test=develop (#4389)

PALM @ 5426f750

-Subproject commit 5426f75073cf5bd416622dbe71b146d3dc8fffb6

ERNIE @ 30b892e3

-Subproject commit 30b892e3c029bff706337f269e6c158b0a223f60

PaddleNLP/README.md

@@ -10,7 +10,7 @@
 
 - **丰富而全面的NLP任务支持：**
 
-  - PaddleNLP为您提供了多粒度，多场景的应用支持。涵盖了从[分词](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[词性标注](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[命名实体识别](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)等NLP基础技术，到[文本分类](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/sentiment_classification)，[文本相似度计算](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/similarity_net)，[语义表示](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK)，[文本生成](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleTextGEN)等NLP核心技术。同时，PaddleNLP还提供了针对常见NLP大型应用系统（如[阅读理解](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMRC)，[对话系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleDialogue)，[机器翻译系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMT)等）的特定核心技术和工具组件，模型和预训练参数等，让您在NLP领域畅通无阻。
+  - PaddleNLP为您提供了多粒度，多场景的应用支持。涵盖了从[分词](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[词性标注](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[命名实体识别](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)等NLP基础技术，到[文本分类](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/sentiment_classification)，[文本相似度计算](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/similarity_net)，[语义表示](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models)，[文本生成](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/seq2seq)等NLP核心技术。同时，PaddleNLP还提供了针对常见NLP大型应用系统（如[阅读理解](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_reading_comprehension)，[对话系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/dialogue_system)，[机器翻译系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_translation)等）的特定核心技术和工具组件，模型和预训练参数等，让您在NLP领域畅通无阻。
 
 - **稳定可靠的NLP模型和强大的预训练参数：**
 
@@ -55,11 +55,11 @@ cd models/PaddleNLP/sentiment_classification
 |                    **语言模型**                    | [Language_model](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/language_model) | 基于循环神经网络（RNN）的经典神经语言模型（neural language model）。 |
 |                 **情感分类**:fire:                 | [Senta](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/sentiment_classification)，[EmotionDetection](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/emotion_detection) | Senta（Sentiment Classification，简称Senta）和EmotionDetection两个项目分别提供了面向*通用场景*和*人机对话场景专用*的情感倾向性分析模型。 |
 |              **文本相似度计算**:fire:              | [SimNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/similarity_net) | SimNet，又称为Similarity Net，为您提供高效可靠的文本相似度计算工具和预训练模型。 |
-|                 **语义表示**:fire:                 | [PaddleLARK](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK) | PaddleLARK，全称为Paddle LAngauge Representation Toolkit，集成了ELMO，BERT，ERNIE 1.0，ERNIE 2.0，XLNet等热门中英文预训练模型。 |
-|                    **文本生成**                    | [PaddleTextGEN](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleTextGEN) | Paddle Text Generation为您提供了一些列经典文本生成模型案例，如vanilla seq2seq，seq2seq with attention，variational seq2seq模型等。 |
-|                    **阅读理解**                    | [PaddleMRC](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMRC) | PaddleMRC，全称为Paddle Machine Reading Comprehension，集合了百度在阅读理解领域相关的模型，工具，开源数据等一系列工作。包括DuReader (百度开源的基于真实搜索用户行为的中文大规模阅读理解数据集)，KT-Net (结合知识的阅读理解模型，SQuAD以及ReCoRD曾排名第一), D-Net (预训练-微调框架，在EMNLP2019 MRQA国际阅读理解评测获得第一)，等。 |
-|                    **对话系统**                    | [PaddleDialogue](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleDialogue) | 包括：1）DGU（Dialogue General Understanding，通用对话理解模型）覆盖了包括**检索式聊天系统**中context-response matching任务和**任务完成型对话系统**中**意图识别**，**槽位解析**，**状态追踪**等常见对话系统任务，在6项国际公开数据集中都获得了最佳效果。<br/> 2) knowledge-driven dialogue：百度开源的知识驱动的开放领域对话数据集，发表于ACL2019。<br/>3）ADEM（Auto Dialogue Evaluation Model）：对话自动评估模型，可用于自动评估不同对话生成模型的回复质量。 |
-|                    **机器翻译**                    | [PaddleMT](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMT) | 全称为Paddle Machine Translation，基于Transformer的经典机器翻译模型。 |
+|                 **语义表示**:fire:                 | [pretrain_langauge_models](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models) | 集成了ELMO，BERT，ERNIE 1.0，ERNIE 2.0，XLNet等热门中英文预训练模型。 |
+|                    **文本生成**                    | [seq2seq](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleTextGEN) | seq2seq为您提供了一些列经典文本生成模型案例，如vanilla seq2seq，seq2seq with attention，variational seq2seq模型等。 |
+|                    **阅读理解**                    | [machine_reading_comprehension](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_reading_comprehension) | Paddle Machine Reading Comprehension，集合了百度在阅读理解领域相关的模型，工具，开源数据等一系列工作。包括DuReader (百度开源的基于真实搜索用户行为的中文大规模阅读理解数据集)，KT-Net (结合知识的阅读理解模型，SQuAD以及ReCoRD曾排名第一), D-Net (预训练-微调框架，在EMNLP2019 MRQA国际阅读理解评测获得第一)，等。 |
+|                    **对话系统**                    | [dialogue_system](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/dialogue_system) | 包括：1）DGU（Dialogue General Understanding，通用对话理解模型）覆盖了包括**检索式聊天系统**中context-response matching任务和**任务完成型对话系统**中**意图识别**，**槽位解析**，**状态追踪**等常见对话系统任务，在6项国际公开数据集中都获得了最佳效果。<br/> 2) knowledge-driven dialogue：百度开源的知识驱动的开放领域对话数据集，发表于ACL2019。<br/>3）ADEM（Auto Dialogue Evaluation Model）：对话自动评估模型，可用于自动评估不同对话生成模型的回复质量。 |
+|                    **机器翻译**                    | [machine_translation](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_translation) | 全称为Paddle Machine Translation，基于Transformer的经典机器翻译模型。 |
 |                  **其他前沿工作**                  | [Research](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/Research) |                    百度最新前沿工作开源。                    |
 
 
@@ -70,13 +70,13 @@ cd models/PaddleNLP/sentiment_classification
 ```text
 .
 ├── Research                          # 百度NLP在research方面的工作集合
-├── PaddleMT                          # 机器翻译相关代码，数据，预训练模型
-├── PaddleDialogue                    # 对话系统相关代码，数据，预训练模型
-├── PaddleMRC                         # 阅读理解相关代码，数据，预训练模型
-├── PaddleLARK                        # 语言表示工具箱
+├── machine_translation               # 机器翻译相关代码，数据，预训练模型
+├── dialogue_system                   # 对话系统相关代码，数据，预训练模型
+├── machcine_reading_comprehension    # 阅读理解相关代码，数据，预训练模型
+├── pretrain_langauge_models          # 语言表示工具箱
 ├── language_model                    # 语言模型
 ├── lexical_analysis                  # LAC词法分析
-├── models                            # 共享网络
+├── shared_modules/models             # 共享网络
 │   ├── __init__.py
 │   ├── classification
 │   ├── dialogue_model_toolkit
@@ -87,7 +87,7 @@ cd models/PaddleNLP/sentiment_classification
 │   ├── representation
 │   ├── sequence_labeling
 │   └── transformer_encoder.py
-├── preprocess                        # 共享文本预处理工具
+├── shared_modules/preprocess        # 共享文本预处理工具
 │   ├── __init__.py
 │   ├── ernie
 │   ├── padding.py

PaddleNLP/dialogue_domain_classification/run_classifier.py

@@ -16,7 +16,6 @@
 # limitations under the License.
 """
 
-
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
@@ -40,43 +39,55 @@ import math
 np.random.seed(0)
 random.seed(0)
 
-
 parser = argparse.ArgumentParser(__doc__)
 DEV_COUNT = 1
 model_g = ArgumentGroup(parser, "model", "model configuration and paths.")
-model_g.add_arg("init_checkpoint", str, None, "Init checkpoint to resume training from.")
