Merge pull request #1707 from frankwhzhang/fix_bug

fix bug

fluid/PaddleRec/gru4rec/README.md

@@ -79,7 +79,7 @@ SessionId    ItemId    Time
 2    214757407    1396850438.247
 ```
 
-数据格式需要转换 运行脚本
+数据格式需要转换, 运行脚本如下
 ```
 python convert_format.py
 ```
@@ -101,7 +101,7 @@ python convert_format.py
 
 根据训练和测试文件生成字典和对应的paddle输入文件
 
-注意需要将训练文件放到一个目录下面，测试文件放到一个目录下面,同时支持多训练文件
+需要将训练文件放到目录raw_train_data下，测试文件放到目录raw_test_data下，并生成对应的train_data，test_data和vocab.txt文件
 ```
 python text2paddle.py raw_train_data/ raw_test_data/ train_data test_data vocab.txt
 ```

fluid/PaddleRec/gru4rec/net.py

@@ -171,7 +171,8 @@ def train_cross_entropy_network(vocab_size, neg_size, hid_size, drop_out=0.2):
     ele_mul = fluid.layers.elementwise_mul(emb_label_drop, gru)
     red_sum = fluid.layers.reduce_sum(input=ele_mul, dim=1, keep_dim=True)
 
-    pre = fluid.layers.sequence_reshape(input=red_sum, new_dim=(neg_size + 1))
+    pre_ = fluid.layers.sequence_reshape(input=red_sum, new_dim=(neg_size + 1))
+    pre = fluid.layers.softmax(input=pre_)
 
     cost = fluid.layers.cross_entropy(input=pre, label=pos_label)
     cost_sum = fluid.layers.reduce_sum(input=cost)

fluid/PaddleRec/gru4rec/train_sample_neg.py

@@ -68,9 +68,11 @@ def train():
 
     # Train program
     if args.loss == 'bpr':
+        print('bpr loss')
         src, pos_label, label, avg_cost = net.train_bpr_network(
             neg_size=args.neg_size, vocab_size=vocab_size, hid_size=hid_size)
     else:
+        print('cross-entory loss')
         src, pos_label, label, avg_cost = net.train_cross_entropy_network(
             neg_size=args.neg_size, vocab_size=vocab_size, hid_size=hid_size)
 

fluid/PaddleRec/gru4rec/utils.py

@@ -45,8 +45,8 @@ def to_lodtensor_bpr(raw_data, neg_size, vocab_size, place):
     neg_data = np.tile(pos_data, neg_size)
     np.random.shuffle(neg_data)
     for ii in range(length * neg_size):
-        if neg_data[ii] == pos_data[ii / neg_size]:
-            neg_data[ii] = pos_data[length - 1 - ii / neg_size]
+        if neg_data[ii] == pos_data[ii // neg_size]:
+            neg_data[ii] = pos_data[length - 1 - ii // neg_size]
 
     label_data = np.column_stack(
         (pos_data.reshape(length, 1), neg_data.reshape(length, neg_size)))

fluid/PaddleRec/ssr/infer.py

@@ -81,7 +81,7 @@ def infer(args, vocab_size, test_reader):
         start_up_program = fluid.Program()
         with fluid.program_guard(main_program, start_up_program):
             acc = model(vocab_size, emb_size, hid_size)
-            for epoch in xrange(start_index, last_index + 1):
+            for epoch in range(start_index, last_index + 1):
                 copy_program = main_program.clone()
                 model_path = model_dir + "/epoch_" + str(epoch)
                 fluid.io.load_params(