A proof-of-concept for training tinygrad models on the edge.

Works by leveraging the tinygrad C backend to generate c code for the entire train step.

run CLANG=1 python compile.py first to export the kernels.

drop a config.h file in esp32/main/ with the following contents:

#define WIFI_SSID "your wifi ssid"
#define WIFI_PWD "your wifi password"
#define SERVER_IP "dataserver address"
#define SERVER_PORT "dataserver port"

then build the esp-idf project in esp32/ and flash.

make sure the data server is running on your computer.

python dataserver.py

There isn't enough flash on the esp32 to actually store the entire dataset, so we are streaming it from somewhere.