caffe2fluid: fix bug in leaky-relu and add interface for prelu conver… (#1086)

* caffe2fluid: fix bug in leaky-relu and add interface for prelu convertion
* caffe2fluid:fix code style problem in graph.py

fluid/image_classification/caffe2fluid/examples/imagenet/compare.py

@@ -24,15 +24,10 @@ def calc_diff(f1, f2):
     #print d2.shape
     #print d1[0, 0, 0:10, 0:10]
     #print d2[0, 0, 0:10, 0:10]
-    #d1 = d1[:, :, 1:-2, 1:-2]
-    #d2 = d2[:, :, 1:-2, 1:-2]
 
     d1 = d1.flatten()
     d2 = d2.flatten()
 
-    #print d1[:10]
-    #print d2[:10]
-
     d1_num = reduce(lambda x, y: x * y, d1.shape)
     d2_num = reduce(lambda x, y: x * y, d2.shape)
     if d1_num != d2_num:
@@ -41,7 +36,11 @@ def calc_diff(f1, f2):
         assert (d1_num == d2_num), "their shape is not consistent"
 
     try:
+        mask = np.abs(d1) >= np.abs(d2)
+        mask = mask.astype('int32')
+
         df = np.abs(d1 - d2)
+        df = df / (1.0e-10 + np.abs(d1) * mask + np.abs(d2) * (1 - mask))
         max_df = np.max(df)
         sq_df = np.mean(df * df)
         return max_df, sq_df

fluid/image_classification/caffe2fluid/kaffe/layers.py

@@ -39,6 +39,7 @@ LAYER_DESCRIPTORS = {
     'Pooling': shape_pool,
     'Power': shape_identity,
     'ReLU': shape_identity,
+    'PReLU': shape_identity,
     'Scale': shape_identity,
     'Sigmoid': shape_identity,
     'SigmoidCrossEntropyLoss': shape_scalar,

fluid/image_classification/caffe2fluid/kaffe/paddle/network.py

@@ -240,10 +240,16 @@ class Network(object):
     @layer
     def relu(self, input, name):
         fluid = import_fluid()
-        output = fluid.layers.relu(
-            name=self.get_unique_output_name(name, 'relu'), x=input)
+        output = fluid.layers.relu(input)
         return output
 
+    @layer
+    def prelu(self, input, channel_shared, name):
+        #fluid = import_fluid()
+        #output = fluid.layers.relu(input)
+        #return output
+        raise NotImplementedError('prelu not implemented')
+
     def pool(self, pool_type, input, k_h, k_w, s_h, s_w, ceil_mode, padding,
              name):
         # Get the number of channels in the input
@@ -382,7 +388,8 @@ class Network(object):
                             name,
                             scale_offset=True,
                             eps=1e-5,
-                            relu=False):
+                            relu=False,
+                            relu_negative_slope=0.0):
         # NOTE: Currently, only inference is supported
         fluid = import_fluid()
         prefix = name + '_'
@@ -392,6 +399,15 @@ class Network(object):
             name=prefix + 'offset')
         mean_name = prefix + 'mean'
         variance_name = prefix + 'variance'
+
+        leaky_relu = False
+        act = 'relu'
+        if relu is False:
+            act = None
+        elif relu_negative_slope != 0.0:
+            leaky_relu = True
+            act = None
+
         output = fluid.layers.batch_norm(
             name=self.get_unique_output_name(name, 'batch_norm'),
             input=input,
@@ -401,7 +417,10 @@ class Network(object):
             moving_mean_name=mean_name,
             moving_variance_name=variance_name,
             epsilon=eps,
-            act='relu' if relu is True else None)
+            act=act)
+
+        if leaky_relu:
+            output = fluid.layers.leaky_relu(output, alpha=relu_negative_slope)
 
         return output
 

fluid/image_classification/caffe2fluid/kaffe/paddle/transformer.py

@@ -112,6 +112,13 @@ class PaddleMapper(NodeMapper):
     def map_relu(self, node):
         return PaddleNode('relu')
 
+    def map_prelu(self, node):
+        channel_shared = getattr(node.parameters, 'channel_shared', False)
+        return PaddleNode('prelu', channel_shared)
+
+    def map_tanh(self, node):
+        return PaddleNode('tanh')
+
     def map_pooling(self, node):
         pool_type = node.parameters.pool
         if pool_type == 0: