hello,
I am in the process of optimizing and quantizing a model for depth prediction, but even the smallest FP32 model variant is 150mb. I can either target the ESP32 or my PiZero, but id rather the ESP32 because it has many advantages for my particular project (and inference time is not a problem, even if it takes 20 minutes).
However, I DO need to know upfront whether its even worth the effort to quantize the model with ESP32 as target. So, I have an FP32 model at 150mb and an FP16 model at 75mb. lets say int8 quantization gets me a 4x size reduction. so we're looking at 37.5mb and 18.75mb for the quantized FP32 and FP16 models, respectively.
I only have an ESP32S3 with 8mb PSRAM. so, that leaves me with two final questions:
1. Does the ESP32 allow for (slow) VRAM swap space to be allocated on the SD card like linux does?
2. if not, i've read that ESP32S3 supports a maximum of 32mb PSRAM. does anyone know of a manufacturer actually making that configuration at the moment?
thanks