Replace TinyCNN with pico-faces on the hackathon branch

.github/workflows/build.yml

@@ -2,7 +2,8 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 # CI for the three-step flow (cbuild setup, create_ai_layer.py, cbuild)
-# followed by an FVP run, with each of the three toolchains.
+# followed by an FVP run that checks the generated face against the host
+# reference, with each of the three toolchains.
 name: Build and Run
 
 on:
@@ -17,11 +18,77 @@ env:
   TARGET: SSE-320-U85
 
 jobs:
+  # Step 1 and 2 of the flow, once: the AI layer is the same for every
+  # toolchain and both target-types. The model data (ai_layer/model_pte.c) is
+  # generated here rather than committed; the job also renders the host's
+  # fake-quant reference image of the boot demo for the FVP check below.
+  export:
+    name: Create the AI layer
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout repo
+        uses: actions/checkout@v7
+
+      - name: Install tools
+        uses: ARM-software/cmsis-actions/vcpkg@v1
+
+      # cbuild setup of CMSIS-Toolbox 2.15 configures the CMake project, which
+      # probes the solution's compiler (AC6).
+      - name: Activate Arm tool license
+        uses: ARM-software/cmsis-actions/armlm@v1
+
+      - name: Cache packs
+        uses: actions/cache@v6
+        with:
+          key: cmsis-packs-${{ hashFiles('*.csolution.yml') }}
+          path: /home/runner/.cache/arm/packs
+
+      # torch + executorch is a multi-gigabyte install and otherwise dominates
+      # the job. Keyed on the requirements files so a pin change rebuilds it.
+      - name: Cache model-export venv
+        uses: actions/cache@v6
+        with:
+          key: venv-${{ runner.os }}-${{ hashFiles('requirements*.txt', 'setup_venv.py') }}
+          path: .venv
+
+      - name: Cache pico-faces checkpoints
+        uses: actions/cache@v6
+        with:
+          key: pico-faces-${{ hashFiles('setup_venv.py') }}
+          path: model/pico_faces
+
+      - name: Create model-export venv and fetch the checkpoints
+        run: ./setup_venv.sh
+
+      # The committed AI layer lists ai_layer/model_pte.c, which only step 2
+      # generates, and csolution refuses a layer whose files are missing. An
+      # empty placeholder lets step 1 run; create_ai_layer.py overwrites it.
+      - name: Generate the MLOps information
+        run: |
+          touch ai_layer/model_pte.c
+          cbuild setup ${SOLUTION}.csolution.yml --active ${TARGET} --packs
+
+      - name: Create the AI layer
+        run: python3 create_ai_layer.py ${SOLUTION}.cbuild-mlops.yml
+
+      - name: Render the host reference of the boot demo
+        run: python3 model/verify_export.py --stage fakequant --seed 3 --class 1 --w 4 --k 4 --out reference/pf_fakequant.png
+
+      - name: Upload the AI layer
+        uses: actions/upload-artifact@v4
+        with:
+          name: ai-layer
+          path: |
+            ai_layer/
+            reference/
+
   # Full build plus an FVP run, once per toolchain. Linux only: the FVP needs
   # an Arm user-based license, so matrixing this across hosts is not cheap.
   build-and-run:
     name: Build and run on Corstone-320 (${{ matrix.toolchain }})
     runs-on: ubuntu-latest
+    needs: export
 
     strategy:
       fail-fast: false
@@ -32,6 +99,11 @@ jobs:
       - name: Checkout repo
         uses: actions/checkout@v7
 
+      - name: Download the AI layer
+        uses: actions/download-artifact@v4
+        with:
+          name: ai-layer
+
       - name: Install tools
         uses: ARM-software/cmsis-actions/vcpkg@v1
 
@@ -41,7 +113,8 @@ jobs:
       # The FVP loads the runner's libpython3.11, whose standard library is
       # not where that build's prefix says; the model then dies at start-up
       # with "Fatal Python error: init_fs_encoding". PYTHONHOME (set on the
-      # FVP step only) points it at a complete Python 3.11.
+      # FVP step only) points it at a complete Python 3.11, which also runs
+      # the image comparison.
       - name: Set up Python 3.11 for the FVP
         uses: actions/setup-python@v5
         with:
@@ -53,25 +126,8 @@ jobs:
           key: cmsis-packs-${{ hashFiles('*.csolution.yml') }}
           path: /home/runner/.cache/arm/packs
 
-      # torch + executorch is a multi-gigabyte install and otherwise dominates
-      # the job. Keyed on the requirements files so a pin change rebuilds it.
-      - name: Cache model-export venv
-        uses: actions/cache@v6
-        with:
-          key: venv-${{ runner.os }}-${{ hashFiles('requirements*.txt', 'setup_venv.py') }}
-          path: .venv
-
-      - name: Create model-export venv
-        run: ./setup_venv.sh
-
-      - name: Generate the MLOps information
-        run: cbuild setup ${SOLUTION}.csolution.yml --active ${TARGET} --packs
-
-      - name: Create the AI layer
-        run: python3 create_ai_layer.py ${SOLUTION}.cbuild-mlops.yml
-
       - name: Build
-        run: cbuild ${SOLUTION}.csolution.yml --active ${TARGET} --toolchain ${{ matrix.toolchain }}
+        run: cbuild ${SOLUTION}.csolution.yml --active ${TARGET} --toolchain ${{ matrix.toolchain }} --packs
 
