use kaiming initialization for conv layers (#98)

use kaiming initialization for convolutional layers in dacapo models.
This is important for our standard UNets since the default
initialization is pretty terrible when following convolutions with
ReLU's.
might want to move this into individual architectures if you are playing
around with non ReLU activations

dacapo/experiments/model.py

@@ -40,6 +40,13 @@ def __init__(
         )
         self.eval_activation = eval_activation
 
+        # UPDATE WEIGHT INITIALIZATION TO USE KAIMING
+        # TODO: put this somewhere better, there might be
+        # conv layers that aren't follwed by relus?
+        for _name, layer in self.named_modules():
+            if isinstance(layer, torch.nn.modules.conv._ConvNd):
+                torch.nn.init.kaiming_normal_(layer.weight, nonlinearity="relu")
+
     def forward(self, x):
         result = self.chain(x)
         if not self.training and self.eval_activation is not None: