Merge pull request #1 from lucasb-eyer/always-return-params

Default `Module.parameters` always return params, regardless of mode.

Author: lucasb-eyer

DeepFried2/layers/Module.py

@@ -22,12 +22,12 @@ def zero_grad_parameters(self):
     def parameters(self):
         params, grads = [], []
 
-        if self.training_mode and hasattr(self, 'weight'):
+        if hasattr(self, 'weight'):
             assert hasattr(self, 'grad_weight'), "The layer {} has a `weight` variable but no `grad_weight`, you probably forget to implement it.".format(type(self))
             params += [self.weight]
             grads += [self.grad_weight]
 
-        if self.training_mode and hasattr(self, 'bias'):
+        if hasattr(self, 'bias'):
             assert hasattr(self, 'grad_bias'), "The layer {} has a `bias` variable but no `grad_bias`, you probably forget to implement it.".format(type(self))
             params += [self.bias]
             grads += [self.grad_bias]

examples/MNIST/run.py

@@ -1,3 +1,4 @@
+import DeepFried2 as df
 import DeepFried2.optimizers as optim
 from mnist import *
 from train import *
@@ -12,7 +13,7 @@ def main(params):
 
     model = lenet()
 
-    criterion = bb8.ClassNLLCriterion()
+    criterion = df.ClassNLLCriterion()
 
     optimiser = optim.SGD(lr=params['lr'])