTS + OpenVoice V2 for advanced TTS with voice cloning

Pranav (@0xprv) 's Twitter Profile Photo

Why Voip only support mono channel and 8K hrtz voice transmission, on call, i think it should be increased to 44K with stereo channel

Pranav (@0xprv) 's Twitter Profile Photo

Just a quick practice in ml: use torch.load(PATH, mmap=True) to speed up PyTorch model loading. setting mmap=True enables memory-mapped files, mapping file regions directly to virtual memory. This avoids slow I/O operations by accessing data via memory addresses

Just a quick practice in ml: 

use torch.load(PATH, mmap=True) to speed up PyTorch model loading. 

setting mmap=True enables memory-mapped files, 

mapping file regions directly to virtual memory. This avoids slow I/O operations by accessing data via memory addresses