SR-GNN: some optimization (#2676)

use all data
network:
    support variable batch_size
    add some shape comment
infer:
    use pyreader

PaddleRec/gnn/infer.py

@@ -19,6 +19,7 @@ import os
 import paddle
 import paddle.fluid as fluid
 import reader
+import network
 
 logging.basicConfig(format='%(asctime)s - %(levelname)s - %(message)s')
 logger = logging.getLogger("fluid")
@@ -31,6 +32,8 @@ def parse_args():
         '--model_path', type=str, default='./saved_model/', help="path of model parameters")
     parser.add_argument(
         '--test_path', type=str, default='./data/diginetica/test.txt', help='dir of test file')
+    parser.add_argument(
+        '--config_path', type=str, default='./data/diginetica/config.txt', help='dir of config')
     parser.add_argument(
         '--use_cuda', type=int, default=1, help='whether to use gpu')
     parser.add_argument(
@@ -39,40 +42,51 @@ def parse_args():
         '--start_index', type=int, default='0', help='start index')
     parser.add_argument(
         '--last_index', type=int, default='10', help='end index')
+    parser.add_argument(
+        '--hidden_size', type=int, default=100, help='hidden state size')
+    parser.add_argument(
+        '--step', type=int, default=1, help='gnn propogation steps')
     return parser.parse_args()
 
 
-def infer(epoch_num):
-    args = parse_args()
+def infer(args):
     batch_size = args.batch_size
+    items_num = reader.read_config(args.config_path)
     test_data = reader.Data(args.test_path, False)
     place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
+    loss, acc, py_reader, feed_datas = network.network(items_num, args.hidden_size, args.step)
+    exe.run(fluid.default_startup_program())
+    infer_program = fluid.default_main_program().clone(for_test=True)
 
-    model_path = args.model_path + "epoch_" + str(epoch_num)
-    try:
-        [infer_program, feed_names, fetch_targets] = fluid.io.load_inference_model(
-            model_path, exe)
-        feeder = fluid.DataFeeder(
-            feed_list=feed_names, place=place, program=infer_program)
+    for epoch_num in range(args.start_index, args.last_index + 1):
+        model_path = args.model_path + "epoch_" + str(epoch_num)
+        try:
+            if not os.path.exists(model_path):
+                raise ValueError()
+            fluid.io.load_persistables(executor=exe, dirname=model_path,
+                    main_program=infer_program)
 
-        loss_sum = 0.0
-        acc_sum = 0.0
-        count = 0
-        for data in test_data.reader(batch_size, batch_size, False)():
-            res = exe.run(infer_program,
-                          feed=feeder.feed(data),
-                          fetch_list=fetch_targets)
-            loss_sum += res[0]
-            acc_sum += res[1]
-            count += 1
-        logger.info("TEST --> loss: %.4lf, Recall@20: %.4lf" %
-                    (loss_sum / count, acc_sum / count))
-    except ValueError as e:
-        logger.info("TEST --> error: there is no model in " + model_path)
+            loss_sum = 0.0
+            acc_sum = 0.0
+            count = 0
+            py_reader.decorate_paddle_reader(test_data.reader(batch_size, batch_size*20, False))
+            py_reader.start()
+            try:
+                while True:
+                    res = exe.run(infer_program,
+                                  fetch_list=[loss.name, acc.name], use_program_cache=True)
+                    loss_sum += res[0]
+                    acc_sum += res[1]
+                    count += 1
+            except fluid.core.EOFException:
+                py_reader.reset()
+            logger.info("TEST --> loss: %.4lf, Recall@20: %.4lf" %
+                        (loss_sum / count, acc_sum / count))
+        except ValueError as e:
+            logger.info("TEST --> error: there is no model in " + model_path)
 
 
 if __name__ == "__main__":
     args = parse_args()
-    for index in range(args.start_index, args.last_index + 1):
-        infer(index)
+    infer(args)

PaddleRec/gnn/network.py

@@ -19,44 +19,37 @@ import paddle.fluid as fluid
 import paddle.fluid.layers as layers
 
