use a sync batchnorm as way of normalizing all channels of the HRRR, …

…applied to target, but not to predictions

README.md

@@ -73,8 +73,9 @@ surface_target, hrrr_target, precipitation_target = metnet3(
 ## Todo
 
 - [x] figure out all the cross entropy and MSE losses
+- [x] auto-handle normalization across all the channels of the HRRR by tracking a running mean and variance of targets during training (using sync batchnorm as hack)
 
-- [ ] auto-handle normalization across all the channels of the HRRR by tracking a running mean and variance of targets during training, as well as allow researcher to pass in their own normalization variables
+- [ ] allow researcher to pass in their own normalization variables for HRRR
 - [ ] figure out the topological embedding, consult a neural weather researcher
 
 ## Citations

metnet3_pytorch/metnet3_pytorch.py

@@ -1,3 +1,4 @@
+from contextlib import contextmanager
 from functools import partial
 from collections import namedtuple
 
@@ -40,6 +41,21 @@ def MaybeSyncBatchnorm2d(is_distributed = None):
     is_distributed = default(is_distributed, dist.is_initialized() and dist.get_world_size() > 1)
     return nn.SyncBatchNorm if is_distributed else nn.BatchNorm2d
 
+@contextmanager
+def freeze_batchnorm(bn):
+    assert not exists(next(bn.parameters(), None))
+
+    was_training = bn.training
+    was_tracking_stats = bn.track_running_stats # in some versions of pytorch, running mean and variance still gets updated even in eval mode it seems..
+
+    bn.eval()
+    bn.track_running_stats = False
+
+    yield bn
+
+    bn.train(was_training)
+    bn.track_running_stats = was_tracking_stats
+
 # loss scaling in section 4.3.2
 
 class LossScaleFunction(Function):
@@ -686,6 +702,8 @@ def __init__(
             nn.Conv2d(dim, precipitation_target_channels, 1)
         )
 
+        self.batchnorm_hrrr = MaybeSyncBatchnorm2d()(hrrr_target_channels, affine = False)
+
         self.mse_loss_scaler = LossScaler()
 
     def forward(
@@ -779,9 +797,18 @@ def forward(
 
         # calculate HRRR mse loss
 
-        hrrr_pred = self.mse_loss_scaler(hrrr_pred)
+        # use a batchnorm to normalize each channel to mean zero and unit variance
+
+        normed_hrrr_target = self.batchnorm_hrrr(hrrr_target)
+
+        with freeze_batchnorm(self.batchnorm_hrrr) as frozen_batchnorm:
+            normed_hrrr_pred = frozen_batchnorm(hrrr_pred)
+
+        # proposed loss gradient rescaler from section 4.3.2
+
+        normed_hrrr_pred = self.mse_loss_scaler(normed_hrrr_pred)
 
-        hrrr_loss = F.mse_loss(hrrr_pred, hrrr_target)
+        hrrr_loss = F.mse_loss(normed_hrrr_pred, normed_hrrr_target)
 
         # total loss
 

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'metnet3-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.0.2',
+  version = '0.0.3',
   license='MIT',
   description = 'MetNet 3 - Pytorch',
   author = 'Phil Wang',