-model_g.add_arg("checkpoints", str, "./checkpoints", "Path to save checkpoints.")
+model_g.add_arg("init_checkpoint", str, None,
+                "Init checkpoint to resume training from.")
+model_g.add_arg("checkpoints", str, "./checkpoints",
+                "Path to save checkpoints.")
 model_g.add_arg("config_path", str, "./data/input/model.conf", "Model conf.")
 model_g.add_arg("build_dict", bool, False, "Build dict.")
 
 train_g = ArgumentGroup(parser, "training", "training options.")
 train_g.add_arg("cpu_num", int, 3, "Number of Threads.")
 train_g.add_arg("epoch", int, 100, "Number of epoches for training.")
-train_g.add_arg("learning_rate", float, 0.1, "Learning rate used to train with warmup.")
-train_g.add_arg("save_steps", int, 1000, "The steps interval to save checkpoints.")
-train_g.add_arg("validation_steps", int, 100, "The steps interval to evaluate model performance.")
+train_g.add_arg("learning_rate", float, 0.1,
+                "Learning rate used to train with warmup.")
+train_g.add_arg("save_steps", int, 1000,
+                "The steps interval to save checkpoints.")
+train_g.add_arg("validation_steps", int, 100,
+                "The steps interval to evaluate model performance.")
 train_g.add_arg("random_seed", int, 7, "random seed")
-train_g.add_arg("threshold", float, 0.1, "When the confidence exceeds the threshold, the corresponding label is given.")
+train_g.add_arg(
+    "threshold", float, 0.1,
+    "When the confidence exceeds the threshold, the corresponding label is given."
+)
 
 log_g = ArgumentGroup(parser, "logging", "logging related.")
 log_g.add_arg("skip_steps", int, 10, "The steps interval to print loss.")
 
-data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options")
+data_g = ArgumentGroup(parser, "data",
+                       "Data paths, vocab paths and data processing options")
 data_g.add_arg("data_dir", str, "./data/input/", "Path to training data.")
 data_g.add_arg("save_dir", str, "./data/output/", "Path to save.")
-data_g.add_arg("max_seq_len", int, 50, "Tokens' number of the longest seqence allowed.")
-data_g.add_arg("batch_size", int, 64, "The total number of examples in one batch for training.")
+data_g.add_arg("max_seq_len", int, 50,
+               "Tokens' number of the longest seqence allowed.")
+data_g.add_arg("batch_size", int, 64,
+               "The total number of examples in one batch for training.")
 
 run_type_g = ArgumentGroup(parser, "run_type", "running type options.")
 run_type_g.add_arg("use_cuda", bool, False, "If set, use GPU for training.")
 # run_type_g.add_arg("use_fast_executor", bool, False, "If set, use fast parallel executor (in experiment).")
-run_type_g.add_arg("do_train", bool, True, "Whether to perform evaluation on test data set.")
-run_type_g.add_arg("do_eval", bool, True, "Whether to perform evaluation on test data set.")
-run_type_g.add_arg("do_test", bool, True, "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("do_train", bool, True,
+                   "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("do_eval", bool, True,
+                   "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("do_test", bool, True,
+                   "Whether to perform evaluation on test data set.")
 args = parser.parse_args()
 
 
-
 def get_score(pred_result, label, eval_phase):
     """[get precision recall and f-score]
     
@@ -139,7 +150,7 @@ def train(args, train_exe, build_res, place):
     pred_label = build_res["pred_label"]
     label = build_res["label"]
     fetch_list = [cost.name, prediction.name, pred_label.name, label.name]
-    train_pyreader = build_res["train_pyreader"]
+    train_data_loader = build_res["train_data_loader"]
     train_prog = build_res["train_prog"]
     steps = 0
     time_begin = time.time()
@@ -147,22 +158,24 @@ def train(args, train_exe, build_res, place):
     logger.info("Begin training")
     for i in range(args.epoch):
         try:
-            for data in train_pyreader(): 
+            for data in train_data_loader():
                 avg_cost_np, avg_pred_np, pred_label, label = train_exe.run(feed=data, program=compiled_prog, \
                                                                             fetch_list=fetch_list)
                 steps += 1
                 if steps % int(args.skip_steps) == 0:
                     time_end = time.time()
                     used_time = time_end - time_begin
-                    get_score(pred_label, label, eval_phase = "Train")
+                    get_score(pred_label, label, eval_phase="Train")
                     logger.info('loss is {}'.format(avg_cost_np))
-                    logger.info("epoch: %d, step: %d, speed: %f steps/s" % (i, steps, args.skip_steps / used_time))
+                    logger.info("epoch: %d, step: %d, speed: %f steps/s" %
+                                (i, steps, args.skip_steps / used_time))
                     time_begin = time.time()
                 if steps % args.save_steps == 0:
                     save_path = os.path.join(args.checkpoints,
                                              "step_" + str(steps))
-                    fluid.io.save_persistables(train_exe, save_path, train_prog)
-                    logger.info("[save]step %d : save at %s" % (steps, save_path))
+                    fluid.io.save(train_prog, save_path)
+                    logger.info("[save]step %d : save at %s" %
+                                (steps, save_path))
                 if steps % args.validation_steps == 0:
                     if args.do_eval:
                         evaluate(args, test_exe, build_res, "eval")
@@ -173,11 +186,16 @@ def train(args, train_exe, build_res, place):
             logger.error("Train error : %s" % str(e))
             exit(1)
     save_path = os.path.join(args.checkpoints, "step_" + str(steps))
-    fluid.io.save_persistables(train_exe, save_path, train_prog)
+    fluid.io.save(train_prog, save_path)
     logger.info("[save]step %d : save at %s" % (steps, save_path))
 
 
-def evaluate(args, test_exe, build_res, eval_phase, save_result=False, id2intent=None):
+def evaluate(args,
+             test_exe,
+             build_res,
+             eval_phase,
+             save_result=False,
+             id2intent=None):
     """[evaluate on dev/test dataset]
     
     Arguments:
@@ -203,14 +221,14 @@ def evaluate(args, test_exe, build_res, eval_phase, save_result=False, id2intent
     total_cost, total_acc, pred_prob_list, pred_label_list, label_list = [], [], [], [], []
     if eval_phase == "eval":
         test_prog = build_res["eval_compiled_prog"]
-        test_pyreader = build_res["eval_pyreader"]
+        test_data_loader = build_res["eval_data_loader"]
     elif eval_phase == "test":
         test_prog = build_res["test_compiled_prog"]
-        test_pyreader = build_res["test_pyreader"]
+        test_data_loader = build_res["test_data_loader"]
     else:
         exit(1)
     logger.info("-----------------------------------------------------------")
-    for data in test_pyreader():
+    for data in test_data_loader():
         avg_cost_np, avg_pred_np, pred_label, label= test_exe.run(program=test_prog, fetch_list=fetch_list, feed=data, \
             return_numpy=True)
         total_cost.append(avg_cost_np)
@@ -219,13 +237,18 @@ def evaluate(args, test_exe, build_res, eval_phase, save_result=False, id2intent
         label_list.extend(label)
 
     if save_result:
-        logger.info("save result at : %s" % args.save_dir + "/" + eval_phase + ".rst")
+        logger.info("save result at : %s" % args.save_dir + "/" + eval_phase +
+                    ".rst")
         save_dir = args.save_dir
         if not os.path.exists(save_dir):
             logger.warning("save dir not exists, and create it")
             os.makedirs(save_dir)
-        fin = codecs.open(os.path.join(args.data_dir, eval_phase + ".txt"), "r", encoding="utf8")
-        fout = codecs.open(args.save_dir + "/" + eval_phase + ".rst", "w", encoding="utf8")
+        fin = codecs.open(
+            os.path.join(args.data_dir, eval_phase + ".txt"),
+            "r",
+            encoding="utf8")
+        fout = codecs.open(
+            args.save_dir + "/" + eval_phase + ".rst", "w", encoding="utf8")
         for line in pred_prob_list:
             query = fin.readline().rsplit("\t", 1)[0]
             res = []
