Speed up CI runtime (#3189)

* Speedup test_ucf101 (#2623 

* Speedup Cmake builds (#3186)

* Speedup test_autoaugment (#3190)

* Speedup DeformConvTester (#3191)

* Speedup InceptionV3 and GoogleNet on Windows (#3196)

.circleci/unittest/windows/scripts/environment.yml

@@ -13,6 +13,6 @@ dependencies:
   - pip:
     - future
     - pillow>=4.1.1
-    - scipy==1.4.1
+    - scipy
     - av
     - dataclasses

packaging/build_cmake.sh

 #!/bin/bash
 set -ex
 
+PARALLELISM=8
+if [ -n "$MAX_JOBS" ]; then
+    PARALLELISM=$MAX_JOBS
+fi
+
 if [[ "$(uname)" != Darwin && "$OSTYPE" != "msys" ]]; then
     eval "$(./conda/bin/conda shell.bash hook)"
     conda activate ./env
@@ -40,11 +45,11 @@ cmake .. -DTorch_DIR=$TORCH_PATH/share/cmake/Torch -DWITH_CUDA=$CMAKE_USE_CUDA
 
 # Compile and install libtorchvision
 if [[ "$OSTYPE" == "msys" ]]; then
-    "$script_dir/windows/internal/vc_env_helper.bat" "$script_dir/windows/internal/build_cmake.bat"
+    "$script_dir/windows/internal/vc_env_helper.bat" "$script_dir/windows/internal/build_cmake.bat" $PARALLELISM
     CONDA_PATH=$(dirname $(which python))
     cp -r "C:/Program Files (x86)/torchvision/include/torchvision" $CONDA_PATH/include
 else
-    make
+    make -j$PARALLELISM
     make install
 
     if [[ "$(uname)" == Darwin ]]; then
@@ -71,12 +76,12 @@ cp fasterrcnn_resnet50_fpn.pt build
 cd build
 cmake .. -DTorch_DIR=$TORCH_PATH/share/cmake/Torch -DWITH_CUDA=$CMAKE_USE_CUDA
 if [[ "$OSTYPE" == "msys" ]]; then
-    "$script_dir/windows/internal/vc_env_helper.bat" "$script_dir/windows/internal/build_frcnn.bat"
+    "$script_dir/windows/internal/vc_env_helper.bat" "$script_dir/windows/internal/build_frcnn.bat" $PARALLELISM
     mv fasterrcnn_resnet50_fpn.pt Release
     cd Release
     export PATH=$(cygpath "C:/Program Files (x86)/torchvision/bin"):$(cygpath $TORCH_PATH)/lib:$PATH
 else
-    make
+    make -j$PARALLELISM
 fi
 
 # Run traced program
@@ -91,10 +96,10 @@ cd build
 cmake .. -DTorch_DIR=$TORCH_PATH/share/cmake/Torch
 
 if [[ "$OSTYPE" == "msys" ]]; then
-    "$script_dir/windows/internal/vc_env_helper.bat" "$script_dir/windows/internal/build_cpp_example.bat"
+    "$script_dir/windows/internal/vc_env_helper.bat" "$script_dir/windows/internal/build_cpp_example.bat" $PARALLELISM
     cd Release
 else
-    make
+    make -j$PARALLELISM
 fi
 
 # Run CPP example

packaging/windows/internal/build_cmake.bat

 @echo on
-msbuild "-p:Configuration=Release" torchvision.vcxproj
-msbuild "-p:Configuration=Release" INSTALL.vcxproj
+msbuild "-p:Configuration=Release" "-p:BuildInParallel=true" "-p:MultiProcessorCompilation=true" "-p:CL_MPCount=%1" torchvision.vcxproj -maxcpucount:%1
+msbuild "-p:Configuration=Release" "-p:BuildInParallel=true" "-p:MultiProcessorCompilation=true" "-p:CL_MPCount=%1" INSTALL.vcxproj -maxcpucount:%1

packaging/windows/internal/build_cpp_example.bat

 @echo on
 set CL=/I"C:\Program Files (x86)\torchvision\include"
-msbuild "-p:Configuration=Release" hello-world.vcxproj
+msbuild "-p:Configuration=Release" "-p:BuildInParallel=true" "-p:MultiProcessorCompilation=true" "-p:CL_MPCount=%1" hello-world.vcxproj -maxcpucount:%1

packaging/windows/internal/build_frcnn.bat

 @echo on
 set CL=/I"C:\Program Files (x86)\torchvision\include"
-msbuild "-p:Configuration=Release" test_frcnn_tracing.vcxproj
+msbuild "-p:Configuration=Release" "-p:BuildInParallel=true" "-p:MultiProcessorCompilation=true" "-p:CL_MPCount=%1" test_frcnn_tracing.vcxproj -maxcpucount:%1

test/common_utils.py

@@ -39,7 +39,6 @@ def set_rng_seed(seed):
 