       - name: Run on the FVP
         env:
@@ -80,7 +136,7 @@ jobs:
           IMAGE=$(find out -name "${SOLUTION}.axf" -o -name "${SOLUTION}.elf" | head -1)
           echo "Running ${IMAGE}"
           # stdio is retargeted to semihosting, so the output arrives on the
-          # FVP's own stdout.
+          # FVP's own stdout; the application ends the simulation itself.
           FVP_Corstone_SSE-320 \
               -f board/Corstone-320/fvp_config.txt \
               -a "${IMAGE}" \
@@ -90,23 +146,33 @@ jobs:
           echo "Checking simulation output"
           grep -q "Test_result: PASS" fvp_stdout.log
 
+      # The run writes the generated image to out/fvp_image.bin through
+      # semihosting; it has to match what the quantized graphs compute on the
+      # host (the NPU's integer arithmetic differs from the fake-quant graphs
+      # only by rounding).
+      - name: Compare the image with the host reference
+        run: |
+          python -m pip install --quiet numpy pillow
+          python model/verify_export.py --compare out/fvp_image.bin reference/pf_fakequant.png --min-psnr 35
+
       - name: Upload artifacts
         if: always()
         uses: actions/upload-artifact@v4
         with:
           name: corstone-320-${{ matrix.toolchain }}
           path: |
             fvp_stdout.log
+            out/fvp_image.bin
+            out/fvp_result.txt
             out/**/*.axf
             out/**/*.elf
-            ai_layer/
 
   # The same solution for the Alif Ensemble E8 DevKit (target-type DevKit-E8),
-  # once per toolchain. Build only: the hardware is not in CI. The shipped
-  # ai_layer/ is used, so no venv is needed.
+  # once per toolchain. Build only: the hardware is not in CI.
   build-devkit-e8:
     name: Build for the Alif Ensemble E8 DevKit (${{ matrix.toolchain }})
     runs-on: ubuntu-latest
+    needs: export
 
     strategy:
       fail-fast: false
@@ -117,6 +183,11 @@ jobs:
       - name: Checkout repo
         uses: actions/checkout@v7
 
+      - name: Download the AI layer
+        uses: actions/download-artifact@v4
+        with:
+          name: ai-layer
+
       - name: Install tools
         uses: ARM-software/cmsis-actions/vcpkg@v1
 
@@ -126,14 +197,11 @@ jobs:
       - name: Cache packs
         uses: actions/cache@v6
         with:
-          key: cmsis-packs
+          key: cmsis-packs-${{ hashFiles('*.csolution.yml') }}
           path: /home/runner/.cache/arm/packs
 
-      - name: Generate the MLOps information
-        run: cbuild setup ${SOLUTION}.csolution.yml --active DevKit-E8 --packs
-
       - name: Build
-        run: cbuild ${SOLUTION}.csolution.yml --active DevKit-E8 --toolchain ${{ matrix.toolchain }}
+        run: cbuild ${SOLUTION}.csolution.yml --active DevKit-E8 --toolchain ${{ matrix.toolchain }} --packs
 
       - name: Upload artifacts
         if: always()
@@ -142,6 +210,7 @@ jobs:
           name: devkit-e8-${{ matrix.toolchain }}
           path: |
             out/**/*.axf
+            out/**/*.elf
             out/**/*.hex
 
   # The export flow must work on all three host OSes -- that is the point of a
@@ -186,6 +255,8 @@ jobs:
           # No cbuild-mlops.yml without cbuild; importing the script checks
           # that it parses under this interpreter.
           "${PY}" -c "import create_ai_layer; print('create_ai_layer.py OK')"
+          # The checkpoints setup_venv.py downloaded load on this host.
+          "${PY}" -c "import sys; sys.path.insert(0, 'model'); import model; model.load_checkpoint(); print('checkpoints OK')"
 
       - name: Report resolved versions
         shell: bash

.gitignore

@@ -3,9 +3,15 @@
 __pycache__/
 *.pyc
 
-# The exported ExecuTorch program. create_ai_layer.py writes it next to the
-# generated AI layer; the C array in ai_layer/model_pte.c is what gets built.
+# The exported ExecuTorch program and its C array. create_ai_layer.py writes
+# both into the generated AI layer; at ~2.8 MB (a ~14 MB C source) they are
+# regenerated by "Create AI layer" rather than committed. The clayer and the
+# headers next to them are committed.
 *.pte
+ai_layer/model_pte.c
+
+# The pico-faces checkpoints setup_venv.py downloads (see PICO_FACES_FILES).
+model/pico_faces/
 
 # CMSIS-Toolbox build outputs. RTE/ is tracked in full: the layers' configuration
 # files and the per-context RTE/_<build>_<target>/ headers the toolbox writes.