@@ -245,9 +268,14 @@ def evaluate(args, test_exe, build_res, eval_phase, save_result=False, id2intent
     logger.info("-----------------------------------------------------------")
 
 
-
-def create_net(args, flow_data, class_dim, dict_dim, place, model_name="textcnn_net", is_infer=False):
-    """[create network and pyreader]
+def create_net(args,
+               flow_data,
+               class_dim,
+               dict_dim,
+               place,
+               model_name="textcnn_net",
+               is_infer=False):
+    """[create network and loader]
     
     Arguments:
         flow_data {[type]} -- [description]
@@ -266,11 +294,23 @@ def create_net(args, flow_data, class_dim, dict_dim, place, model_name="textcnn_
         model = textcnn_net_multi_label
     else:
         return
-    char_list = fluid.data(name="char", shape=[None, args.max_seq_len, 1], dtype="int64", lod_level=0)
-    label = fluid.data(name="label", shape=[None, class_dim], dtype="float32", lod_level=0)  # label data
-    reader = fluid.io.PyReader(feed_list=[char_list, label], capacity=args.batch_size * 10, iterable=True, \
+    char_list = fluid.data(
+        name="char",
+        shape=[None, args.max_seq_len, 1],
+        dtype="int64",
+        lod_level=0)
+    label = fluid.data(
+        name="label", shape=[None, class_dim], dtype="float32",
+        lod_level=0)  # label data
+    data_loader = fluid.io.DataLoader.from_generator(
+        feed_list=[char_list, label],
+        capacity=args.batch_size * 10,
+        iterable=True,
         return_list=False)
-    output = model(char_list, label, dict_dim,
+    output = model(
+        char_list,
+        label,
+        dict_dim,
         emb_dim=flow_data["model"]["emb_dim"],
         hid_dim=flow_data["model"]["hid_dim"],
         hid_dim2=flow_data["model"]["hid_dim2"],
@@ -281,14 +321,15 @@ def create_net(args, flow_data, class_dim, dict_dim, place, model_name="textcnn_
         max_seq_len=args.max_seq_len)
     if is_infer:
         prediction = output
-        return [reader, prediction]
+        return [data_loader, prediction]
     else:
-        avg_cost, prediction, pred_label, label = output[0], output[1], output[2], output[3]
-        return [reader, avg_cost, prediction, pred_label, label]
+        avg_cost, prediction, pred_label, label = output[0], output[1], output[
+            2], output[3]
+        return [data_loader, avg_cost, prediction, pred_label, label]
 
 
-def build_data_reader(args, char_dict, intent_dict):
-    """[decorate samples for pyreader]
+def build_data_loader(args, char_dict, intent_dict):
+    """[decorate samples for dataloader]
     
     Arguments:
         args {[type]} -- [description]
@@ -298,20 +339,22 @@ def build_data_reader(args, char_dict, intent_dict):
     Returns:
         [type] -- [description]
     """
-    reader_res = {}
+    loader_res = {}
     if args.do_train:
         train_processor = DataReader(char_dict, intent_dict, args.max_seq_len)
         train_data_generator = train_processor.prepare_data(
             data_path=args.data_dir + "train.txt",
             batch_size=args.batch_size,
             mode='train')
-        reader_res["train_data_generator"] = train_data_generator
+        loader_res["train_data_generator"] = train_data_generator
         num_train_examples = train_processor._get_num_examples()
         logger.info("Num train examples: %d" % num_train_examples)
         logger.info("Num train steps: %d" % (math.ceil(num_train_examples * 1.0 / args.batch_size) * \
                                             args.epoch // DEV_COUNT))
-        if math.ceil(num_train_examples * 1.0 / args.batch_size) // DEV_COUNT <= 0:
-            logger.error("Num of train steps is less than 0  or equals to 0, exit")
+        if math.ceil(num_train_examples * 1.0 /
+                     args.batch_size) // DEV_COUNT <= 0:
+            logger.error(
+                "Num of train steps is less than 0  or equals to 0, exit")
             exit(1)
     if args.do_eval:
         eval_processor = DataReader(char_dict, intent_dict, args.max_seq_len)
@@ -319,7 +362,7 @@ def build_data_reader(args, char_dict, intent_dict):
             data_path=args.data_dir + "eval.txt",
             batch_size=args.batch_size,
             mode='eval')
-        reader_res["eval_data_generator"] = eval_data_generator
+        loader_res["eval_data_generator"] = eval_data_generator
         num_eval_examples = eval_processor._get_num_examples()
         logger.info("Num eval examples: %d" % num_eval_examples)
     if args.do_test:
@@ -328,11 +371,12 @@ def build_data_reader(args, char_dict, intent_dict):
             data_path=args.data_dir + "test.txt",
             batch_size=args.batch_size,
             mode='test')
-        reader_res["test_data_generator"] = test_data_generator
-    return reader_res
+        loader_res["test_data_generator"] = test_data_generator
+    return loader_res
 
 
-def build_graph(args, model_config, num_labels, dict_dim, place, test_place, reader_res):
+def build_graph(args, model_config, num_labels, dict_dim, place, test_place,
+                loader_res):
     """[build paddle graph]
     
     Arguments:
@@ -341,7 +385,7 @@ def build_graph(args, model_config, num_labels, dict_dim, place, test_place, rea
         num_labels {[type]} -- [description]
         dict_dim {[type]} -- [description]
         place {[type]} -- [description]
-        reader_res {[type]} -- [description]
+        loader_res {[type]} -- [description]
     
     Returns:
         [type] -- [description]
@@ -358,36 +402,42 @@ def build_graph(args, model_config, num_labels, dict_dim, place, test_place, rea
     if args.do_train:
         with fluid.program_guard(train_prog, startup_prog):
             with fluid.unique_name.guard():
-                train_pyreader, cost, prediction, pred_label, label = create_net(args, model_config, num_labels, \
+                train_data_loader, cost, prediction, pred_label, label = create_net(args, model_config, num_labels, \
                                                             dict_dim, place, model_name="textcnn_net")
-                train_pyreader.decorate_sample_list_generator(reader_res['train_data_generator'], places=place)
-                res["train_pyreader"] = train_pyreader
-                sgd_optimizer = fluid.optimizer.SGD(learning_rate=fluid.layers.exponential_decay(
-                                learning_rate=args.learning_rate, decay_steps=1000, decay_rate=0.5, staircase=True))
+                train_data_loader.set_sample_list_generator(
+                    loader_res['train_data_generator'], places=place)
+                res["train_data_loader"] = train_data_loader
+                sgd_optimizer = fluid.optimizer.SGD(
+                    learning_rate=fluid.layers.exponential_decay(
+                        learning_rate=args.learning_rate,
+                        decay_steps=1000,
+                        decay_rate=0.5,
+                        staircase=True))
                 sgd_optimizer.minimize(cost)
     if args.do_eval:
         with fluid.program_guard(eval_prog, startup_prog):
             with fluid.unique_name.guard():
-                eval_pyreader, cost, prediction, pred_label, label = create_net(args, model_config, num_labels, \
+                eval_data_loader, cost, prediction, pred_label, label = create_net(args, model_config, num_labels, \
                                                              dict_dim, test_place, model_name="textcnn_net")
-                eval_pyreader.decorate_sample_list_generator(reader_res['eval_data_generator'], places=test_place)
-                res["eval_pyreader"] = eval_pyreader
+                eval_data_loader.set_sample_list_generator(
+                    loader_res['eval_data_generator'], places=test_place)
+                res["eval_data_loader"] = eval_data_loader
     if args.do_test:
         with fluid.program_guard(test_prog, startup_prog):
             with fluid.unique_name.guard():
-                test_pyreader, cost, prediction, pred_label, label = create_net(args, model_config, num_labels, \
+                test_data_loader, cost, prediction, pred_label, label = create_net(args, model_config, num_labels, \
                                                             dict_dim, test_place, model_name="textcnn_net")
-                test_pyreader.decorate_sample_list_generator(reader_res['test_data_generator'], places=test_place)