 ACCEPT = os.getenv('EXPECTTEST_ACCEPT')
 TEST_WITH_SLOW = os.getenv('PYTORCH_TEST_WITH_SLOW', '0') == '1'
-# TEST_WITH_SLOW = True  # TODO: Delete this line once there is a PYTORCH_TEST_WITH_SLOW aware CI job
 
 
 parser = argparse.ArgumentParser(add_help=False)

test/test_datasets.py

@@ -264,12 +264,15 @@ class Tester(unittest.TestCase):
 
     @unittest.skipIf(not HAS_PYAV, "PyAV unavailable")
     def test_ucf101(self):
+        cached_meta_data = None
         with ucf101_root() as (root, ann_root):
             for split in {True, False}:
                 for fold in range(1, 4):
                     for length in {10, 15, 20}:
-                        dataset = torchvision.datasets.UCF101(
-                            root, ann_root, length, fold=fold, train=split)
+                        dataset = torchvision.datasets.UCF101(root, ann_root, length, fold=fold, train=split,
+                                                              num_workers=2, _precomputed_metadata=cached_meta_data)
+                        if cached_meta_data is None:
+                            cached_meta_data = dataset.metadata
                         self.assertGreater(len(dataset), 0)
 
                         video, audio, label = dataset[0]

test/test_models.py

@@ -268,7 +268,7 @@ class ModelTester(TestCase):
             f = 2 ** sum(i)
             self.assertEqual(out.shape, (1, 2048, 7 * f, 7 * f))
 
-    def test_mobilenetv2_residual_setting(self):
+    def test_mobilenet_v2_residual_setting(self):
         model = models.__dict__["mobilenet_v2"](inverted_residual_setting=[[1, 16, 1, 1], [6, 24, 2, 2]])
         model.eval()
         x = torch.rand(1, 3, 224, 224)
@@ -286,7 +286,7 @@ class ModelTester(TestCase):
         self.assertFalse(any(isinstance(x, nn.BatchNorm2d) for x in model.modules()))
         self.assertTrue(any(isinstance(x, nn.GroupNorm) for x in model.modules()))
 
-    def test_inceptionv3_eval(self):
+    def test_inception_v3_eval(self):
         # replacement for models.inception_v3(pretrained=True) that does not download weights
         kwargs = {}
         kwargs['transform_input'] = True

test/test_ops.py

@@ -5,6 +5,7 @@ import unittest
 import numpy as np
 
 import torch
+from functools import lru_cache
 from torch import Tensor
 from torch.autograd import gradcheck
 from torch.nn.modules.utils import _pair
@@ -496,6 +497,7 @@ class DeformConvTester(OpTester, unittest.TestCase):
         out += bias.view(1, n_out_channels, 1, 1)
         return out
 
+    @lru_cache(maxsize=None)
     def get_fn_args(self, device, contiguous, batch_sz, dtype):
         n_in_channels = 6
         n_out_channels = 2
@@ -614,9 +616,11 @@ class DeformConvTester(OpTester, unittest.TestCase):
         gradcheck(lambda z, off, wei, bi: script_func_no_mask(z, off, wei, bi, stride, padding, dilation),
                   (x, offset, weight, bias), nondet_tol=1e-5)
 
+    @unittest.skipIf(not torch.cuda.is_available(), "CUDA unavailable")
+    def test_compare_cpu_cuda_grads(self):
         # Test from https://github.com/pytorch/vision/issues/2598
         # Run on CUDA only
-        if "cuda" in device.type:
+        for contiguous in [False, True]:
             # compare grads computed on CUDA with grads computed on CPU
             true_cpu_grads = None
 

test/test_transforms_tensor.py

@@ -630,15 +630,16 @@ class Tester(TransformsTester):
         tensor = torch.randint(0, 256, size=(3, 44, 56), dtype=torch.uint8, device=self.device)
         batch_tensors = torch.randint(0, 256, size=(4, 3, 44, 56), dtype=torch.uint8, device=self.device)
 
+        s_transform = None
         for policy in T.AutoAugmentPolicy:
             for fill in [None, 85, (10, -10, 10), 0.7, [0.0, 0.0, 0.0], [1, ], 1]:
-                for _ in range(100):
                 transform = T.AutoAugment(policy=policy, fill=fill)
                 s_transform = torch.jit.script(transform)
-
+                for _ in range(100):
                     self._test_transform_vs_scripted(transform, s_transform, tensor)
                     self._test_transform_vs_scripted_on_batch(transform, s_transform, batch_tensors)
 
+        if s_transform is not None:
             with get_tmp_dir() as tmp_dir:
                 s_transform.save(os.path.join(tmp_dir, "t_autoaugment.pt"))