Expand model sources and validate a real network corpus

.github/workflows/collection.yml

@@ -12,7 +12,8 @@ permissions:
 jobs:
   collection:
     name: Base installation / Python ${{ matrix.python }}
-    runs-on: ubuntu-latest
+    # Newer runner images no longer provide the Python 3.9 compatibility target.
+    runs-on: ubuntu-22.04
     strategy:
       matrix:
         python: ['3.9', '3.12']

.github/workflows/model-sources.yml

@@ -0,0 +1,132 @@
+name: Real network sources
+
+on:
+  pull_request:
+    branches: [master, main]
+    paths: ['src/**', 'tests/**', 'configs/**', 'catalogs/**', 'pyproject.toml', '.github/workflows/model-sources.yml']
+  push:
+    branches: [master, main]
+    paths: ['src/**', 'tests/**', 'configs/**', 'catalogs/**', 'pyproject.toml', '.github/workflows/model-sources.yml']
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  source-smoke:
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Install fixed CPU export environment
+        run: |
+          python -m pip install torch==2.9.1 torchvision==0.24.1 --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install -e '.[networks,test]' -c configs/constraints-networks.txt
+          python -m pip check
+      - name: Check all source adapters and PyTorch/ONNX parity
+        env:
+          OMP_NUM_THREADS: '1'
+          OPENBLAS_NUM_THREADS: '1'
+          MKL_NUM_THREADS: '1'
+        run: |
+          python -m pytest tests/unit tests/operators -q --junitxml=.validation/network-tests.xml
+          infinitensor-test-expand --manifest configs/network-source-smoke.json --data-dir .validation/source-smoke --offline --minimum-real-models 3 --max-size-mb 160 --jobs 1
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: real-network-source-evidence
+          include-hidden-files: true
+          path: |
+            .validation/network-tests.xml
+            .validation/source-smoke/expansion/inventory.json
+            .validation/source-smoke/expansion/results.json
+            .validation/source-smoke/expansion/batches/**/worker-*.log
+
+  diverse-smoke:
+    runs-on: ubuntu-latest
+    timeout-minutes: 25
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Install fixed CPU multi-task environment
+        run: |
+          python -m pip install torch==2.9.1 torchvision==0.24.1 torchaudio==2.9.1 --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install -e '.[diverse,test]' -c configs/constraints-diverse.txt
+          python -m pip check
+      - name: Validate real multi-task networks and input contracts
+        env:
+          OMP_NUM_THREADS: '1'
+          OPENBLAS_NUM_THREADS: '1'
+          MKL_NUM_THREADS: '1'
+        run: |
+          python -m pytest tests/integration/test_diverse_sources.py -q --junitxml=.validation/diverse-tests.xml
+          infinitensor-test-expand --manifest configs/diverse-source-smoke.json --data-dir .validation/diverse-smoke --offline --minimum-real-models 6 --max-size-mb 768 --jobs 1
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: diverse-network-source-evidence
+          include-hidden-files: true
+          path: |
+            .validation/diverse-tests.xml
+            .validation/diverse-smoke/expansion/inventory.json
+            .validation/diverse-smoke/expansion/results.json
+            .validation/diverse-smoke/expansion/batches/**/worker-*.log
+
+  paddle-smoke:
+    runs-on: ubuntu-latest
+    timeout-minutes: 25
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Install independent CPU Paddle environment
+        shell: bash
+        run: |
+          # onnxoptimizer 0.3.13 has no cp312 wheel and vendors CMake 3.1 policies.
+          python -m pip install cmake==3.31.10
+          # PEP 517 isolation hides the cmake Python module from its wrapper.
+          export PATH="$(python -c 'import cmake; print(cmake.CMAKE_BIN_DIR)'):$PATH"
+          cmake --version
+          if ! python -m pip install -e '.[paddle,test]' -c configs/constraints-paddle.txt 2>&1 | tee paddle-install.log; then
+            python - <<'PY'
+          from pathlib import Path
+          lines = Path('paddle-install.log').read_text().splitlines()
+          starts = [i for i, line in enumerate(lines) if any(term in line for term in ('CMake Error', 'ModuleNotFoundError', 'Traceback'))]
+          detail = [line for i in starts[:4] for line in lines[max(0, i-2):i+16]]
+          text = '\n'.join(detail + lines[-30:])
+          text = text.replace('%', '%25').replace('\r', '%0D').replace('\n', '%0A')
+          print('::error title=Paddle dependency installation failed::' + text)
+          PY
+            exit 1
+          fi
+      - name: Check independent Paddle dependencies
+        run: python -m pip check
+      - name: Prepare pinned Paddle sources
+        run: |
+          infinitensor-test-paddle-setup --directory .validation/paddle-repos
+      - name: Validate Paddle exports and four real tasks
+        env:
+          ITEST_PADDLE_REPOS_DIR: ${{ github.workspace }}/.validation/paddle-repos
+          OMP_NUM_THREADS: '1'
+          OPENBLAS_NUM_THREADS: '1'
+          MKL_NUM_THREADS: '1'
+        run: |
+          python -m pytest tests/integration/test_paddle_exports.py -q --junitxml=.validation/paddle-tests.xml
+          infinitensor-test-expand --manifest configs/paddle-source-smoke.json --data-dir .validation/paddle-smoke --offline --minimum-real-models 4 --max-size-mb 768 --jobs 2 --timeout 300
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: paddle-network-source-evidence
+          include-hidden-files: true
+          path: |
+            paddle-install.log
+            .validation/paddle-tests.xml
+            .validation/paddle-smoke/expansion/inventory.json
+            .validation/paddle-smoke/expansion/results.json
+            .validation/paddle-smoke/expansion/batches/**/worker-*.log