-                res["test_pyreader"] = test_pyreader
+                test_data_loader.set_sample_list_generator(
+                    loader_res['test_data_generator'], places=test_place)
+                res["test_data_loader"] = test_data_loader
     res["cost"] = cost
     res["prediction"] = prediction
     res["label"] = label
     res["pred_label"] = pred_label
-    res["train_prog"] =train_prog 
+    res["train_prog"] = train_prog
     res["eval_prog"] = eval_prog
     res["test_prog"] = test_prog
 
-  
     return res
 
 
@@ -421,8 +471,9 @@ def main(args):
         id2intent[int(value)] = key
     num_labels = len(intent_dict)
     # build model
-    reader_res = build_data_reader(args, char_dict, intent_dict)
-    build_res = build_graph(args, model_config, num_labels, dict_dim, place, test_place, reader_res)
+    loader_res = build_data_loader(args, char_dict, intent_dict)
+    build_res = build_graph(args, model_config, num_labels, dict_dim, place,
+                            test_place, loader_res)
     build_res["place"] = place
     build_res["test_place"] = test_place
     if not (args.do_train or args.do_eval or args.do_test):
@@ -432,11 +483,13 @@ def main(args):
     exe.run(startup_prog)
     if args.init_checkpoint and args.init_checkpoint != "None":
         try:
-            init_checkpoint(exe, args.init_checkpoint, main_program=startup_prog)
+            init_checkpoint(
+                exe, args.init_checkpoint, main_program=startup_prog)
             logger.info("Load model from %s" % args.init_checkpoint)
         except Exception as e:
             logger.exception(str(e))
-            logger.error("Faild load model from %s [%s]" % (args.init_checkpoint, str(e)))
+            logger.error("Faild load model from %s [%s]" %
+                         (args.init_checkpoint, str(e)))
     build_strategy = fluid.compiler.BuildStrategy()
     build_strategy.fuse_all_reduce_ops = False
     exec_strategy = fluid.ExecutionStrategy()
@@ -449,10 +502,12 @@ def main(args):
                                                                     exec_strategy=exec_strategy)
         build_res["compiled_prog"] = compiled_prog
     if args.do_test:
-        test_compiled_prog = fluid.compiler.CompiledProgram(build_res["test_prog"])
+        test_compiled_prog = fluid.compiler.CompiledProgram(build_res[
+            "test_prog"])
         build_res["test_compiled_prog"] = test_compiled_prog
     if args.do_eval:
-        eval_compiled_prog = fluid.compiler.CompiledProgram(build_res["eval_prog"])
+        eval_compiled_prog = fluid.compiler.CompiledProgram(build_res[
+            "eval_prog"])
         build_res["eval_compiled_prog"] = eval_compiled_prog
 
     if args.do_train:
@@ -465,7 +520,6 @@ def main(args):
                   save_result=True, id2intent=id2intent)
 
 
-
 if __name__ == "__main__":
     logger.info("the paddle version is %s" % paddle.__version__)
     check_version('1.6.0')

PaddleNLP/dialogue_domain_classification/utils.py

@@ -32,7 +32,6 @@ try:
 except ImportError:
     import ConfigParser as cp
 
-
 random_seed = 7
 logger = logging.getLogger()
 format = "%(asctime)s - %(name)s - %(levelname)s -%(filename)s-%(lineno)4d -%(message)s"
@@ -77,6 +76,7 @@ class ArgumentGroup(object):
     Arguments:
         object {[type]} -- [description]
     """
+
     def __init__(self, parser, title, des):
         self._group = parser.add_argument_group(title=title, description=des)
 
@@ -107,6 +107,7 @@ class DataReader(object):
     Returns:
         [type] -- [description]
     """
+
     def __init__(self, char_vocab, intent_dict, max_len):
         self._char_vocab = char_vocab
         self._intent_dict = intent_dict
@@ -128,12 +129,17 @@ class DataReader(object):
         #     word_dict_path), "The given word dictionary dose not exist."
         assert os.path.exists(data_path), "The given data file does not exist."
         if mode == "train":
-            train_reader = fluid.io.batch(paddle.reader.shuffle(self.data_reader(data_path, self.max_len, shuffle=True),
-                                        buf_size=batch_size * 100), batch_size)
+            train_reader = fluid.io.batch(
+                paddle.reader.shuffle(
+                    self.data_reader(
+                        data_path, self.max_len, shuffle=True),
+                    buf_size=batch_size * 100),
+                batch_size)
             # train_reader = fluid.io.batch(self.data_reader(data_path), batch_size)                   
             return train_reader
         else:
-            test_reader = fluid.io.batch(self.data_reader(data_path, self.max_len), batch_size)
+            test_reader = fluid.io.batch(
+                self.data_reader(data_path, self.max_len), batch_size)
             return test_reader
 
     def data_reader(self, file_path, max_len, shuffle=False):
@@ -150,7 +156,8 @@ class DataReader(object):
             char_id_list = list(map(lambda x: 0 if x not in self._char_vocab else int(self._char_vocab[x]), \
                             list(query)))
             if len(char_id_list) < max_len:
-                char_id_list.extend([self.padding_id] * (max_len - len(char_id_list)))
+                char_id_list.extend([self.padding_id] *
+                                    (max_len - len(char_id_list)))
             char_id_list = char_id_list[:max_len]
             intent_id_list = [self.padding_id] * self.intent_size
             for item in intent.split('\2'):
@@ -159,6 +166,7 @@ class DataReader(object):
         if shuffle:
             random.seed(random_seed)
             random.shuffle(self.all_data)
+
         def reader():
             """
             reader
@@ -166,6 +174,7 @@ class DataReader(object):
             for char_id_list, intent_id_list in self.all_data:
                 # print char_id_list, intent_id
                 yield char_id_list, intent_id_list
+
         return reader
 
 
@@ -178,6 +187,7 @@ class DataProcesser(object):
     Returns:
         [type] -- [description]
     """
+
     @staticmethod
     def read_dict(filename):
         """
@@ -227,7 +237,8 @@ class DataProcesser(object):
                     intent_dict[intent] = 0
                 intent_dict[intent] += 1
         #   save char dict
-        with codecs.open("%s/char.dict" % save_dir, "w", encoding="utf8") as f_out:
+        with codecs.open(
+                "%s/char.dict" % save_dir, "w", encoding="utf8") as f_out:
             f_out.write("PAD\0020\n")
             f_out.write("OOV\0021\n")
             char_id = 2
@@ -238,7 +249,8 @@ class DataProcesser(object):
                     f_out.write("%s\002%d\n" % (key, char_id))
                     char_id += 1
         #   save intent dict
-        with codecs.open("%s/domain.dict" % save_dir, "w", encoding="utf8") as f_out:
+        with codecs.open(
+                "%s/domain.dict" % save_dir, "w", encoding="utf8") as f_out:
             f_out.write("SYS_OTHER\0020\n")
             intent_id = 1
             for key, value in intent_dict.items():
@@ -249,7 +261,6 @@ class DataProcesser(object):
                     intent_id += 1
 
 
-
 class ConfigReader(object):
     """[read model config file]
     
@@ -282,49 +293,13 @@ class ConfigReader(object):
         return flow_data
 
 
-def init_pretraining_params(exe,
-                            pretraining_params_path,
-                            main_program,
-                            use_fp16=False):
-    """load params of pretrained model, NOT including moment, learning_rate"""
-    assert os.path.exists(pretraining_params_path
-                          ), "[%s] cann't be found." % pretraining_params_path
-
-    def _existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        return os.path.exists(os.path.join(pretraining_params_path, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        pretraining_params_path,
-        main_program=main_program,
-        predicate=_existed_params)
-    print("Load pretraining parameters from {}.".format(
-        pretraining_params_path))
-
-
 def init_checkpoint(exe, init_checkpoint_path, main_program):
     """
     Init CheckPoint
     """
-    assert os.path.exists(
-        init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
-
-    def existed_persitables(var):
-        """
-        If existed presitabels
