State dict support (#6)

1 files changed, 45 insertions, 1 deletions

diff --git a/tests/test_ema.py b/tests/test_ema.py
index ad6ee37..67a14dc 100644
--- a/tests/test_ema.py
+++ b/tests/test_ema.py
@@ -1,5 +1,7 @@
 import pytest
 
+import copy
+
 import torch
 
 from torch_ema import ExponentialMovingAverage
@@ -40,6 +42,7 @@ def test_val_error(decay, use_num_updates, explicit_params):
     model.eval()
     logits = model(x_val)
     loss_orig = torch.nn.functional.cross_entropy(logits, y_val)
+    print(f"Original loss: {loss_orig}")
 
     # Validation: with EMA
     # First save original parameters before replacing with EMA version
@@ -55,6 +58,7 @@ def test_val_error(decay, use_num_updates, explicit_params):
     logits = model(x_val)
     loss_ema = torch.nn.functional.cross_entropy(logits, y_val)
 
+    print(f"EMA loss: {loss_ema}")
     assert loss_ema < loss_orig, "EMA loss wasn't lower"
 
     # Test restore
@@ -131,4 +135,44 @@ def test_explicit_params():
         model2.weight.fill_(1.0)
     ema.update(model2.parameters())
     ema.copy_to()
-    assert not torch.all(model.weight == 0.0)
\ No newline at end of file
+    assert not torch.all(model.weight == 0.0)
+
+
+@pytest.mark.parametrize("decay", [0.995])
+@pytest.mark.parametrize("use_num_updates", [True, False])
+@pytest.mark.parametrize("explicit_params", [True, False])
+def test_state_dict(decay, use_num_updates, explicit_params):
+    model = torch.nn.Linear(10, 2, bias=False)
+    with torch.no_grad():
+        model.weight.fill_(0.0)
+    ema = ExponentialMovingAverage(
+        model.parameters(),
+        decay=decay,
+        use_num_updates=False
+    )
+    state_dict = copy.deepcopy(ema.state_dict())
+
+    model2 = torch.nn.Linear(10, 2, bias=False)
+    ema2 = ExponentialMovingAverage(model2.parameters(), decay=0.0)
+    ema2.load_state_dict(state_dict)
+    assert ema2.decay == decay
+    assert torch.allclose(ema2.shadow_params[0], ema.shadow_params[0])
+
+    with torch.no_grad():
+        model2.weight.fill_(1.0)
+    if explicit_params:
+        ema2.update(model2.parameters())
+    else:
+        ema2.update()
+    assert torch.all(model2.weight == 1.0), "ema.update changed model weights"
+
+    ema.load_state_dict(ema2.state_dict())
+
+    if explicit_params:
+        ema.copy_to(model.parameters())
+    else:
+        ema.copy_to()
+    assert torch.allclose(
+        model.weight,
+        torch.full(size=(1,), fill_value=(1.0 - decay))
+    ), "average was wrong"