Summary: Pico2 demo of simple neural network (MNIST)

examples/raspberry_pi/pico2/CMakeLists.txt

@@ -64,11 +64,14 @@ set(MODEL_STAMP "${CMAKE_CURRENT_BINARY_DIR}/model_pte.stamp")
 
 add_custom_command(
   OUTPUT ${MODEL_STAMP}
-  COMMAND python3 ${EXECUTORCH_ROOT}/executorch/examples/rpi/pte_to_array.py
-          --model ${INPUT_MODEL} --file ${MODEL_PTE_C}
+  COMMAND
+    python3
+    ${EXECUTORCH_ROOT}/executorch/examples/raspberry_pi/pico2/pte_to_array.py
+    --model ${INPUT_MODEL} --file ${MODEL_PTE_C}
   COMMAND ${CMAKE_COMMAND} -E touch ${MODEL_STAMP}
-  DEPENDS ${INPUT_MODEL}
-          ${EXECUTORCH_ROOT}/executorch/examples/rpi/pte_to_array.py
+  DEPENDS
+    ${INPUT_MODEL}
+    ${EXECUTORCH_ROOT}/executorch/examples/raspberry_pi/pico2/pte_to_array.py
   COMMENT "Injecting PTE data from '${INPUT_MODEL}' into model_pte.c"
   VERBATIM
 )

examples/raspberry_pi/pico2/README.md

@@ -1,51 +1,142 @@
 # Overview
-This document outlines the steps required to run a simple Add Module on the Pico2 microcontroller using executorch.
 
-## (Pre-requisistes) Prepare the Environment for Arm
+This document outlines the steps required to run a simple MNIST digit recognition neural network on the Pico2 microcontroller using ExecuTorch.
 
-1. Setup executorch development environment, Also see  <a href="https://docs.pytorch.org/executorch/main/tutorial-arm-ethos-u.html#software"/> for instructions on setting up the environment for Arm.
-2. Make sure you have the toolchain configured correctly.
+## Demo Model: Hand-crafted MNIST Classifier
+
+The included `export_mlp_mnist.py` creates a demonstration model with hand-crafted weights (not production-trained). This tiny MLP recognizes digits 0, 1, 4, and 7 using manually designed feature detectors.
+Note: This is a proof-of-concept. For production use, train your model on real MNIST data.
+
+## Bring Your Own Model
+
+This demo demonstrates ExecuTorch's ability to bring your own PyTorch model and deploy it to Pico2 with one simple script. The complete pipeline works from any PyTorch model to a runnable binary:
+
+### Train your PyTorch model
+
+Export using `torch.export()` and `to_edge()`
+Build firmware with one command
+Deploy directly to Pico2
+
+#### Important Caveats:
+
+- Memory constraints - Models must fit in 520KB SRAM
+- Missing operators - Some ops may not be supported
+- Selective builds - Include only operators your model uses
+
+## Memory Constraints & Optimization
+
+- Critical: Pico2 has limited memory:
+- 520KB SRAM (on-chip static RAM)
+- 4MB QSPI Flash (onboard storage)
+
+### Always apply optimization techniques on large models that do not fit in Pico2 memory:
+
+Large models will not fit. Keep your `.pte` files small!
+- Quantization (INT8, INT4)
+- Model pruning
+- Operator fusion
+- Selective builds (include only needed operators)
+For more details , refer to the [ExecuTorch Quantization Optimization Guide](https://docs.pytorch.org/executorch/1.0/quantization-optimization.html), [Model Export & Lowering](https://docs.pytorch.org/executorch/1.0/using-executorch-export.html) and [Selective Build support](https://docs.pytorch.org/executorch/1.0/kernel-library-selective-build.html)
+
+## (Prerequisites) Prepare the Environment for Arm
+
+Setup executorch development environment. Also see instructions for setting up the environment for Arm.
+Make sure you have the toolchain configured correctly. Refer to this [setup](https://docs.pytorch.org/executorch/1.0/backends-arm-ethos-u.html#development-requirements) for more details.
 
 ```bash
 which arm-none-eabi-gcc
---> return something like executorch/examples/arm/ethos-u-scratch/arm-gnu-toolchain-13.3.rel1-x86_64-arm-none-eabi/bin/arm-none-eabi-gcc
+# Should return: executorch/examples/arm/ethos-u-scratch/arm-gnu-toolchain-13.3.rel1-x86_64-arm-none-eabi/bin/arm-none-eabi-gcc
 ```
 