-        """
-        if not fluid.io.is_persistable(var):
-            return False
-        return os.path.exists(os.path.join(init_checkpoint_path, var.name))
+    fluid.load(main_program, init_checkpoint_path, exe)
+    print("Load model from {}".format(init_checkpoint_path))
 
-    fluid.io.load_vars(
-        exe,
-        init_checkpoint_path,
-        main_program=main_program,
-        predicate=existed_persitables)
-    print ("Load model from {}".format(init_checkpoint_path))
 
 def print_arguments(args):
     """
@@ -350,5 +325,3 @@ def check_version(version='1.6.0'):
     except Exception as e:
         logger.error(err)
         sys.exit(1)
-
-

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/_ce.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/_ce.py

@@ -21,8 +21,10 @@ from kpi import DurationKpi
 
 train_loss_card1 = CostKpi('train_loss_card1', 0.03, 0, actived=True)
 train_loss_card4 = CostKpi('train_loss_card4', 0.03, 0, actived=True)
-train_duration_card1 = DurationKpi('train_duration_card1', 0.01, 0, actived=True)
-train_duration_card4 = DurationKpi('train_duration_card4', 0.01, 0, actived=True)
+train_duration_card1 = DurationKpi(
+    'train_duration_card1', 0.01, 0, actived=True)
+train_duration_card4 = DurationKpi(
+    'train_duration_card4', 0.01, 0, actived=True)
 
 tracking_kpis = [
     train_loss_card1,

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/ade/prepare_data_and_model.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/ade/prepare_data_and_model.py

@@ -20,22 +20,25 @@ import sys
 import io
 import os
 
-URLLIB=urllib
+URLLIB = urllib
 if sys.version_info >= (3, 0):
     import urllib.request
-    URLLIB=urllib.request
+    URLLIB = urllib.request
 
-DATA_MODEL_PATH = {"DATA_PATH": "https://baidu-nlp.bj.bcebos.com/auto_dialogue_evaluation_dataset-1.0.0.tar.gz", 
-                   "TRAINED_MODEL": "https://baidu-nlp.bj.bcebos.com/auto_dialogue_evaluation_models.2.0.0.tar.gz"} 
+DATA_MODEL_PATH = {
+    "DATA_PATH":
+    "https://baidu-nlp.bj.bcebos.com/auto_dialogue_evaluation_dataset-1.0.0.tar.gz",
+    "TRAINED_MODEL":
+    "https://baidu-nlp.bj.bcebos.com/auto_dialogue_evaluation_models.2.0.0.tar.gz"
+}
 
-PATH_MAP = {'DATA_PATH': "./data/input", 
-            'TRAINED_MODEL': './data/saved_models'}
+PATH_MAP = {'DATA_PATH': "./data/input", 'TRAINED_MODEL': './data/saved_models'}
 
 
 def un_tar(tar_name, dir_name):
     try:
         t = tarfile.open(tar_name)
-        t.extractall(path = dir_name)
+        t.extractall(path=dir_name)
         return True
     except Exception as e:
         print(e)
@@ -51,7 +54,8 @@ def download_model_and_data():
         shutil.rmtree(path)
     for path_key in DATA_MODEL_PATH:
         filename = os.path.basename(DATA_MODEL_PATH[path_key])
-        URLLIB.urlretrieve(DATA_MODEL_PATH[path_key], os.path.join("./", filename))
+        URLLIB.urlretrieve(DATA_MODEL_PATH[path_key],
+                           os.path.join("./", filename))
         state = un_tar(filename, PATH_MAP[path_key])
         if not state:
             print("Tar %s error....." % path_key)

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/ade/utils/save_load_io.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/ade/utils/save_load_io.py

@@ -122,5 +122,3 @@ def save_param(args, exe, program, dirname):
     print("save parameters at %s" % (os.path.join(param_dir, dirname)))
 
     return True
-
-

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/ade_net.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/ade_net.py

@@ -21,8 +21,7 @@ import paddle
 import paddle.fluid as fluid
 
 
-def create_net(
-    is_training,
+def create_net(is_training,
                model_input,
                args,
                clip_value=10.0,
@@ -52,14 +51,12 @@ def create_net(
             initializer=fluid.initializer.Normal(scale=0.1)))
 
     #fc to fit dynamic LSTM
-    context_fc = fluid.layers.fc(
-        input=context_emb,
+    context_fc = fluid.layers.fc(input=context_emb,
                                  size=args.hidden_size * 4,
                                  param_attr=fluid.ParamAttr(name='fc_weight'),
                                  bias_attr=fluid.ParamAttr(name='fc_bias'))
 
-    response_fc = fluid.layers.fc(
-        input=response_emb,
+    response_fc = fluid.layers.fc(input=response_emb,
                                   size=args.hidden_size * 4,
                                   param_attr=fluid.ParamAttr(name='fc_weight'),
                                   bias_attr=fluid.ParamAttr(name='fc_bias'))
@@ -106,7 +103,5 @@ def set_word_embedding(word_emb, place, word_emb_name="shared_word_emb"):
     """
     Set word embedding
     """
-    word_emb_param = fluid.global_scope().find_var(
-        word_emb_name).get_tensor()
+    word_emb_param = fluid.global_scope().find_var(word_emb_name).get_tensor()
     word_emb_param.set(word_emb, place)
-

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/inference_model.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/inference_model.py

@@ -42,22 +42,24 @@ def do_save_inference_model(args):
         with fluid.unique_name.guard():
 
             context_wordseq = fluid.data(
-                    name='context_wordseq', shape=[-1, 1], dtype='int64', lod_level=1)
+                name='context_wordseq',
+                shape=[-1, 1],
+                dtype='int64',
+                lod_level=1)
             response_wordseq = fluid.data(
-                    name='response_wordseq', shape=[-1, 1], dtype='int64', lod_level=1)
-            labels = fluid.data(
-                    name='labels', shape=[-1, 1], dtype='int64')
+                name='response_wordseq',
+                shape=[-1, 1],
+                dtype='int64',
+                lod_level=1)
+            labels = fluid.data(name='labels', shape=[-1, 1], dtype='int64')
 
             input_inst = [context_wordseq, response_wordseq, labels]
             input_field = InputField(input_inst)
-            data_reader = fluid.io.PyReader(feed_list=input_inst, 
-                        capacity=4, iterable=False)
+            data_reader = fluid.io.PyReader(
+                feed_list=input_inst, capacity=4, iterable=False)
 
             logits = create_net(
-                    is_training=False,
-                    model_input=input_field, 
-                    args=args
-                )
+                is_training=False, model_input=input_field, args=args)
 
     if args.use_cuda:
         place = fluid.CUDAPlace(0)
@@ -81,9 +83,7 @@ def do_save_inference_model(args):
             input_field.context_wordseq.name,
             input_field.response_wordseq.name,
         ],
-            target_vars=[
-                logits,
-            ],
+        target_vars=[logits, ],
         executor=exe,
         main_program=test_prog,
         model_filename="model.pdmodel",

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/main.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/main.py

@@ -26,7 +26,6 @@ from inference_model import do_save_inference_model
 
 from ade.utils.configure import PDConfig
 
-
 if __name__ == "__main__":
 
     args = PDConfig(yaml_file="./data/config/ade.yaml")

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/predict.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/predict.py

@@ -46,22 +46,24 @@ def do_predict(args):
         with fluid.unique_name.guard():
 
             context_wordseq = fluid.data(
-                    name='context_wordseq', shape=[-1, 1], dtype='int64', lod_level=1)
+                name='context_wordseq',
+                shape=[-1, 1],
+                dtype='int64',
+                lod_level=1)
             response_wordseq = fluid.data(
-                    name='response_wordseq', shape=[-1, 1], dtype='int64', lod_level=1)
-            labels = fluid.data(
-                    name='labels', shape=[-1, 1], dtype='int64')
+                name='response_wordseq',
+                shape=[-1, 1],
+                dtype='int64',
+                lod_level=1)
+            labels = fluid.data(name='labels', shape=[-1, 1], dtype='int64')
 
             input_inst = [context_wordseq, response_wordseq, labels]
             input_field = InputField(input_inst)
-            data_reader = fluid.io.PyReader(feed_list=input_inst, 
-                        capacity=4, iterable=False)
+            data_reader = fluid.io.PyReader(
