Use new save/load api and change fluid.layers.data to fluid.data (#3547)

* Use new save/load api and change fluid.layers.data to fluid.data

* add check for paddle version

PaddleNLP/language_model/README.md

@@ -5,6 +5,8 @@
 ## 1. 任务说明
 本文主要介绍基于lstm的语言的模型的实现，给定一个输入词序列（中文分词、英文tokenize），计算其ppl（语言模型困惑度，用户表示句子的流利程度），基于循环神经网络语言模型的介绍可以[参阅论文](https://arxiv.org/abs/1409.2329)。相对于传统的方法，基于循环神经网络的方法能够更好的解决稀疏词的问题。
 
+**目前语言模型要求使用PaddlePaddle 1.6及以上版本或适当的develop版本。**
+
 同时推荐用户参考[IPython Notebook demo](https://aistudio.baidu.com/aistudio/projectDetail/122290)
 
 ## 2. 效果说明

PaddleNLP/language_model/args.py

@@ -72,6 +72,11 @@ def parse_args():
         type=str,
         default="models",
         help='dir of the saved model.')
+    parser.add_argument(
+        '--init_from_pretrain_model',
+        type=str,
+        default=None,
+        help='dir to init model.')
     parser.add_argument('--enable_ce', action='store_true')
     parser.add_argument('--batch_size', type=int, default=0, help='batch size')
     parser.add_argument('--max_epoch', type=int, default=0, help='max epoch')

PaddleNLP/language_model/run.sh

@@ -6,7 +6,8 @@ function run_train() {
     python train.py \
         --data_path data/simple-examples/data/ \
         --model_type small \
-        --use_gpu True \
+        --use_gpu True
+        #--init_from_pretrain_model models/0/params
 }
 
 run_train

PaddleNLP/language_model/train.py

@@ -40,7 +40,7 @@ import os
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
 from args import *
-from models.model_check import check_cuda
+from models.model_check import check_cuda, check_version
 from models.language_model import lm_model
 from config import RNNConfig
 import logging
@@ -88,7 +88,10 @@ def save_para_npz(train_prog, train_exe):
 def main():
     args = parse_args()
 
+    # check if set use_gpu=True in paddlepaddle cpu version
     check_cuda(args.use_gpu)
+    # check if paddlepaddle version is satisfied
+    check_version()
 
     logger = logging.getLogger("lm")
     logger.setLevel(logging.INFO)
@@ -168,6 +171,15 @@ def main():
     exe = Executor(place)
     exe.run(startup_program)
 
+    if args.init_from_pretrain_model:
+        if not os.path.exists(args.init_from_pretrain_model + '.pdparams'):
+            print(args.init_from_pretrain_model)
+            raise Warning("The pretrained params do not exist.")
+            return
+        fluid.load(main_program, args.init_from_pretrain_model)
+        print("finish initing model from pretrained params from %s" %
+              (args.init_from_pretrain_model))
+
     device_count = len(fluid.cuda_places()) if args.use_gpu else len(
         fluid.cpu_places())
 
@@ -280,7 +292,6 @@ def main():
                 epoch_id=epoch_id,
                 with_lr=True,
                 device_count=device_count)
-
             batch_start_time = time.time()
             fetch_outs = exe.run(train_program,
                                  feed=input_data_feed,
@@ -304,7 +315,6 @@ def main():
                 print(
                     "-- Epoch:[%d]; Batch:[%d]; Time: %.5f s; ppl: %.5f, lr: %.5f"
                     % (epoch_id, batch_id, batch_time, ppl[0], lr[0]))
-
         ppl = np.exp(total_loss / iters)
         return ppl
 
@@ -434,9 +444,9 @@ def main():
                     format(
                         len(valid_data), config.batch_size, config.num_steps))
 
-            save_model_dir = os.path.join(args.save_model_dir, str(epoch_id))
-            fluid.io.save_persistables(
-                executor=exe, dirname=save_model_dir, main_program=main_program)
+            save_model_dir = os.path.join(args.save_model_dir,
+                                          str(epoch_id), "params")
+            fluid.save(main_program, save_model_dir)
             print("Saved model to: %s.\n" % save_model_dir)
 
     with profile_context(args.profile):

PaddleNLP/models/language_model/lm_model.py

@@ -255,16 +255,10 @@ def lm_model(hidden_size,
         return real_res, last_hidden, last_cell
 
     batch_size_each = batch_size // fluid.core.get_cuda_device_count()
-    x = layers.data(
-        name="x",
-        shape=[batch_size_each, num_steps, 1],
-        dtype='int64',
-        append_batch_size=False)
-    y = layers.data(
-        name="y",
-        shape=[batch_size_each * num_steps, 1],
-        dtype='int64',
-        append_batch_size=False)
+    x = fluid.data(
+        name="x", shape=[batch_size_each, num_steps, 1], dtype='int64')
+    y = fluid.data(
+        name="y", shape=[batch_size_each * num_steps, 1], dtype='int64')
 
     if use_dataloader:
         dataloader = fluid.io.DataLoader.from_generator(
@@ -273,16 +267,14 @@ def lm_model(hidden_size,
             iterable=False,
             use_double_buffer=True)
 
-    init_hidden = layers.data(
+    init_hidden = fluid.data(
         name="init_hidden",
         shape=[num_layers, batch_size_each, hidden_size],
-        dtype='float32',
-        append_batch_size=False)
-    init_cell = layers.data(
+        dtype='float32')
+    init_cell = fluid.data(
         name="init_cell",
         shape=[num_layers, batch_size_each, hidden_size],
-        dtype='float32',
-        append_batch_size=False)
+        dtype='float32')
 
     init_cell.persistable = True
     init_hidden.persistable = True

PaddleNLP/models/model_check.py

@@ -22,6 +22,10 @@ def check_cuda(use_cuda, err = \
     "\nYou can not set use_cuda = True in the model because you are using paddlepaddle-cpu.\n \
     Please: 1. Install paddlepaddle-gpu to run your models on GPU or 2. Set use_cuda = False to run models on CPU.\n"
                                                                                                                      ):
+    """
+    Log error and exit when set use_gpu=true in paddlepaddle
+    cpu version.
+    """
     try:
         if use_cuda == True and fluid.is_compiled_with_cuda() == False:
             print(err)
@@ -29,6 +33,23 @@ def check_cuda(use_cuda, err = \
     except Exception as e:
         pass
 
+
+def check_version():
+    """
+    Log error and exit when the installed version of paddlepaddle is
+    not satisfied.
+    """
+    err = "PaddlePaddle version 1.6 or higher is required, " \
+        "or a suitable develop version is satisfied as well. \n" \
+        "Please make sure the version is good with your code." \
+
+    try:
+        fluid.require_version('1.6.0')
+    except Exception as e:
+        print(err)
+        sys.exit(1)
+
+
 def check_version():
     """
     Log error and exit when the installed version of paddlepaddle is