Arm64 CI setup with TravisCI (scikit-learn#17996)

.travis.yml

@@ -23,21 +23,31 @@ matrix:
     # installed from their CI wheels in a virtualenv with the Python
     # interpreter provided by travis.
     - python: 3.7
-      env: CHECK_WARNINGS="true"
+      env:
+        - CHECK_WARNINGS="true"
+        - CI_CPU_COUNT="3"
       if: type = cron OR commit_message =~ /\[scipy-dev\]/
-    
+
     # As above but build scikit-learn with Intel C compiler (ICC).
     - python: 3.7
       env:
         - CHECK_WARNING="true"
         - BUILD_WITH_ICC="true"
+        - CI_CPU_COUNT="3"
       if: type = cron OR commit_message =~ /\[icc-build\]/
 
+    - python: 3.7
+      env:
+        - CI_CPU_COUNT="8"
+      os: linux
+      arch: arm64
+      if: type = cron OR commit_message =~ /\[arm64\]/
+
 install: source build_tools/travis/install.sh
 script:
-  - bash build_tools/travis/test_script.sh
-  - bash build_tools/travis/test_docs.sh
-  - bash build_tools/travis/test_pytest_soft_dependency.sh
+  - bash build_tools/travis/test_script.sh || travis_terminate 1
+  - bash build_tools/travis/test_docs.sh || travis_terminate 1
+  - bash build_tools/travis/test_pytest_soft_dependency.sh || travis_terminate 1
 after_success: source build_tools/travis/after_success.sh
 notifications:
   webhooks:

build_tools/travis/install.sh

@@ -14,6 +14,12 @@
 set -e
 
 # Fail fast
+echo "CPU Arch: ${TRAVIS_CPU_ARCH}"
+
+# jq is used in travis_fastfail.sh, it's already pre-installed in non arm64
+# environments
+sudo apt-get install jq
+
 build_tools/travis/travis_fastfail.sh
 
 # Imports get_dep
@@ -35,28 +41,40 @@ ccache --max-size 100M --show-stats
 # If Travvis has language=generic, deactivate does not exist. `|| :` will pass.
 deactivate || :
 
+
 # Install miniconda
-fname=Miniconda3-latest-Linux-x86_64.sh
-wget https://repo.continuum.io/miniconda/$fname -O miniconda.sh
+if [[ "$TRAVIS_CPU_ARCH" == "arm64" ]]; then
+    wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-aarch64.sh -O miniconda.sh
+else
+    fname=Miniconda3-latest-Linux-x86_64.sh
+    wget https://repo.continuum.io/miniconda/$fname -O miniconda.sh
+fi
 MINICONDA_PATH=$HOME/miniconda
 chmod +x miniconda.sh && ./miniconda.sh -b -p $MINICONDA_PATH
 export PATH=$MINICONDA_PATH/bin:$PATH
 conda update --yes conda
 
 # Create environment and install dependencies
 conda create -n testenv --yes python=3.7
+
 source activate testenv
 
-pip install --upgrade pip setuptools
-echo "Installing numpy and scipy master wheels"
-dev_anaconda_url=https://pypi.anaconda.org/scipy-wheels-nightly/simple
-pip install --pre --upgrade --timeout=60 --extra-index $dev_anaconda_url numpy scipy pandas
-pip install --pre cython
-echo "Installing joblib master"
-pip install https://github.com/joblib/joblib/archive/master.zip
-echo "Installing pillow master"
-pip install https://github.com/python-pillow/Pillow/archive/master.zip
-pip install $(get_dep pytest $PYTEST_VERSION) pytest-cov
+if [[ "$TRAVIS_CPU_ARCH" == "amd64" ]]; then
+    pip install --upgrade pip setuptools
+    echo "Installing numpy and scipy master wheels"
+    dev_anaconda_url=https://pypi.anaconda.org/scipy-wheels-nightly/simple
+    pip install --pre --upgrade --timeout=60 --extra-index $dev_anaconda_url numpy scipy pandas
+    pip install --pre cython
+    echo "Installing joblib master"
+    pip install https://github.com/joblib/joblib/archive/master.zip
+    echo "Installing pillow master"
+    pip install https://github.com/python-pillow/Pillow/archive/master.zip
+else
+    conda install -y scipy numpy pandas cython
+    pip install joblib threadpoolctl
+fi
+
+pip install $(get_dep pytest $PYTEST_VERSION) pytest-cov pytest-xdist
 
 # Build scikit-learn in the install.sh script to collapse the verbose
 # build output in the travis output when it succeeds.
@@ -76,11 +94,11 @@ if [[ "$BUILD_WITH_ICC" == "true" ]]; then
     # The build_clib command is implicitly used to build libsvm-skl. To compile
     # with a different compiler we also need to specify the compiler for this
     # command.
-    python setup.py build_ext --compiler=intelem -i -j 3 build_clib --compiler=intelem
+    python setup.py build_ext --compiler=intelem -i -j "${CI_CPU_COUNT}" build_clib --compiler=intelem
 else
     # Use setup.py instead of `pip install -e .` to be able to pass the -j flag
     # to speed-up the building multicore CI machines.
-    python setup.py build_ext --inplace -j 3
+    python setup.py build_ext --inplace -j "${CI_CPU_COUNT}"
 fi
 
 python setup.py develop

build_tools/travis/test_docs.sh

@@ -9,4 +9,4 @@ if [[ "$BUILD_WITH_ICC" == "true" ]]; then
     source /opt/intel/inteloneapi/setvars.sh
 fi
 