+                feed_list=input_inst, capacity=4, iterable=False)
 
             logits = create_net(
-                    is_training=False,
-                    model_input=input_field, 
-                    args=args
-                )
+                is_training=False, model_input=input_field, args=args)
             logits.persistable = True
 
             fetch_list = [logits.name]
@@ -89,10 +91,7 @@ def do_predict(args):
         batch_size=args.batch_size)
 
     batch_generator = processor.data_generator(
-        place=place,
-        phase="test",
-        shuffle=False, 
-        sample_pro=1)
+        place=place, phase="test", shuffle=False, sample_pro=1)
     num_test_examples = processor.get_num_examples(phase='test')
 
     data_reader.decorate_batch_generator(batch_generator)
@@ -107,7 +106,7 @@ def do_predict(args):
             data_reader.reset()
             break
 
-    scores = scores[: num_test_examples]
+    scores = scores[:num_test_examples]
     print("Write the predicted results into the output_prediction_file")
     fw = io.open(args.output_prediction_file, 'w', encoding="utf8")
     for index, score in enumerate(scores):

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/train.py → PaddleNLP/dialogue_system/auto_dialogue_evaluation/train.py

@@ -49,22 +49,24 @@ def do_train(args):
 
         with fluid.unique_name.guard():
             context_wordseq = fluid.data(
-                    name='context_wordseq', shape=[-1, 1], dtype='int64', lod_level=1)
+                name='context_wordseq',
+                shape=[-1, 1],
+                dtype='int64',
+                lod_level=1)
             response_wordseq = fluid.data(
-                    name='response_wordseq', shape=[-1, 1], dtype='int64', lod_level=1)
-            labels = fluid.data(
-                    name='labels', shape=[-1, 1], dtype='int64')
+                name='response_wordseq',
+                shape=[-1, 1],
+                dtype='int64',
+                lod_level=1)
+            labels = fluid.data(name='labels', shape=[-1, 1], dtype='int64')
 
             input_inst = [context_wordseq, response_wordseq, labels]
             input_field = InputField(input_inst)
-            data_reader = fluid.io.PyReader(feed_list=input_inst, 
-                        capacity=4, iterable=False)
+            data_reader = fluid.io.PyReader(
+                feed_list=input_inst, capacity=4, iterable=False)
 
             loss = create_net(
-                    is_training=True,
-                    model_input=input_field, 
-                    args=args
-                )
+                is_training=True, model_input=input_field, args=args)
             loss.persistable = True
             # gradient clipping
             fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByValue(
@@ -74,7 +76,8 @@ def do_train(args):
 
             if args.use_cuda:
                 dev_count = fluid.core.get_cuda_device_count()
-                place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0')))
+                place = fluid.CUDAPlace(
+                    int(os.getenv('FLAGS_selected_gpus', '0')))
             else:
                 dev_count = int(os.environ.get('CPU_NUM', 1))
                 place = fluid.CPUPlace()
@@ -114,9 +117,14 @@ def do_train(args):
     if args.word_emb_init:
         print("start loading word embedding init ...")
         if six.PY2:
-            word_emb = np.array(pickle.load(io.open(args.word_emb_init, 'rb'))).astype('float32')
+            word_emb = np.array(
+                pickle.load(io.open(args.word_emb_init, 'rb'))).astype(
+                    'float32')
         else:
-            word_emb = np.array(pickle.load(io.open(args.word_emb_init, 'rb'), encoding="bytes")).astype('float32')
+            word_emb = np.array(
+                pickle.load(
+                    io.open(args.word_emb_init, 'rb'),
+                    encoding="bytes")).astype('float32')
         set_word_embedding(word_emb, place)
         print("finish init word embedding  ...")
 
@@ -147,15 +155,20 @@ def do_train(args):
                     used_time = time_end - time_begin
                     current_time = time.strftime('%Y-%m-%d %H:%M:%S',
                                                  time.localtime(time.time()))
-                    print('%s epoch: %d, step: %s, avg loss %s, speed: %f steps/s' % (current_time, epoch_step, steps, sum_loss / args.print_steps, args.print_steps / used_time))
+                    print(
+                        '%s epoch: %d, step: %s, avg loss %s, speed: %f steps/s'
+                        % (current_time, epoch_step, steps, sum_loss /
+                           args.print_steps, args.print_steps / used_time))
                     sum_loss = 0.0
                     time_begin = time.time()
 
                 if steps % args.save_steps == 0:
                     if args.save_checkpoint:
-                        save_load_io.save_checkpoint(args, exe, train_prog, "step_" + str(steps))
+                        save_load_io.save_checkpoint(args, exe, train_prog,
+                                                     "step_" + str(steps))
                     if args.save_param:
-                        save_load_io.save_param(args, exe, train_prog, "step_" + str(steps))
+                        save_load_io.save_param(args, exe, train_prog,
+                                                "step_" + str(steps))
                 steps += 1
             except fluid.core.EOFException:
                 data_reader.reset()

PaddleNLP/PaddleDialogue/dialogue_general_understanding/_ce.py → PaddleNLP/dialogue_system/dialogue_general_understanding/_ce.py

@@ -20,12 +20,18 @@ from kpi import CostKpi
 from kpi import DurationKpi
 from kpi import AccKpi
 
-each_step_duration_atis_slot_card1 = DurationKpi('each_step_duration_atis_slot_card1', 0.01, 0, actived=True)
-train_loss_atis_slot_card1 = CostKpi('train_loss_atis_slot_card1', 0.08, 0, actived=True)
-train_acc_atis_slot_card1 = CostKpi('train_acc_atis_slot_card1', 0.01, 0, actived=True)
-each_step_duration_atis_slot_card4 = DurationKpi('each_step_duration_atis_slot_card4', 0.06, 0, actived=True)
-train_loss_atis_slot_card4 = CostKpi('train_loss_atis_slot_card4', 0.03, 0, actived=True)
-train_acc_atis_slot_card4 = CostKpi('train_acc_atis_slot_card4', 0.01, 0, actived=True)
+each_step_duration_atis_slot_card1 = DurationKpi(
+    'each_step_duration_atis_slot_card1', 0.01, 0, actived=True)
+train_loss_atis_slot_card1 = CostKpi(
+    'train_loss_atis_slot_card1', 0.08, 0, actived=True)
+train_acc_atis_slot_card1 = CostKpi(
+    'train_acc_atis_slot_card1', 0.01, 0, actived=True)
+each_step_duration_atis_slot_card4 = DurationKpi(
+    'each_step_duration_atis_slot_card4', 0.06, 0, actived=True)
+train_loss_atis_slot_card4 = CostKpi(
+    'train_loss_atis_slot_card4', 0.03, 0, actived=True)
+train_acc_atis_slot_card4 = CostKpi(
+    'train_acc_atis_slot_card4', 0.01, 0, actived=True)
 
 tracking_kpis = [
     each_step_duration_atis_slot_card1,

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/batching.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/batching.py

@@ -100,8 +100,12 @@ def prepare_batch_data(task_name,
 
     if isinstance(insts[0][3], list):
         if task_name == "atis_slot":
-            labels_list = [inst[3] + [0] * (max_len - len(inst[3])) for inst in insts]
-            labels_list = [np.array(labels_list).astype("int64").reshape([-1, max_len])]
+            labels_list = [
+                inst[3] + [0] * (max_len - len(inst[3])) for inst in insts
+            ]
+            labels_list = [
+                np.array(labels_list).astype("int64").reshape([-1, max_len])
+            ]
         elif task_name == "dstc2":
             labels_list = [inst[3] for inst in insts]
             labels_list = [np.array(labels_list).astype("int64")]
@@ -124,10 +128,7 @@ def prepare_batch_data(task_name,
         out = batch_src_ids
     # Second step: padding
     src_id, self_input_mask = pad_batch_data(
-        out, 
-        max_len, 
-        pad_idx=pad_id, 
-        return_input_mask=True)
+        out, max_len, pad_idx=pad_id, return_input_mask=True)
     pos_id = pad_batch_data(
         batch_pos_ids,
         max_len,
@@ -163,13 +164,13 @@ def pad_batch_data(insts,
     corresponding position data and attention bias.
