I haven’t quantized the model yet, but it shows 8-bit for every single layer. Is this correct, or am I using the wrong command for the profiler?
The command I used was: hailo profiler XXX(model before quantized).har.
I also tried using the Hailo profiler tool after applying the model script (to check if some layers use 16-bit weights), but I couldn’t see any changes. Only the compiled HAR file showed the correct results.
When you profile a model before quantization using the command hailo profiler XXX(model before quantized).har, it’s typical to see 8-bit for every layer. This happens because the profiler provides a preview of how the model would be quantized by default.
To check if certain layers are using 16-bit weights, you need to profile the model after it has been quantized and compiled. The fact that you see the correct results only in the compiled HAR file aligns with this explanation.
To accurately view the quantized model, including any 16-bit layers, Profile the compiled HAR file.