.gitignore

@@ -18,6 +18,7 @@ htmlcov/
 venv/
 env/
 .venv/
+.venv-networks/
 
 # IDE
 .idea/
@@ -31,6 +32,10 @@ data/models/
 data/cache/
 data/results/
 data/reports/
+data/real-networks/
+data/paddle-repos/
+data/paddle-networks/
+.venv-paddle/
 *.log
 
 # Model files (NEVER commit)

README.md

@@ -1,5 +1,12 @@
 # InfiniTensor-Test
 
+The [general corpus](catalogs/real-networks.json) and [Paddle corpus](catalogs/paddle-networks.json)
+contain **1,422 deduplicated network configurations across 21 tasks** (1,409 + 13),
+including OCR, table structure, image matting, text, audio, video and optical flow,
+validated with ONNX checks, ONNX Runtime and all-output PyTorch/ONNX comparison.
+The 548 synthetic positives and 18 negatives are counted separately. See
+[multi-task sources](docs/task-expansion.md) and [Paddle sources and replay commands](docs/paddle-models.md).
+
 This change combines lazy sources, reproducible package collection and model
 validity repairs: 548 synthetic positives must pass ONNX checks and CPU reference
 execution; 18 declared negatives must be rejected for their specified reasons.

README.zh-CN.md

@@ -1,5 +1,10 @@
 # InfiniTensor-Test
 
+[通用清单](catalogs/real-networks.json)与[飞桨清单](catalogs/paddle-networks.json)合计
+**1422 个去重网络配置、21 种任务**（1409 + 13），新增 OCR、表格结构、图像抠图等任务，
+全部通过 ONNX 检查、参考执行及原始框架/ONNX 全部输出比较。原有 548 个合成正例与 18 个负例另计。
+安装与复验见[多任务网络说明](docs/task-expansion.md)和[飞桨模型接入说明](docs/paddle-models.md)。
+
 本分支合并采集基础与模型有效性修复：来源按需加载、确定性完整包缓存，
 以及 548 个合成正例的 ONNX 检查和 CPU 参考执行、18 个明确负例的按原因拒绝验证。
 请先阅读[能力提升及复验命令](docs/pr-t01-collection.md)与[PR 审查说明](docs/pr-t01-review.md)。