Merge pull request #1446 from kuke/dam_fixs

Fix test & feed type and disable grad clip temporarily

fluid/PaddleNLP/deep_attention_matching_net/model.py

@@ -25,7 +25,7 @@ class Net(object):
         # turns ids
         shapes = [[-1, self._max_turn_len, 1]
                   for i in six.moves.xrange(self._max_turn_num)]
-        dtypes = ["int32" for i in six.moves.xrange(self._max_turn_num)]
+        dtypes = ["int64" for i in six.moves.xrange(self._max_turn_num)]
         # turns mask
         shapes += [[-1, self._max_turn_len, 1]
                    for i in six.moves.xrange(self._max_turn_num)]
@@ -34,7 +34,7 @@ class Net(object):
         # response ids, response mask, label
         shapes += [[-1, self._max_turn_len, 1], [-1, self._max_turn_len, 1],
                    [-1, 1]]
-        dtypes += ["int32", "float32", "float32"]
+        dtypes += ["int64", "float32", "float32"]
 
         py_reader = fluid.layers.py_reader(
             capacity=capacity,
@@ -60,7 +60,7 @@ class Net(object):
         for i in six.moves.xrange(self._max_turn_num):
             name = "turn_%d" % i
             turn = fluid.layers.data(
-                name=name, shape=[self._max_turn_len, 1], dtype="int32")
+                name=name, shape=[self._max_turn_len, 1], dtype="int64")
             self.turns_data.append(turn)
             self._feed_names.append(name)
 
@@ -73,7 +73,7 @@ class Net(object):
             self._feed_names.append(name)
 
         self.response = fluid.layers.data(
-            name="response", shape=[self._max_turn_len, 1], dtype="int32")
+            name="response", shape=[self._max_turn_len, 1], dtype="int64")
         self.response_mask = fluid.layers.data(
             name="response_mask",
             shape=[self._max_turn_len, 1],
@@ -141,7 +141,7 @@ class Net(object):
                     mask_cache=mask_cache)
                 Hu_stack.append(Hu)
 
-            # cross attention 
+            # cross attention
             r_a_t_stack = []
             t_a_r_stack = []
             for index in six.moves.xrange(self._stack_num + 1):
@@ -183,7 +183,7 @@ class Net(object):
                 t_a_r = fluid.layers.concat(input=t_a_r_stack, axis=1)
                 r_a_t = fluid.layers.concat(input=r_a_t_stack, axis=1)
 
-            # sim shape: [batch_size, 2*(stack_num+1), max_turn_len, max_turn_len]    
+            # sim shape: [batch_size, 2*(stack_num+1), max_turn_len, max_turn_len]
             sim = fluid.layers.matmul(
                 x=t_a_r, y=r_a_t, transpose_y=True, alpha=1 / np.sqrt(200.0))
             sim_turns.append(sim)

fluid/PaddleNLP/deep_attention_matching_net/test_and_evaluate.py

@@ -126,6 +126,7 @@ def test(args):
     dam = Net(args.max_turn_num, args.max_turn_len, args.vocab_size,
               args.emb_size, args.stack_num, args.channel1_num,
               args.channel2_num)
+    dam.create_data_layers()
     loss, logits = dam.create_network()
 
     loss.persistable = True
@@ -144,7 +145,7 @@ def test(args):
             staircase=True))
     optimizer.minimize(loss)
 
-    # The fethced loss is wrong when mem opt is enabled 
+    # The fethced loss is wrong when mem opt is enabled
     fluid.memory_optimize(fluid.default_main_program())
 
     if args.use_cuda:
@@ -191,7 +192,8 @@ def test(args):
         feed_list = []
         for dev in six.moves.xrange(dev_count):
             index = it * dev_count + dev
-            feed_dict = reader.make_one_batch_input(test_batches, index)
+            batch_data = reader.make_one_batch_input(test_batches, index)
+            feed_dict = dict(zip(dam.get_feed_names(), batch_data))
             feed_list.append(feed_dict)
 
         predicts = test_exe.run(feed=feed_list, fetch_list=[logits.name])

fluid/PaddleNLP/deep_attention_matching_net/train_and_evaluate.py

@@ -203,8 +203,8 @@ def train(args):
             loss.persistable = True
             logits.persistable = True
             # gradient clipping
-            fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByValue(
-                max=1.0, min=-1.0))
+            #fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByValue(
+            #    max=1.0, min=-1.0))
 
             optimizer = fluid.optimizer.Adam(
                 learning_rate=fluid.layers.exponential_decay(
@@ -324,7 +324,7 @@ def train(args):
                 evaluate(score_path, result_file_path)
         return step
 
-    # train on one epoch with pyreader 
+    # train on one epoch with pyreader
     def train_with_pyreader(step):
         def data_provider():
             for index in six.moves.xrange(batch_num):