4 ms·
I appreciate your answer and your take on the subject, I'll definitely check VoVNet2 out. I think the practice of taking whatever model is currently achieving S
by quibono 6y ago
I appreciate your answer and your take on the subject, I'll definitely check VoVNet2 out. I think the practice of taking whatever model is currently achieving SotA results and fine-tuning it on one's data is bringing results, but at the same time it makes it too easy to just blindly apply those models.
I think part of the problem is that it's currently too easy to reach for the big models - why train one from scratch where you can just tweak a few layers and get better results.
- m0zg 6y agoThat's the thing, if your dataset is not similar to the dataset on which your model was pre-trained (which it often isn't), the initial weights may actually prove to be either worthless, or even detrimental. That is, they may be worse than random initialization.