 
-def network(batch_size, items_num, hidden_size, step):
+def network(items_num, hidden_size, step):
     stdv = 1.0 / math.sqrt(hidden_size)
 
     items = layers.data(
         name="items",
-        shape=[batch_size, items_num, 1],
-        dtype="int64",
-        append_batch_size=False)  #[bs, uniq_max, 1]
+        shape=[1, 1],
+        dtype="int64") #[batch_size, uniq_max, 1]
     seq_index = layers.data(
         name="seq_index",
-        shape=[batch_size, items_num],
-        dtype="int32",
-        append_batch_size=False)  #[-1(seq_max)*batch_size, 1]
+        shape=[1],
+        dtype="int32") #[batch_size, seq_max]
     last_index = layers.data(
         name="last_index",
-        shape=[batch_size],
-        dtype="int32",
-        append_batch_size=False)  #[batch_size, 1]
+        shape=[1],
+        dtype="int32") #[batch_size, 1]
     adj_in = layers.data(
         name="adj_in",
-        shape=[batch_size, items_num, items_num],
-        dtype="float32",
-        append_batch_size=False)
+        shape=[1,1],
+        dtype="float32") #[batch_size, seq_max, seq_max]
     adj_out = layers.data(
         name="adj_out",
-        shape=[batch_size, items_num, items_num],
-        dtype="float32",
-        append_batch_size=False)
+        shape=[1,1],
+        dtype="float32") #[batch_size, seq_max, seq_max]
     mask = layers.data(
         name="mask",
-        shape=[batch_size, -1, 1],
-        dtype="float32",
-        append_batch_size=False)
+        shape=[1, 1],
+        dtype="float32") #[batch_size, seq_max, 1]
     label = layers.data(
         name="label",
-        shape=[batch_size, 1],
-        dtype="int64",
-        append_batch_size=False)
+        shape=[1],
+        dtype="int64") #[batch_size, 1]
 
     datas = [items, seq_index, last_index, adj_in, adj_out, mask, label]
     py_reader = fluid.layers.create_py_reader_by_data(
@@ -71,11 +64,12 @@ def network(batch_size, items_num, hidden_size, step):
             initializer=fluid.initializer.Uniform(
                 low=-stdv, high=stdv)),
         size=[items_num, hidden_size])  #[batch_size, uniq_max, h]
+    items_emb_shape = layers.shape(items_emb)
 
     pre_state = items_emb
     for i in range(step):
         pre_state = layers.reshape(
-            x=pre_state, shape=[batch_size, -1, hidden_size])
+            x=pre_state, shape=[-1, 1, hidden_size], actual_shape=items_emb_shape)
         state_in = layers.fc(
             input=pre_state,
             name="state_in",
@@ -114,15 +108,20 @@ def network(batch_size, items_num, hidden_size, step):
                 x=pre_state, shape=[-1, hidden_size]),
             size=3 * hidden_size)
 
-    final_state = pre_state
+    final_state = pre_state #[batch_size * uniq_max, h]
+
+    seq_origin_shape = layers.assign(np.array([0,0,hidden_size-1]).astype("int32"))
+    seq_origin_shape += layers.shape(layers.unsqueeze(seq_index,[2])) #value: [batch_size, seq_max, h]
+    seq_origin_shape.stop_gradient = True
+
     seq_index = layers.reshape(seq_index, shape=[-1])
-    seq = layers.gather(final_state, seq_index)  #[batch_size*-1(seq_max), h]
+    seq = layers.gather(final_state, seq_index)  #[batch_size * seq_max, h]
     last = layers.gather(final_state, last_index)  #[batch_size, h]
 
     seq = layers.reshape(
-        seq, shape=[batch_size, -1, hidden_size])  #[batch_size, -1(seq_max), h]
+        seq, shape=[-1, 1, hidden_size], actual_shape=seq_origin_shape)  #[batch_size, seq_max, h]
     last = layers.reshape(
-        last, shape=[batch_size, hidden_size])  #[batch_size, h]
+        last, shape=[-1, hidden_size])  #[batch_size, h]
 
