fix build

README.md

@@ -20,7 +20,7 @@ It's Yet Another TF wrapper, but different in:
 	+ See [tensorpack/benchmarks](https://github.com/tensorpack/benchmarks) for the benchmark scripts.
 
 2. Focus on __large datasets__.
-	+ It's painful to read/preprocess data through TF.
+	+ It's unnecessary to read/preprocess data with a new language called TF.
 		Tensorpack helps you load large datasets (e.g. ImageNet) in __pure Python__ with autoparallelization.
 
 3. It's not a model wrapper.

examples/SimilarityLearning/mnist-embeddings.py

@@ -247,7 +247,7 @@ class SiameseModel(EmbeddingModel):
 
         # tag the embedding of 'input' with name 'emb', just for inference later on
         with tf.variable_scope(tf.get_variable_scope(), reuse=True):
-            tf.identity(self.embed(inputs[0]), name="emb")
+            tf.identity(self.embed(x), name="emb")
 
         # compute the actual loss
         cost, pos_dist, neg_dist = contrastive_loss(x, y, label, 5., extra=True, scope="loss")
@@ -263,7 +263,7 @@ class CosineModel(SiameseModel):
         x, y = self.embed([x, y])
 
         with tf.variable_scope(tf.get_variable_scope(), reuse=True):
-            tf.identity(self.embed(inputs[0]), name="emb")
+            tf.identity(self.embed(x), name="emb")
 
         cost = siamese_cosine_loss(x, y, label, scope="loss")
         cost = tf.identity(cost, name="cost")
@@ -290,7 +290,7 @@ class TripletModel(EmbeddingModel):
         a, p, n = self.embed([a, p, n])
 
         with tf.variable_scope(tf.get_variable_scope(), reuse=True):
-            tf.identity(self.embed(inputs[0]), name="emb")
+            tf.identity(self.embed(a), name="emb")
 
         cost, pos_dist, neg_dist = self.loss(a, p, n)
 
@@ -321,7 +321,7 @@ class CenterModel(EmbeddingModel):
 
         # tag the embedding of 'input' with name 'emb', just for inference later on
         with tf.variable_scope(tf.get_variable_scope(), reuse=True):
-            tf.identity(self.embed(inputs[0]), name="emb")
+            tf.identity(self.embed(x), name="emb")
 
         # compute the embedding loss
         emb_cost = center_loss(x, label, 10, 0.01)