| # Build and inspect |
|
|
| Use **ESP-IDF v5.5.1**, its **riscv32-esp-elf GCC 14.2.0_20241119** toolchain, |
| and **esptool 5.6.0**. The release was built with CMake 3.30.2 and Ninja 1.12.1. |
| The complete configuration is included as `source/firmware/sdkconfig`, alongside |
| `sdkconfig.defaults`. Reproducible-build mode and project version 1.0.0 are set. |
| |
| After installing/activating official ESP-IDF, from the repository root: |
| |
| ```sh |
| idf.py -C source/firmware build |
| ``` |
| |
| Build your own merged image with official esptool: |
| |
| ```sh |
| python -m esptool --chip esp32c3 merge-bin -o rebuilt-firmware.bin --flash-mode dio --flash-freq 40m --flash-size 4MB --pad-to-size 4MB 0x0 source/firmware/build/bootloader/bootloader.bin 0x8000 source/firmware/build/partition_table/partition-table.bin 0x10000 source/firmware/build/flower_c3.bin 0x70000 data/model-int8.bin |
| ``` |
| |
| This writes no device. Flash the resulting file at `0x0` using the same guide. |
| Reproducible-build mode removes variable build metadata; byte identity across |
| independent build environments has not been independently verified. |
| |
| `data/model-int8.json` describes the weights as signed INT8 OHWI tensors, with |
| byte offsets, little-endian INT32 biases, and power-of-two quantization exponents. |
| `source/generated/model-meta.h` contains layer descriptors, lookup tables, and the |
| startup latent. The source retains all trained values from `data/original.espdl`. |
| |
| For desktop inference with a C11 compiler: |
| |
| ```sh |
| cc -O3 -std=c11 -Isource/generated -Isource/native source/native/flower.c source/native/render.c source/native/host_main.c -o host-model |
| ./host-model data/model-int8.bin output-int8.bin 2000 |
| ``` |
| |
| The output is 320 × 176 interleaved signed INT8 RGB, plus a packed `.epaper` frame. |
| Seed 2000 has CRC-32 `534cfdc5`; the embedded startup seed has CRC-32 `4033cceb`. |
| Original seed-2000 float reference tensors and a seven-seed test report are in |
| `data/`. Training source is in `training/`; reproducing training requires the |
| upstream dataset and dependencies listed in `data/original-export.json`. |
| |