.vscode.d/tasks.json

@@ -29,7 +29,7 @@
         },
         {
             "label": "Create AI layer",
-            "detail": "Export model/model.py for the NPU in cmsis-executorch.cbuild-mlops.yml and write ai_layer/ (run a CMSIS build or 'cbuild setup' first)",
+            "detail": "Export model/model.py (pico-faces) for the NPU in cmsis-executorch.cbuild-mlops.yml and write ai_layer/ (run a CMSIS build or 'cbuild setup' first; takes a few minutes)",
             "type": "shell",
             "command": "python3",
             "args": ["create_ai_layer.py", "cmsis-executorch.cbuild-mlops.yml"],
@@ -43,12 +43,14 @@
         },
         {
             "label": "Alif: Install M55_HP debug stubs (DevKit-E8, single core configuration)",
-            "detail": "Programs the ATOC/MRAM stubs with Alif SETOOLS so J-Link can load and debug the M55_HP core. Overwrites build/config/M55_HP_mram_cfg.json and build/images/M55_HP_mram_stub.bin in the SETOOLS tree, then runs app-gen-toc and app-write-mram; stops at the first failing step. Set SW4 to SEUART, set 'alif.setools.root' in your VS Code settings.",
+            "detail": "Programs the ATOC/MRAM stubs with Alif SETOOLS so J-Link can load and debug the M55_HP core. Selects the DevKit-E8's part in SETOOLS first (E8 AE822FA0E5597LS0, revision A0): app-gen-toc writes it into the ATOC, and the Secure Enclave does not boot an ATOC built for another part. The part and revision are written into SETOOLS' utils/global-cfg.db before tools-config runs, because tools-config refuses to start while another project's 'tools-config -p' without '-r' has left a revision the part does not have ('Revision is invalid!'). Overwrites build/config/M55_HP_mram_cfg.json and build/images/M55_HP_mram_stub.bin in the SETOOLS tree, then runs app-gen-toc and app-write-mram; stops at the first failing step. Set SW4 to SEUART, set 'alif.setools.root' in your VS Code settings.",
             "type": "shell",
             "command": [
                 "cp './.alif/M55_HP_mram_cfg.json' '${config:alif.setools.root}/build/config/M55_HP_mram_cfg.json' &&",
                 "cp './.alif/M55_HP_mram_stub.bin' '${config:alif.setools.root}/build/images/M55_HP_mram_stub.bin' &&",
                 "cd '${config:alif.setools.root}' &&",
+                "sed -i.bak -E -e 's|(\"Part#\": *\")[^\"]*|\\1E8 (AE822FA0E5597LS0) - 5.5 MRAM / 9.75 SRAM|' -e 's|(\"Revision\": *\")[^\"]*|\\1A0|' utils/global-cfg.db &&",
+                "./tools-config -p 'E8 (AE822FA0E5597LS0) - 5.5 MRAM / 9.75 SRAM' -r A0 &&",
                 "./app-gen-toc -f 'build/config/M55_HP_mram_cfg.json' &&",
                 "./app-write-mram -p ${input:discoverCOM}"
             ],
@@ -57,6 +59,8 @@
                     "cp './.alif/M55_HP_mram_cfg.json' '${config:alif.setools.root}/build/config/M55_HP_mram_cfg.json'; if (-not $?) { exit 1 };",
                     "cp './.alif/M55_HP_mram_stub.bin' '${config:alif.setools.root}/build/images/M55_HP_mram_stub.bin'; if (-not $?) { exit 1 };",
                     "cd '${config:alif.setools.root}'; if (-not $?) { exit 1 };",
+                    "(Get-Content utils/global-cfg.db -Raw) -replace '(\\x22Part#\\x22:\\s*\\x22)[^\\x22]*', '$1E8 (AE822FA0E5597LS0) - 5.5 MRAM / 9.75 SRAM' -replace '(\\x22Revision\\x22:\\s*\\x22)[^\\x22]*', '$1A0' | Set-Content utils/global-cfg.db -NoNewline; if (-not $?) { exit 1 };",
+                    "./tools-config -p 'E8 (AE822FA0E5597LS0) - 5.5 MRAM / 9.75 SRAM' -r A0; if (-not $?) { exit 1 };",
                     "./app-gen-toc -f 'build/config/M55_HP_mram_cfg.json'; if (-not $?) { exit 1 };",
                     "./app-write-mram -p ${input:discoverCOM}"
                 ],
@@ -112,7 +116,7 @@
         {
             "id": "setupPythonVersion",
             "type": "promptString",
-            "description": "Python version for uv (>=3.10,<3.15), e.g. 3.12 or 3.12.10",
+            "description": "Python version for uv (>=3.10,<3.14), e.g. 3.12 or 3.12.10",
             "default": "3.12"
         }
     ]