3 ms·
> The weights aren’t binary. The weights are typically distributed as a binary blob containing IEEE floating point values, or fixed point values; and these num
by amelius 3y ago
> The weights aren’t binary.
The weights are typically distributed as a binary blob containing IEEE floating point values, or fixed point values; and these numbers are typically not meant for direct human consumption.
> The same code used to train the model / generate the weights will not produce the same model weights if run again, it is not deterministic.
This is not true, except in contrived situations like when you use PRNGs and you deliberately throw away their seeds.
> Also you will need millions of dollars to retrain
This is besides the point.