Fix SiLU breakage in RegNets for PT < 1.7

.circleci/config.yml

@@ -126,10 +126,10 @@ jobs:
             # fallback to using the latest cache if no exact match is found
             - v7-cpu-dependencies-
 
-      - <<: *install_dep
-
       - <<: *install_dev_dep
 
+      - <<: *install_dep
+
       - <<: *pip_list
 
       - save_cache:
@@ -177,10 +177,10 @@ jobs:
             # fallback to using the latest cache if no exact match is found
             - v4-gpu-dependencies-
 
-      - <<: *install_dep
-
       - <<: *install_dev_dep
 
+      - <<: *install_dep
+
       - <<: *pip_list
 
       - <<: *check_cuda_available
@@ -214,10 +214,10 @@ jobs:
             # fallback to using the latest cache if no exact match is found
             - v2-gpu-bc-dependencies-
 
-      - <<: *install_dep_bc
-
       - <<: *install_dev_dep
 
+      - <<: *install_dep_bc
+
       - <<: *pip_list
 
       - <<: *check_cuda_available

classy_vision/models/regnet.py

@@ -447,17 +447,15 @@ class RegNet(ClassyModel):
     def __init__(self, params: RegNetParams):
         super().__init__()
 
-        if params.activation_type == ActivationType.SILU and get_torch_version() < [
-            1,
-            7,
-        ]:
-            raise RuntimeError("SiLU activation is only supported since PyTorch 1.7")
-
+        silu = None if get_torch_version() < [1, 7] else nn.SiLU()
         activation = {
             ActivationType.RELU: nn.ReLU(params.relu_in_place),
-            ActivationType.SILU: nn.SiLU(),
+            ActivationType.SILU: silu,
         }[params.activation_type]
 
+        if activation is None:
+            raise RuntimeError("SiLU activation is only supported since PyTorch 1.7")
+
         # Ad hoc stem
         self.stem = {
             StemType.RES_STEM_CIFAR: ResStemCifar,

test/optim_sharded_sgd_test.py

@@ -25,25 +25,29 @@
 from classy_vision.optim.zero import ZeRO
 
 
-def dist_init(rank, world_size):
-    os.environ["MASTER_ADDR"] = "localhost"
-    os.environ["MASTER_PORT"] = "29500"
-    dist.init_process_group(backend=dist.Backend.GLOO, rank=rank, world_size=world_size)
+def dist_init(rank, world_size, filename):
+    dist.init_process_group(
+        init_method="file://" + filename,
+        backend=dist.Backend.GLOO,
+        rank=rank,
+        world_size=world_size,
+    )
 
 
 class TestOptimizerStateShardingIntegration(unittest.TestCase, TestOptimizer):
     @staticmethod
-    def _maybe_destro_dist():
+    def _maybe_destroy_dist():
         if dist.is_initialized():
             logging.debug("Destroy previous torch dist process group")
             dist.destroy_process_group()
 
     def setUp(self):
-        self._maybe_destro_dist()
-        dist_init(0, 1)
+        self._maybe_destroy_dist()
+        self.filename = tempfile.NamedTemporaryFile(delete=True).name
+        dist_init(0, 1, self.filename)
 
     def tearDown(self):
-        self._maybe_destro_dist()
+        self._maybe_destroy_dist()
 
     def _get_config(self):
         return {"name": "zero", "base_optimizer": {"name": "sgd"}, "num_epochs": 3}