Merge pull request #98 from lucasb-eyer/bn-momentum

BN momentum

Author: lucasb-eyer

DeepFried2/layers/BatchNormalization.py

@@ -6,13 +6,15 @@
 
 
 class BatchNormalization(df.Module):
-    def __init__(self, n_features, eps=1e-5):
+    def __init__(self, n_features, momentum=False, eps=1e-5):
         """
         - `n_features` may be an integer (#features, #feature-maps for images) or a tuple.
             - If a single integer, it indicates the size of the 1-axis, i.e. first feature-axis.
               This is the only axis that will be normalized using statistics across all other axes.
             - If a tuple, it indicates the sizes of multiple axes (starting at 1) which are
               considered feature-axes and will consequently be normalized over statistics across all other axes.
+        - `momentum` means statistics are collected as running (geometric) statistics
+          during training. It should be the decay value within (0,1).
         - `eps` is a small number which is added to the variance in order to
           avoid computing sqrt(0) for features with zero variance.
         """
@@ -30,6 +32,7 @@ def __init__(self, n_features, eps=1e-5):
         self.buf_var = df.th.shared(_np.full(n_features, 1, df.floatX), name='BN_var_{}'.format(n_features))
         self.buf_mean = df.th.shared(_np.full(n_features, 0, df.floatX), name='BN_mean_{}'.format(n_features))
         self.buf_count = df.th.shared(_np.asarray(0, dtype=df.floatX), name='BN_count_{}'.format(n_features))
+        self.momentum = momentum
 
         self.eps = eps or 1e-5
 
@@ -74,6 +77,16 @@ def get_stat_updates(self):
             (self.buf_count, self.buf_count + 1.0),
         ]
 
+    def get_extra_updates(self):
+        if self._mode == 'train' and self.momentum != False:
+            return [
+                (self.buf_mean, self.momentum*self.buf_mean + (1-self.momentum)*self.batch_mean),
+                (self.buf_var, self.momentum*self.buf_var + (1-self.momentum)*self.batch_var),
+                (self.buf_count, 1.0),
+            ]
+        else:
+            return []
+
     def training(self):
         df.Module.training(self)
         self.buf_count.set_value(0)

DeepFried2/zoo/resnet.py

@@ -11,13 +11,13 @@ def symb_forward(self, symb_inputs):
         return s
 
 
-def block(nchan, fs=(3,3), body=None):
+def block(nchan, fs=(3,3), body=None, bnmom=False):
     return df.Sequential(
         df.RepeatInput(
             df.Sequential(
-                df.BatchNormalization(nchan), df.ReLU(),
+                df.BatchNormalization(nchan, bnmom), df.ReLU(),
                 df.SpatialConvolutionCUDNN(nchan, nchan, fs, border='same', init=df.init.prelu(), bias=False),
-                df.BatchNormalization(nchan), df.ReLU(),
+                df.BatchNormalization(nchan, bnmom), df.ReLU(),
                 df.SpatialConvolutionCUDNN(nchan, nchan, fs, border='same', init=df.init.prelu(), bias=False)
             ) if body is None else body,
             df.Identity()
@@ -26,13 +26,13 @@ def block(nchan, fs=(3,3), body=None):
     )
 
 
-def block_proj(nin, nout, fs=(3,3), body=None):
+def block_proj(nin, nout, fs=(3,3), body=None, bnmom=False):
     return df.Sequential(
         df.RepeatInput(
             df.Sequential(
-                df.BatchNormalization(nin), df.ReLU(),
+                df.BatchNormalization(nin, bnmom), df.ReLU(),
                 df.SpatialConvolutionCUDNN(nin, nout, fs, border='same', init=df.init.prelu(), bias=False),
-                df.BatchNormalization(nout), df.ReLU(),
+                df.BatchNormalization(nout, bnmom), df.ReLU(),
                 df.SpatialConvolutionCUDNN(nout, nout, fs, border='same', init=df.init.prelu(), bias=False)
             ) if body is None else body,
             df.SpatialConvolutionCUDNN(nin, nout, (1,)*len(fs)),