-
|
I have a specific use case where I should use Float32 with a small model. Can I force hls4ml to generate an hls project using floats for my model? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Not at the moment. This used to be possible with some caveats, but was not implemented in the newer codebase. Perhaps in the future. What you can do is manually change the precision data types in the generated |
Beta Was this translation helpful? Give feedback.
Not at the moment. This used to be possible with some caveats, but was not implemented in the newer codebase. Perhaps in the future. What you can do is manually change the precision data types in the generated
defines.handparameters.hfiles, but you may run into issues with some layers (pooling comes to mind, perhaps others). I hope you have a strong incentive for doing this because floating-point operations on FPGAs come with lots of issues related to latency. Quantization is the way to go.