Fix AffineChannel op param initialization

Summary:
The `AffineChannel` op wrapper did not use the officially sanctioned
version of parameter creation, namely using `ModelHelper.create_param()`
Presumably this was because of the `share_with` option where you could share
parameters.

While doing this though, the parameters were never initialized! This probably
doesn't matter for the algorithm (those params get overwritten pretty quickly)
but it's a bug nevertheless. I decided to kill the `share_with` option since
this isn't used anyway and resort to the standard way of creating weights and
biases.

Reviewed By: rbgirshick

Differential Revision: D6826914

fbshipit-source-id: 65628c59b085b9ab160006b003dd40dbefa2f7c1

lib/modeling/ResNet.py

@@ -92,7 +92,7 @@ def add_ResNet_convX_body(model, block_counts, freeze_at=2):
     X = 4 or 5)."""
     assert freeze_at in [0, 2, 3, 4, 5]
     p = model.Conv('data', 'conv1', 3, 64, 7, pad=3, stride=2, no_bias=1)
-    p = model.AffineChannel(p, 'res_conv1_bn', inplace=True)
+    p = model.AffineChannel(p, 'res_conv1_bn', dim=64, inplace=True)
     p = model.Relu(p, p)
     p = model.MaxPool(p, 'pool1', kernel=3, pad=1, stride=2)
     dim_in = 64
@@ -204,7 +204,7 @@ def add_shortcut(model, prefix, blob_in, dim_in, dim_out, stride):
         stride=stride,
         no_bias=1
     )
-    return model.AffineChannel(c, prefix + '_branch1_bn')
+    return model.AffineChannel(c, prefix + '_branch1_bn', dim=dim_out)
 
 
 # ------------------------------------------------------------------------------

lib/modeling/detector.py

@@ -26,6 +26,8 @@ import logging
 from caffe2.python import cnn
 from caffe2.python import core
 from caffe2.python import workspace
+from caffe2.python.modeling import initializers
+from caffe2.python.modeling.parameter_info import ParameterTags
 
 from core.config import cfg
 from ops.collect_and_distribute_fpn_rpn_proposals \
@@ -75,26 +77,27 @@ class DetectionModelHelper(cnn.CNNModelHelper):
                  str(p).find('gpu_{}'.format(gpu_id)) == 0)
             )]
 
-    def AffineChannel(self, blob_in, blob_out, share_with=None, inplace=False):
+    def AffineChannel(self, blob_in, blob_out, dim, inplace=False):
         """Affine transformation to replace BN in networks where BN cannot be
         used (e.g., because the minibatch size is too small).
 
-        The AffineChannel parameters may be shared with another AffineChannelOp
-        by specifying its blob name (excluding the '_{s,b}' suffix) in the
-        share_with argument. The operations can be done in place to save memory.
+        The operations can be done in place to save memory.
         """
         blob_out = blob_out or self.net.NextName()
-        is_not_sharing = share_with is None
-        param_prefix = blob_out if is_not_sharing else share_with
-        scale = core.ScopedBlobReference(
-            param_prefix + '_s', self.param_init_net)
-        bias = core.ScopedBlobReference(
-            param_prefix + '_b', self.param_init_net)
-        if is_not_sharing:
-            self.net.Proto().external_input.extend([str(scale), str(bias)])
-            self.params.extend([scale, bias])
-            self.weights.append(scale)
-            self.biases.append(bias)
+        param_prefix = blob_out
+
+        scale = self.create_param(
+            param_name=param_prefix + '_s',
+            initializer=initializers.Initializer("ConstantFill", value=1.),
+            tags=ParameterTags.WEIGHT,
+            shape=[dim, ],
+        )
+        bias = self.create_param(
+            param_name=param_prefix + '_b',
+            initializer=initializers.Initializer("ConstantFill", value=0.),
+            tags=ParameterTags.BIAS,
+            shape=[dim, ],
+        )
         if inplace:
             return self.net.AffineChannel([blob_in, scale, bias], blob_in)
         else:
@@ -403,7 +406,7 @@ class DetectionModelHelper(cnn.CNNModelHelper):
             no_bias=1
         )
         blob_out = self.AffineChannel(
-            conv_blob, prefix + suffix, inplace=inplace
+            conv_blob, prefix + suffix, dim=dim_out, inplace=inplace
         )
         return blob_out