Zp/roll back conv bias (#1000)

* roll back conv_bias

neon/layers/layer.py

@@ -1794,7 +1794,8 @@ def __init__(self, fshape, init, strides={}, padding={}, dilation={},
                  name=None):
         super(Conv, self).__init__(bias=bias, batch_norm=batch_norm,
                                    activation=activation, name=name)
-        if bias and NervanaObject.be.is_mkl():
+        # temp fall back to old conv and bias for bug of weights save/load
+        if False and bias and NervanaObject.be.is_mkl():
             self.append(Convolution_bias(fshape=fshape, strides=strides, padding=padding,
                                          dilation=dilation, init=init, bsum=batch_norm, bias=bias,
                                          name=name))
@@ -1807,7 +1808,8 @@ def __init__(self, fshape, init, strides={}, padding={}, dilation={},
     def add_postfilter_layers(self):
         self.init_base_name()
         # mklbackend will do conv+bias
-        if self.bias is not None and not NervanaObject.be.is_mkl():
+        # fall back
+        if self.bias is not None:
             name = self.base_name + '_bias'
             self.append(Bias(init=self.bias, name=name))
         if self.batch_norm:

tests/test_optimizer.py

@@ -388,7 +388,8 @@ def test_multi_optimizer(backend_default_mkl):
     layers_to_optimize1 = [l for l in layer_list1 if isinstance(l, ParameterLayer)]
     layers_to_optimize2 = [l for l in layer_list2 if isinstance(l, ParameterLayer)]
     opt.optimize(layers_to_optimize1, 0)
-    if l1[0].be.is_mkl():
+    # temp roll back conv_bias
+    if False and l1[0].be.is_mkl():
         assert opt.map_list[opt_adam][0].__class__.__name__ is 'Convolution_bias'
     else:
         assert opt.map_list[opt_adam][0].__class__.__name__ is 'Convolution'

tests/test_wrapper_layers.py

@@ -28,7 +28,8 @@ def test_conv_wrapper(backend_default):
 
     conv = Conv((4, 4, 3), Uniform(), bias=Uniform())
     assert isinstance(conv, list)
-    if conv[0].be.is_mkl():
+    # temp roll back conv_bias
+    if False and conv[0].be.is_mkl():
         assert len(conv) == 1
         assert isinstance(conv[0], Convolution_bias)
     else:
@@ -44,7 +45,8 @@ def test_conv_wrapper(backend_default):
 
     conv = Conv((4, 4, 3), Uniform(), bias=Uniform(), activation=Rectlin())
     assert isinstance(conv, list)
-    if conv[0].be.is_mkl():
+    # temp roll back conv_bias
+    if False and conv[0].be.is_mkl():
         assert isinstance(conv[0], Convolution_bias)
         assert isinstance(conv[1], Activation)
         assert len(conv) == 2