     """
     return_list = []
-    max_len = max_len_in if max_len_in != -1 else max(len(inst) for inst in insts)
+    max_len = max_len_in if max_len_in != -1 else max(
+        len(inst) for inst in insts)
     # Any token included in dict can be used to pad, since the paddings' loss
     # will be masked out by weights and make no effect on parameter gradients.
 
     inst_data = np.array(
-        [inst + list([pad_idx] * (max_len - len(inst))) for inst in insts
-    ])
+        [inst + list([pad_idx] * (max_len - len(inst))) for inst in insts])
     return_list += [inst_data.astype("int64").reshape([-1, max_len])]
 
     # position data

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/define_predict_pack.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/define_predict_pack.py

@@ -25,18 +25,21 @@ class DefinePredict(object):
     """
     Packaging Prediction Results
     """
+
     def __init__(self):
         """
         init
         """
-        self.task_map = {'udc': 'get_matching_res', 
+        self.task_map = {
+            'udc': 'get_matching_res',
             'swda': 'get_cls_res',
             'mrda': 'get_cls_res',
             'atis_intent': 'get_cls_res',
             'atis_slot': 'get_sequence_tagging',
             'dstc2': 'get_multi_cls_res',
             'dstc2_asr': 'get_multi_cls_res',
-                         'multi-woz': 'get_multi_cls_res'}
+            'multi-woz': 'get_multi_cls_res'
+        }
 
     def get_matching_res(self, probs, params=None):
         """
@@ -79,7 +82,3 @@ class DefinePredict(object):
             label_str = " ".join([str(l) for l in sorted(labels)])
 
         return label_str
-
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/prepare_data_and_model.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/prepare_data_and_model.py

@@ -20,25 +20,29 @@ import sys
 import io
 import os
 
-
-URLLIB=urllib
+URLLIB = urllib
 if sys.version_info >= (3, 0):
     import urllib.request
-    URLLIB=urllib.request
+    URLLIB = urllib.request
 
-DATA_MODEL_PATH = {"DATA_PATH": "https://baidu-nlp.bj.bcebos.com/dmtk_data_1.0.0.tar.gz", 
-                   "PRETRAIN_MODEL": "https://bert-models.bj.bcebos.com/uncased_L-12_H-768_A-12.tar.gz", 
-                   "TRAINED_MODEL": "https://baidu-nlp.bj.bcebos.com/dgu_models_2.0.0.tar.gz"} 
+DATA_MODEL_PATH = {
+    "DATA_PATH": "https://baidu-nlp.bj.bcebos.com/dmtk_data_1.0.0.tar.gz",
+    "PRETRAIN_MODEL":
+    "https://bert-models.bj.bcebos.com/uncased_L-12_H-768_A-12.tar.gz",
+    "TRAINED_MODEL": "https://baidu-nlp.bj.bcebos.com/dgu_models_2.0.0.tar.gz"
+}
 
-PATH_MAP = {'DATA_PATH': "./data/input", 
+PATH_MAP = {
+    'DATA_PATH': "./data/input",
     'PRETRAIN_MODEL': './data/pretrain_model',
-            'TRAINED_MODEL': './data/saved_models'}
+    'TRAINED_MODEL': './data/saved_models'
+}
 
 
 def un_tar(tar_name, dir_name):
     try:
         t = tarfile.open(tar_name)
-        t.extractall(path = dir_name)
+        t.extractall(path=dir_name)
         return True
     except Exception as e:
         print(e)
@@ -48,13 +52,18 @@ def un_tar(tar_name, dir_name):
 def download_model_and_data():
     print("Downloading dgu data, pretrain model and trained models......")
     print("This process is quite long, please wait patiently............")
-    for path in ['./data/input/data', './data/pretrain_model/uncased_L-12_H-768_A-12', './data/saved_models/trained_models']: 
+    for path in [
+            './data/input/data',
+            './data/pretrain_model/uncased_L-12_H-768_A-12',
+            './data/saved_models/trained_models'
+    ]:
         if not os.path.exists(path):
             continue
         shutil.rmtree(path)
     for path_key in DATA_MODEL_PATH:
         filename = os.path.basename(DATA_MODEL_PATH[path_key])
-        URLLIB.urlretrieve(DATA_MODEL_PATH[path_key], os.path.join("./", filename))
+        URLLIB.urlretrieve(DATA_MODEL_PATH[path_key],
+                           os.path.join("./", filename))
         state = un_tar(filename, PATH_MAP[path_key])
         if not state:
             print("Tar %s error....." % path_key)

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/README.md → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/README.md

@@ -19,6 +19,3 @@ python run_build_data.py udc
   python run_build_data.py atis
   生成槽位识别数据在dialogue_general_understanding/data/input/data/atis/atis_slot
   生成意图识别数据在dialogue_general_understanding/data/input/data/atis/atis_intent
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/build_atis_dataset.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/build_atis_dataset.py

@@ -12,7 +12,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 """build swda train dev test dataset"""
 
 import json
@@ -27,6 +26,7 @@ class ATIS(object):
     """
     nlu dataset atis data process
     """
+
     def __init__(self):
         """
         init instance
@@ -73,7 +73,8 @@ class ATIS(object):
             if example[1] not in self.intent_dict:
                 self.intent_dict[example[1]] = self.intent_id
                 self.intent_id += 1
-            fw.write(u"%s\t%s\n" % (self.intent_dict[example[1]], example[0].lower()))
+            fw.write(u"%s\t%s\n" %
+                     (self.intent_dict[example[1]], example[0].lower()))
 
         fw = io.open(self.map_tag_intent, 'w', encoding="utf8")
         for tag in self.intent_dict:
@@ -109,17 +110,19 @@ class ATIS(object):
                     tags_slot.append(str(self.slot_dict[tag]))
                 if i == 0:
                     if start not in [0, 1]:
-                        prefix_num = len(text[: start].strip().split())
+                        prefix_num = len(text[:start].strip().split())
                         tags.extend([str(self.slot_dict['O'])] * prefix_num)
                     tags.extend(tags_slot)
                 else:
-                    prefix_num = len(text[entities[i - 1]['end']: start].strip().split())
+                    prefix_num = len(text[entities[i - 1]['end']:start].strip()
+                                     .split())
                     tags.extend([str(self.slot_dict['O'])] * prefix_num)
                     tags.extend(tags_slot)
             if entities[-1]['end'] < len(text):
                 suffix_num = len(text[entities[-1]['end']:].strip().split())
                 tags.extend([str(self.slot_dict['O'])] * suffix_num)
-            fw.write(u"%s\t%s\n" % (text.encode('utf8'), " ".join(tags).encode('utf8')))
+            fw.write(u"%s\t%s\n" %
+                     (text.encode('utf8'), " ".join(tags).encode('utf8')))
 
         fw = io.open(self.map_tag_slot, 'w', encoding="utf8")
         for slot in self.slot_dict:
@@ -152,7 +155,3 @@ class ATIS(object):
 if __name__ == "__main__":
     atis_inst = ATIS()
     atis_inst.main()
-
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/build_dstc2_dataset.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/build_dstc2_dataset.py

@@ -28,6 +28,7 @@ class DSTC2(object):
     """
     dialogue state tracking dstc2 data process
     """
+
     def __init__(self):
         """
         init instance
@@ -49,7 +50,8 @@ class DSTC2(object):
         self.data_dict = commonlib.load_dict(self.data_list)
         for data_type in self.data_dict:
             for i in range(len(self.data_dict[data_type])):
-                self.data_dict[data_type][i] = os.path.join(self.src_dir, self.data_dict[data_type][i])
+                self.data_dict[data_type][i] = os.path.join(
+                    self.src_dir, self.data_dict[data_type][i])
 
     def _load_ontology(self):
         """
@@ -97,15 +99,25 @@ class DSTC2(object):
                 log_turn = log_json["turns"][i]
                 label_turn = label_json["turns"][i]
                 assert log_turn["turn-index"] == label_turn["turn-index"]
-                labels = ["%s_%s" % (slot, label_turn["goal-labels"][slot]) for slot in label_turn["goal-labels"]]
