fix deeper_gcn

examples/deeper_gcn/README.md

@@ -12,6 +12,14 @@ The datasets contain three citation networks: CORA, PUBMED, CITESEER. The detail
 - paddlepaddle>=1.6
 - pgl
 
+### Performance
+
+We train our models for 200 epochs and report the accuracy on the test dataset.
+
+| Dataset | Accuracy |
+| --- | --- |
+| Cora | ~77% | 
+
 ### How to run
 
 For examples, use gpu to train gat on cora dataset.

examples/deeper_gcn/train.py

@@ -44,7 +44,7 @@ def main(args):
     startup_program = fluid.Program()
     test_program = fluid.Program()
     hidden_size = 64
-    num_layers = 50
+    num_layers = 7
 
     with fluid.program_guard(train_program, startup_program):
         gw = pgl.graph_wrapper.GraphWrapper(
@@ -103,7 +103,7 @@ def main(args):
     
     # get beta param
     beta_param_list = []
-    for param in train_program.global_block().all_parameters():
+    for param in fluid.io.get_program_parameter(train_program):
         if param.name.endswith("_beta"):
             beta_param_list.append(param)
 
@@ -119,7 +119,7 @@ def main(args):
                                         return_numpy=True)
         for param in beta_param_list:
             beta = np.array(fluid.global_scope().find_var(param.name).get_tensor())
-            writer.add_scalar(param.name, beta, epoch)
+            writer.add_scalar("beta/"+param.name, beta, epoch)
 
         if epoch >= 3:
             time_per_epoch = 1.0 * (time.time() - t0)

pgl/message_passing.py

@@ -50,13 +50,14 @@ def max_recv(feat):
     return fluid.layers.sequence_pool(feat, pool_type="max")
 
 
-def lstm_recv(feat):
+def lstm_recv(hidden_dim):
     """doc"""
-    hidden_dim = 128
-    forward, _ = fluid.layers.dynamic_lstm(
-        input=feat, size=hidden_dim * 4, use_peepholes=False)
-    output = fluid.layers.sequence_last_step(forward)
-    return output
+    def lstm_recv_inside(feat):
+        forward, _ = fluid.layers.dynamic_lstm(
+            input=feat, size=hidden_dim * 4, use_peepholes=False)
+        output = fluid.layers.sequence_last_step(forward)
+        return output
+    return lstm_recv_inside
 
 
 def graphsage_sum(gw, feature, hidden_size, act, initializer, learning_rate, name):