Tinkering with machine learning tools
Trying out the latest ML tools in Google Colab
On one hand we hear that advances in AI are controlled by few giant companies — MS via OpenAI, Google and so on. On the other hand the open source community is creating and sharing many tools that is democratising ML/AI application development. Some simple experiments I could carry out without spending any money on API-s
Summarising YouTube Videos
A friend shared this philosophical YouTube video with me. The video is around 2 hours long.
Used the following modules to summarise it
- pytube to extract the audio stream
- whisper to transcribe to a 120 KB text file (high accuracy)
- facebook/bart-large-cnn to summarise from the transcribed text file
The transcript
You have your finger on something that I am haunted by, which is that people are lost, they’re scared in life, they’ve got a negative voice running on repeat. When they get into self-help, they make no lasting changes whatsoever. What’s the answer to that? And how does the subtle art of not giving a fuck play into all this? Starting with the easy questions. I hope not. Well, you know, a question that I’ve been in this industry for 15 years now. And I’ve written thousands of pages worth of content. I’ve recorded tons of videos, done hundreds and hundreds of interviews like this. So most of my career is I’ve existed in the theory…. continues
The summary
The author has written thousands of pages worth of content and recorded tons of videos. He says the only intervention that consistently achieves behavioral change is a good therapist over many years. “I’m very good at changing my mind about things,” he says. I think one of the big things is people have it’s it’s twofold. They have an inability to build desire so they just don’t want anything in their life. And then they don’t know how to use identity statements to hold themselves accountable. I find it very easy to point myself at something new, run a process. And that process then anchors me emotionally. If you are receiving social validation from people for bad behaviors, then one of the most effective things you can do is surround yourself with people who are going to validate you for better behaviors. The most fundamental way to kind of rewire what you care about or what you want in your life is to align yourself with social groups. The internet is solving solving the implementation side, which means solving the validation side. Using the internet to kind of do what you were just describing of like finding communities, finding that desert island group of like, okay, these are my people were together for this explicit reason. I think the most psychologically healthy thing to do is to simply accept all emotional experiences, regardless of whether they feel good or feel bad. …. continues
My comments: Whisper is amazing. GPT4 might perform better at summarisation. The token size limit for summarisation models is a limitation, there might be some method to split the input at the right moments so that the summarisation has the right context
Image editing using text prompts
I used the Grounded-Segment-Anything project to try this out. It connects the following modules
- GoundingDINO for object detection from textual prompt
- segment-anything for segmenting the object
- stable-diffusion-inpainting for replacing the object with a new one
Tried replacing a “clock” with “a bluetooth speaker”. The result looked okay.
Replacing with “a coiled snake” did not work too well :)
Finally generating a hero image for this article
Passed the prompt “tinkering with machine learning tools” to runwayml/stable-diffusion-v1–5
Other interesting projects to check out