-## Build Pico2 Firmware with Executorch
+## Build Pico2 Firmware with ExecuTorch
+
+This involves two steps:
 
-This step involves two sub steps
+### Generate your model:
+
+```bash
+cd examples/raspberry_pi/pico2
+python export_mlp_mnist.py # Creates balanced_tiny_mlp_mnist.pte
+```
 
-1. Cross Compile Executorch for Arm Cortex M, Pico2 target
-2. Build the firmware with the input model provided (If not provided, it will use the default_model.pte)
+### Build firmware:
 
-Use the following command to build the firmware:
-``` bash
-executorch/examples/rpi/build_firmware_pico.sh --model=<path_to_model.pte>
+```bash
+# In the dir examples/raspberry_pi/pico2
+build_firmware_pico.sh --model=balanced_tiny_mlp_mnist.pte # This creates executorch_pico.uf2, a firmware image for Pico2
 ```
 
 ### Flash Firmware
 
-Hold the BOOTSEL button on the Pico2 and connect it to your computer; it will mount as RPI-RP2. Copy the executorch_pico.uf2 file to this drive.
+Hold the BOOTSEL button on Pico2 and connect to your computer. It mounts as `RPI-RP2`. Copy `executorch_pico.uf2` to this drive.
 
 ### Verify Execution
 
-Check that the Pico2's LED blinks 10 times at 500 ms interval to confirm successful firmware execution.
-The Pico2's LED should blink 10 times at 500 ms intervals, indicating successful firmware execution. If connected via serial, you should see:
-
+The Pico2 LED blinks 10 times at 500ms intervals for successful execution. Via serial terminal, you'll see:
 ```bash
-Method loaded [forward]
-Output: 13.000000, 136.000000, 24.000000, 131.000000
+...
+...
+PREDICTED: 4 (Expected: 4) ✅ CORRECT!
+
+==================================================
+
+=== Digit 7 ===
+############################
+############################
+                        ####
+                       ####
+                      ####
+                     ####
+                    ####
+                   ####
+                  ####
+                 ####
+                ####
+               ####
+              ####
+             ####
+            ####
+           ####
+          ####
+         ####
+        ####
+       ####
+      ####
+     ####
+    ####
+   ####
+  ####
+ ####
+####
+###
+
+Input stats: 159 white pixels out of 784 total
+Running neural network inference...
+✅ Neural network results:
+  Digit 0: 370.000
+  Digit 1: 0.000
+  Digit 2: -3.000
+  Digit 3: -3.000
+  Digit 4: 860.000
+  Digit 5: -3.000
+  Digit 6: -3.000
+  Digit 7: 1640.000 ← PREDICTED
+  Digit 8: -3.000
+  Digit 9: -3.000
+
+� PREDICTED: 7 (Expected: 7) ✅ CORRECT!
+
+==================================================
+
+🎉 All tests complete! PyTorch neural network works on Pico2!
 ```
 
 ### Debugging via Serial Terminal
 
-On macOS or Linux, open a serial terminal with:
-
+On macOS/Linux:
 ```bash
 screen /dev/tty.usbmodem1101 115200
 ```
+Replace `/dev/tty.usbmodem1101` with your device path. If LED blinks 10 times at 100ms intervals, check logs for errors, but if it blinks 10 times at 500ms intervals, it is successful!
 
-Replace /dev/tty.usbmodem1101 with your device path. This terminal shows program logs and errors. If
-the LED blinks 10 times at 100 ms intervals, your program hit an error state—check the logs here.
-
-These steps complete running the simple model on Pico2 using ExecuTorch.
+Result: A complete PyTorch → ExecuTorch → Pico2 demo neural network deployment! 🚀

examples/raspberry_pi/pico2/build_firmware_pico.sh

@@ -9,11 +9,11 @@
 # build_firmware_pico.sh
 # Simple script to cross-compile ExecuTorch and build Pico2 firmware with optional model input
 
-set -e
+set -euo pipefail
 
 # Paths
-ROOT_DIR="$(cd "$(dirname "$0")/../.." && pwd)"  # examples/rpi/ -> root dir
-PICO2_DIR="${ROOT_DIR}/examples/rpi/pico2"
+ROOT_DIR="$(cd "$(dirname "$0")/../../.." && pwd)"  # examples/raspberry_pi/ -> root dir
+PICO2_DIR="${ROOT_DIR}/examples/raspberry_pi/pico2"
 BUILD_DIR="${PICO2_DIR}/build"
 EXECUTORCH_BUILD_DIR="${ROOT_DIR}/cmake-out"