doing something
712 subscribers
913 photos
36 videos
36 files
1.95K links
@smlkw doing something, just my notes to keep updates
Download Telegram
Made a colab to test Phi-3.5-vision-instruct model from Microsoft.

In the example, it summarizes frames in a video.

In short:

- Loading the model: ~8 GB
~ Inference peaks: up to 10 GB (for 10 images)

https://colab.research.google.com/drive/1pWgjVlkGpKnBUzfSFBnbOl9YWWk8kui5?usp=sharing

#ai #colab
👍1
doing something
Made a colab to test Phi-3.5-vision-instruct model from Microsoft. In the example, it summarizes frames in a video. In short: - Loading the model: ~8 GB ~ Inference peaks: up to 10 GB (for 10 images) https://colab.research.google.com/drive/1pWgjVlkGpK…
This media is not supported in your browser
VIEW IN TELEGRAM
First 10 seconds summarized as:

In brief, the deck includes the following parts:

- Outdoor Performance:
The presentation revolves around a drummer performing on a rooftop, with a cityscape in the background. The drummer is seated on a red drum set, with a black drum throne and a cymbal stand. The performance is accompanied by a speaker system. The rooftop setting is enclosed by a metal railing, with a glass barrier separating the performer from the audience. The audience includes a man and a woman, who are engaged in conversation. The presentation also includes a slide with a blurred image of a person walking past a drum set.


#demo
🔥1
Rust feeds to listen to used in my node with https://github.com/egorsmkv/feed-supermaster

#rust
Updated https://github.com/egorsmkv/feed-supermaster

Now on the latest Go release and with previews for messages, it's convenient

Also, checked the code with dev tools and fixed some potential bugs

#go
Tested Phi-3 and Phi-3.5 models on Optical Character Recognition task:

- Phi-3 colab: https://colab.research.google.com/drive/1HFvSQ8ktCfkECqqN38Gu7C_zrRN1ldAz?usp=sharing
- Phi-3.5 colab: https://colab.research.google.com/drive/1P7KeiK4RxuhzXiMviPCjbsbhRNSlE3cl?usp=sharing

Phi-3 gives better results - it just returns the text (in first 2 images), Phi-3.5 explains what's on the image :/

#ai #colab

@doing_something
Learned that AdBlocker in my browser Brave is based on Rust

https://github.com/brave/adblock-rust

#rust
💩1