-make test-doc
+PYTEST="pytest -n $CI_CPU_COUNT" make test-doc

build_tools/travis/test_script.sh

@@ -18,7 +18,8 @@ try:
 except ImportError:
     pass
 "
-python -c "import multiprocessing as mp; print('%d CPUs' % mp.cpu_count())"
+python -c "import joblib; print(joblib.cpu_count(), 'CPUs')"
+python -c "import platform; print(platform.machine())"
 
 if [[ "$BUILD_WITH_ICC" == "true" ]]; then
     # the tools in the oneAPI toolkits are configured via environment variables
@@ -36,9 +37,17 @@ run_tests() {
     cp setup.cfg $TEST_DIR
     cd $TEST_DIR
 
-    # Tests that require large downloads over the networks are skipped in CI.
-    # Here we make sure, that they are still run on a regular basis.
-    export SKLEARN_SKIP_NETWORK_TESTS=0
+    if [[ "$TRAVIS_CPU_ARCH" == "arm64" ]]; then
+        # use pytest-xdist for faster tests
+        TEST_CMD="$TEST_CMD -n $CI_CPU_COUNT"
+    else
+        # Tests that require large downloads over the networks are skipped in CI.
+        # Here we make sure, that they are still run on a regular basis.
+        #
+        # Note that using pytest-xdist is currently not compatible
+        # with fetching datasets in tests due to datasets cache corruptions issues.
+        export SKLEARN_SKIP_NETWORK_TESTS=0
+    fi
 
     if [[ "$COVERAGE" == "true" ]]; then
         TEST_CMD="$TEST_CMD --cov sklearn"

conftest.py

@@ -29,15 +29,25 @@ def pytest_addoption(parser):
 
 
 def pytest_collection_modifyitems(config, items):
-
-    # FeatureHasher is not compatible with PyPy
-    if platform.python_implementation() == 'PyPy':
-        skip_marker = pytest.mark.skip(
-            reason='FeatureHasher is not compatible with PyPy')
-        for item in items:
-            if item.name.endswith(('_hash.FeatureHasher',
-                                   'text.HashingVectorizer')):
-                item.add_marker(skip_marker)
+    for item in items:
+        # FeatureHasher is not compatible with PyPy
+        if (item.name.endswith(('_hash.FeatureHasher',
+                                'text.HashingVectorizer'))
+                and platform.python_implementation() == 'PyPy'):
+            marker = pytest.mark.skip(
+                reason='FeatureHasher is not compatible with PyPy')
+            item.add_marker(marker)
+        # Known failure on with GradientBoostingClassifier on ARM64
+        elif (item.name.endswith('GradientBoostingClassifier')
+                and platform.machine() == 'aarch64'):
+
+            marker = pytest.mark.xfail(
+                reason=(
+                    'know failure. See '
+                    'https://github.com/scikit-learn/scikit-learn/issues/17797'  # noqa
+                )
+            )
+            item.add_marker(marker)
 
     # Skip tests which require internet if the flag is provided
     if config.getoption("--skip-network"):

doc/developers/contributing.rst

@@ -477,9 +477,10 @@ message, the following actions are taken.
     ====================== ===================
     Commit Message Marker  Action Taken by CI
     ---------------------- -------------------
-    [scipy-dev]            Add a Travis build with our dependencies (numpy, scipy, etc ...) development builds
     [ci skip]              CI is skipped completely
     [lint skip]            Azure pipeline skips linting
+    [scipy-dev]            Add a Travis build with our dependencies (numpy, scipy, etc ...) development builds
+    [arm64]                Add a Travis build for the ARM64 / aarch64 little endian architecture
     [doc skip]             Docs are not built
     [doc quick]            Docs built, but excludes example gallery plots
     [doc build]            Docs built including example gallery plots

sklearn/cluster/tests/test_k_means.py

@@ -333,7 +333,7 @@ def test_k_means_fit_predict(algo, dtype, constructor, seed, max_iter, tol):
     # using more than one thread, the absolute values of the labels can be
     # different between the 2 strategies but they should correspond to the same
     # clustering.
-    assert v_measure_score(labels_1, labels_2) == 1
+    assert v_measure_score(labels_1, labels_2) == pytest.approx(1, abs=1e-15)
 
 
 def test_minibatch_kmeans_verbose():

sklearn/neural_network/tests/test_mlp.py

@@ -488,7 +488,7 @@ def test_predict_proba_binary():
 
     assert y_proba.shape == (n_samples, n_classes)
     assert_array_equal(proba_max, proba_log_max)
-    assert_array_equal(y_log_proba, np.log(y_proba))
+    assert_allclose(y_log_proba, np.log(y_proba))
 
     assert roc_auc_score(y, y_proba[:, 1]) == 1.0
 
@@ -511,7 +511,7 @@ def test_predict_proba_multiclass():
 
     assert y_proba.shape == (n_samples, n_classes)
     assert_array_equal(proba_max, proba_log_max)
-    assert_array_equal(y_log_proba, np.log(y_proba))
+    assert_allclose(y_log_proba, np.log(y_proba))
 
 
 def test_predict_proba_multilabel():
@@ -535,7 +535,7 @@ def test_predict_proba_multilabel():
 
     assert (y_proba.sum(1) - 1).dot(y_proba.sum(1) - 1) > 1e-10
     assert_array_equal(proba_max, proba_log_max)
-    assert_array_equal(y_log_proba, np.log(y_proba))
+    assert_allclose(y_log_proba, np.log(y_proba))
 
 
 def test_shuffle():