1. For my robotics tool, the response was really good. People enjoyed and even sent feedback and got a couple of issues opened on the repo too. I just think I need more reach for it.
2. Noted, thanks for the heads up!
You can quickly benchmark the model and compare it quickly by moving around/adding layers as you wish. But yeah your point makes sense too.
Makes sense. I too, know of many open-source projects in the same way. But when I compare with other open-source projects in the same niche which offer less features (as an example) getting more popularity. It just makes me think I am doing something wrong so that's why I asked.
0. Personal milestone really. Just want to see how much help I can be to others I guess. Hard to say.
1. One of my projects does almost the same stuff a proprietary software does so I am providing a free alternative. My other project is helpful in the sense that it does something no single-application does. Like there exists a site that can view ML models then there exists softwares that can edit and customize layers in a model. My tool does both in one. But yes you are right, maybe I am not selling my product in the right manner. Good idea to think about.
I got the most traction from the subreddit I posted it on. FYI I am measuring traction based on how many GitHub stars that repo got (which is a bad way to measure I know). But good idea, cheers!
The fact that this solves the cross browser transparent video problem alone makes it interesting. Usually you are forced to serve WebM with alpha for Chrome/Firefox and HEVC with alpha for Safari but even then, trying to build hover states by hand-timing media seeks is a nightmare
How does the learner forget old data? If a model's price drops by 50%, or a new model is swapped into an existing rung, the historical weights learned from the old pricing/capability will actively harm routing decisions. Is there a decay factor on the weights or something else?
For optimizers.py, are you relying purely on boolean arithmetic for these masks, or did you find a more elegant mathematical workaround to avoid the branching overhead entirely?