Merge branch 'develop' of https://github.com/PaddlePaddle/models into develop

PaddleNLP/benchmark/bert/README.md

 # BERT Benchmark with Fleet API
 
-先配置运行环境
-export PYTHONPATH=/home/fangzeyang/PaddleNLP
-export DATA_DIR=/home/fangzeyang/bert_data/wikicorpus_en
+先配置运行环境，clone PaddleNLP的代码，同时下载好预训练的数据到相应的位置
+export PYTHONPATH=${HOME}/models/PaddleNLP
+export DATA_DIR=${HOME}/bert_data/wikicorpus_en
 
 ## NLP 任务中的Pretraining
 

PaddleNLP/benchmark/transformer/README.md

@@ -20,5 +20,5 @@ python3 train.py
 ``` shell
 cd dygraph/
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-python3 -m paddle.distributed.launch --selected_gpus=0,1,2,3,4,5,6,7 train.py
+python3 -m paddle.distributed.launch --gpus "0,1,2,3,4,5,6,7" train.py
 ```

PaddleNLP/examples/language_model/rnnlm/model.py

@@ -87,13 +87,6 @@ class CrossEntropyLossForLm(nn.Layer):
 
 class UpdateModel(paddle.callbacks.Callback):
     # This callback reset model hidden states and update learning rate before each epoch begins 
-    def __init__(self, base_lr, lr_decay, epoch_start_decay):
-        self.base_lr = base_lr
-        self.lr_decay = lr_decay
-        self.epoch_start_decay = epoch_start_decay
-
     def on_epoch_begin(self, epoch=None, logs=None):
         self.model.network.reset_states()
-        new_lr = self.base_lr * (self.lr_decay
-                                 **max(epoch + 1 - self.epoch_start_decay, 0.0))
-        self.model._optimizer.set_lr(new_lr)
+        

PaddleNLP/examples/language_model/rnnlm/train.py

@@ -13,8 +13,9 @@ paddle.seed(102)
 
 
 def create_data_loader(batch_size, num_steps, data_path):
-    train_ds, valid_ds, test_ds = PTBDataset.get_datasets(
-        [batch_size] * 3, [num_steps] * 3, ['train', 'eval', 'test'])
+    train_ds = PTBDataset(batch_size, num_steps, 'train')
+    valid_ds = PTBDataset(batch_size, num_steps, 'eval')
+    test_ds = PTBDataset(batch_size, num_steps, 'test')
 
     train_loader = DataLoader(train_ds, return_list=True, batch_size=None)
     valid_loader = DataLoader(valid_ds, return_list=True, batch_size=None)
@@ -40,15 +41,15 @@ def train(args):
     gloabl_norm_clip = paddle.nn.ClipGradByGlobalNorm(args.max_grad_norm)
     cross_entropy = CrossEntropyLossForLm()
     ppl_metric = Perplexity()
-    callback = UpdateModel(
-        base_lr=args.base_lr,
-        lr_decay=args.lr_decay,
-        epoch_start_decay=args.epoch_start_decay)
-
+    callback = UpdateModel()
+    scheduler = paddle.callbacks.LRScheduler(by_step=False, by_epoch=True)
     model = paddle.Model(network)
-    # FIXME(yuanxiaopeng): Use scheduler instead of callback
-    #scheduler = paddle.optimizer.lr.LambdaDecay(learning_rate=args.base_lr, lr_lambda=lambda x: args.lr_decay**max(x + 1 - args.epoch_start_decay, 0.0), verbose=True)
-    optimizer = paddle.optimizer.SGD(learning_rate=args.base_lr,
+
+    learning_rate = paddle.optimizer.lr.LambdaDecay(
+        learning_rate=args.base_lr,
+        lr_lambda=lambda x: args.lr_decay**max(x + 1 - args.epoch_start_decay, 0.0),
+        verbose=True)
+    optimizer = paddle.optimizer.SGD(learning_rate=learning_rate,
                                      parameters=model.parameters(),
                                      grad_clip=gloabl_norm_clip)
 
@@ -62,7 +63,7 @@ def train(args):
               eval_data=valid_loader,
               epochs=args.max_epoch,
               shuffle=False,
-              callbacks=[callback],
+              callbacks=[callback, scheduler],
               log_freq=max(1, len(train_loader) // 10))
 
     model.save(path='checkpoint/test')  # save for training

PaddleNLP/examples/machine_translation/transformer/README.md

@@ -66,7 +66,7 @@ python train.py
 
 ```sh
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-python -m paddle.distributed.launch --selected_gpus=0,1,2,3,4,5,6,7 train.py
+python -m paddle.distributed.launch --gpus "0,1,2,3,4,5,6,7" train.py
 ```
 
 

PaddleNLP/examples/machine_translation/transformer/reader.py

@@ -49,7 +49,7 @@ def create_infer_loader(args):
             prepare_infer_input,
             bos_idx=args.bos_idx,
             eos_idx=args.eos_idx,
-            pad_idx=args.eos_idx),
+            pad_idx=args.bos_idx),
         num_workers=0,
         return_list=True)
     data_loaders = (data_loader, batch_sampler.__len__)

PaddleNLP/paddlenlp/datasets/squad.py

@@ -445,6 +445,9 @@ class SQuAD(Dataset):
 
         self.examples = examples
 
+    def __len__(self):
+        return len(self.data)
+
     def __getitem__(self, idx):
         feature = self.data[idx]