int32 error compiling for stm32f103c series

[cesarares]

compiling in Arduino IDE for stm32f103c series generate this error: libmaple_types.h:49:13: error: 'int32' has a previous declaration as 'typedef int int32'

[eloquentarduino]

I cannot test because I don't have a board, but at least this fix makes the Sine example compile.

Go to Arduino/libraries/EloquentTinyML/src/tensorflow/lite/kernels/internal/compatibility.h line 84. Replace

using int32 = std::int32_t;

with

#define int32 foo
using int32 = std::int32_t;

I won't integrate this hack into the library because I don't know if it can break somethings. Please report back your results.

[cesarares]

We have a small project where we are testing ML algorithms in small microcontrollers. In particular we are testing a convolutional network model to recognize digits in 8x8 image. The sketch compiled fine for the stm32f103c8t6 (Arm Cortex-M3) development board but the code exceeded the rom by 26.6KiB. We ended up successfully implementing the functions in C to do the inference (we were surprised that our implementation required almost half the ram and rom of this library). We also tested the modification you proposed with an ESP32 development board and It compiled and it worked correctly. I leave below some measurements that we made on the model and library in case they may be useful to you.

Library	MCU (Dev Board)	Data Mem.	Prog. Mem.	Inference Time	Accuracy
Eloquent TinyML	Xtensa LX6 (Esp32)	23.12 KiB	470.6 KiB	2270 us	97,80%
C Implementation	Xtensa LX6 (Esp32)	14.21 KiB	277.1 KiB	606 us	97,80%
Eloquent TinyML	Tensilica L106 (Esp8266)	51.20 KiB	391.5 KiB	11167 us	97,80%
C Implementation	Tensilica L106 (Esp8266)	31.01 KiB	274.0 KiB	7568 us	97,80%
C Implementation	Stm32f103  (Arm Cortex-M3)	2.14 KiB	27.3 KiB	8624 us	97,80%