     seq_fc = layers.fc(
         input=seq,
@@ -133,7 +132,7 @@ def network(batch_size, items_num, hidden_size, step):
         num_flatten_dims=2,
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Uniform(
-            low=-stdv, high=stdv)))  #[batch_size, -1(seq_max), h]
+            low=-stdv, high=stdv)))  #[batch_size, seq_max, h]
     last_fc = layers.fc(
         input=last,
         name="last_fc",
@@ -146,18 +145,18 @@ def network(batch_size, items_num, hidden_size, step):
             low=-stdv, high=stdv)))  #[bathc_size, h]
 
     seq_fc_t = layers.transpose(
-        seq_fc, perm=[1, 0, 2])  #[-1(seq_max), batch_size, h]
+        seq_fc, perm=[1, 0, 2])  #[seq_max, batch_size, h]
     add = layers.elementwise_add(
-        seq_fc_t, last_fc)  #[-1(seq_max), batch_size, h]
+        seq_fc_t, last_fc)  #[seq_max, batch_size, h]
     b = layers.create_parameter(
         shape=[hidden_size],
         dtype='float32',
         default_initializer=fluid.initializer.Constant(value=0.0))  #[h]
-    add = layers.elementwise_add(add, b)  #[-1(seq_max), batch_size, h]
+    add = layers.elementwise_add(add, b)  #[seq_max, batch_size, h]
 
-    add_sigmoid = layers.sigmoid(add) #[-1(seq_max), batch_size, h] 
+    add_sigmoid = layers.sigmoid(add) #[seq_max, batch_size, h] 
     add_sigmoid = layers.transpose(
-        add_sigmoid, perm=[1, 0, 2])  #[batch_size, -1(seq_max), h]
+        add_sigmoid, perm=[1, 0, 2])  #[batch_size, seq_max, h]
 
     weight = layers.fc(
         input=add_sigmoid,
@@ -168,10 +167,10 @@ def network(batch_size, items_num, hidden_size, step):
         bias_attr=False,
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Uniform(
-                low=-stdv, high=stdv)))  #[batch_size, -1, 1]
+                low=-stdv, high=stdv)))  #[batch_size, seq_max, 1]
     weight *= mask
-    weight_mask = layers.elementwise_mul(seq, weight, axis=0)
-    global_attention = layers.reduce_sum(weight_mask, dim=1)
+    weight_mask = layers.elementwise_mul(seq, weight, axis=0) #[batch_size, seq_max, h]
+    global_attention = layers.reduce_sum(weight_mask, dim=1) #[batch_size, h]
 
     final_attention = layers.concat(
         [global_attention, last], axis=1)  #[batch_size, 2*h]

PaddleRec/gnn/reader.py

@@ -98,17 +98,20 @@ class Data():
                     cur_batch = cur_bg[i:i + batch_size]
                     yield self.make_data(cur_batch, batch_size)
 
-            #deal with the remaining, discard at most batch_size data
-            if group_remain < batch_size:
+            #deal with the last batch group
+            if group_remain == 0:
                 return
             remain_data = copy.deepcopy(self.input[-group_remain:])
             if train:
                 remain_data = sorted(
                     remain_data, key=lambda x: len(x[0]), reverse=True)
-            for i in range(0, batch_group_size, batch_size):
-                if i + batch_size <= len(remain_data):
+            for i in range(0, group_remain, batch_size):
+                if i + batch_size <= group_remain:
                     cur_batch = remain_data[i:i + batch_size]
                     yield self.make_data(cur_batch, batch_size)
+                else:
+                    cur_batch = remain_data[i:]
+                    yield self.make_data(cur_batch, group_remain % batch_size)
         return _reader
 
 

PaddleRec/gnn/train.py

@@ -71,7 +71,7 @@ def train():
 
     batch_size = args.batch_size
     items_num = reader.read_config(args.config_path)
-    loss, acc, py_reader, feed_datas = network.network(batch_size, items_num, args.hidden_size,
+    loss, acc, py_reader, feed_datas = network.network(items_num, args.hidden_size,
                                 args.step)
 
     data_reader = reader.Data(args.train_path, True)