-                labels_ids = " ".join([str(self.map_tag_dict.get(label, self.map_tag_dict["%s_none" % label.split('_')[0]])) for label in labels])
+                labels = [
+                    "%s_%s" % (slot, label_turn["goal-labels"][slot])
+                    for slot in label_turn["goal-labels"]
+                ]
+                labels_ids = " ".join([
+                    str(
+                        self.map_tag_dict.get(label, self.map_tag_dict[
+                            "%s_none" % label.split('_')[0]]))
+                    for label in labels
+                ])
                 mach = log_turn['output']['transcript']
                 user = label_turn['transcription']
                 if not labels_ids.strip():
                     labels_ids = self.map_tag_dict['none']
                 out = "%s\t%s\1%s\t%s" % (session_id, mach, user, labels_ids)
-                user_asr = log_turn['input']['live']['asr-hyps'][0]['asr-hyp'].strip()
-                out_asr = "%s\t%s\1%s\t%s" % (session_id, mach, user_asr, labels_ids)
+                user_asr = log_turn['input']['live']['asr-hyps'][0][
+                    'asr-hyp'].strip()
+                out_asr = "%s\t%s\1%s\t%s" % (session_id, mach, user_asr,
+                                              labels_ids)
                 fw.write(u"%s\n" % out.encode('utf8'))
                 fw_asr.write(u"%s\n" % out_asr.encode('utf8'))
 
@@ -144,10 +156,7 @@ class DSTC2(object):
         self.get_test_dataset()
         self.get_labels()
 
+
 if __name__ == "__main__":
     dstc_inst = DSTC2()
     dstc_inst.main()
-
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/build_mrda_dataset.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/build_mrda_dataset.py

@@ -27,6 +27,7 @@ class MRDA(object):
     """
     dialogue act dataset mrda data process
     """
+
     def __init__(self):
         """
         init instance
@@ -67,7 +68,7 @@ class MRDA(object):
         for dadb_key in dadb_list:
             dadb_file = self.dadb_dict[dadb_key]
             fr = io.open(dadb_file, 'r', encoding="utf8")
-            row = csv.reader(fr, delimiter = ',')
+            row = csv.reader(fr, delimiter=',')
             for line in row:
                 elems = line
                 conv_id = elems[2]
@@ -87,7 +88,7 @@ class MRDA(object):
         for trans_key in trans_list:
             trans_file = self.trans_dict[trans_key]
             fr = io.open(trans_file, 'r', encoding="utf8")
-            row = csv.reader(fr, delimiter = ',')
+            row = csv.reader(fr, delimiter=',')
             for line in row:
                 elems = line
                 if len(elems) != 3:
@@ -120,7 +121,8 @@ class MRDA(object):
                 self.tag_id += 1
             caller = elem.split('_')[0].split('-')[-1]
             conv_no = elem.split('_')[0].split('-')[0]
-            out = "%s\t%s\t%s\t%s" % (conv_no, self.map_tag_dict[tag], caller, v_trans[0])
+            out = "%s\t%s\t%s\t%s" % (conv_no, self.map_tag_dict[tag], caller,
+                                      v_trans[0])
             fw.write(u"%s\n" % out)
 
     def get_train_dataset(self):
@@ -158,10 +160,7 @@ class MRDA(object):
         self.get_test_dataset()
         self.get_labels()
 
+
 if __name__ == "__main__":
     mrda_inst = MRDA()
     mrda_inst.main()
-
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/build_swda_dataset.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/build_swda_dataset.py

@@ -27,6 +27,7 @@ class SWDA(object):
     """
     dialogue act dataset swda data process
     """
+
     def __init__(self):
         """
         init instance
@@ -63,7 +64,7 @@ class SWDA(object):
             file_path = self.file_dict[name]
             fr = io.open(file_path, 'r', encoding="utf8")
             idx = 0
-            row = csv.reader(fr, delimiter = ',')
+            row = csv.reader(fr, delimiter=',')
             for r in row:
                 if idx == 0:
                     idx += 1
@@ -224,10 +225,7 @@ class SWDA(object):
         self.get_test_dataset()
         self.get_labels()
 
+
 if __name__ == "__main__":
     swda_inst = SWDA()
     swda_inst.main()
-
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/commonlib.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/commonlib.py

@@ -71,6 +71,3 @@ def load_voc(conf):
         elems = line.split('\t')
         map_dict[elems[0]] = elems[1]
     return map_dict
-
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/scripts/run_build_data.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/scripts/run_build_data.py

@@ -20,7 +20,6 @@ from build_dstc2_dataset import DSTC2
 from build_mrda_dataset import MRDA
 from build_swda_dataset import SWDA
 
-
 if __name__ == "__main__":
     task_name = sys.argv[1]
     task_name = task_name.lower()
@@ -38,11 +37,12 @@ if __name__ == "__main__":
     elif task_name == 'atis':
         atis_inst = ATIS()
         atis_inst.main()
-        shutil.copyfile("../../data/input/data/atis/atis_slot/test.txt", "../../data/input/data/atis/atis_slot/dev.txt")
-        shutil.copyfile("../../data/input/data/atis/atis_intent/test.txt", "../../data/input/data/atis/atis_intent/dev.txt")
+        shutil.copyfile("../../data/input/data/atis/atis_slot/test.txt",
+                        "../../data/input/data/atis/atis_slot/dev.txt")
+        shutil.copyfile("../../data/input/data/atis/atis_intent/test.txt",
+                        "../../data/input/data/atis/atis_intent/dev.txt")
     elif task_name == 'dstc2':
         dstc_inst = DSTC2()
         dstc_inst.main()
     else:
         exit(0)
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/tokenization.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/tokenization.py

@@ -12,7 +12,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 """Tokenization classes."""
 
 from __future__ import absolute_import

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/transformer_encoder.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/transformer_encoder.py

@@ -113,7 +113,7 @@ def multi_head_attention(queries,
         """
         Scaled Dot-Product Attention
         """
-        scaled_q = layers.scale(x=q, scale=d_key ** -0.5)
+        scaled_q = layers.scale(x=q, scale=d_key**-0.5)
         product = layers.matmul(x=scaled_q, y=k, transpose_y=True)
         if attn_bias:
             product += attn_bias

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu/utils/save_load_io.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu/utils/save_load_io.py

@@ -122,5 +122,3 @@ def save_param(args, exe, program, dirname):
     print("save parameters at %s" % (os.path.join(param_dir, dirname)))
 
     return True
-
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/dgu_net.py → PaddleNLP/dialogue_system/dialogue_general_understanding/dgu_net.py

@@ -23,12 +23,7 @@ from dgu.bert import BertModel
 from dgu.utils.configure import JsonConfig
 
 
-def create_net(
-        is_training,
-        model_input,
-        num_labels,
-        paradigm_inst,
-        args): 
+def create_net(is_training, model_input, num_labels, paradigm_inst, args):
     """create dialogue task model"""
 
     src_ids = model_input.src_ids
@@ -48,14 +43,15 @@ def create_net(
         config=bert_conf,
         use_fp16=False)
 
-    params = {'num_labels': num_labels,
+    params = {
+        'num_labels': num_labels,
         'src_ids': src_ids,
         'pos_ids': pos_ids,
         'sent_ids': sent_ids,
         'input_mask': input_mask,
         'labels': labels,
-              'is_training': is_training}
+        'is_training': is_training
+    }
 
     results = paradigm_inst.paradigm(bert, params)
     return results
-

PaddleNLP/PaddleDialogue/dialogue_general_understanding/inference_model.py → PaddleNLP/dialogue_system/dialogue_general_understanding/inference_model.py

@@ -66,7 +66,9 @@ def do_save_inference_model(args):
             sent_ids = fluid.data(
                 name='sent_ids', shape=[-1, args.max_seq_len], dtype='int64')
             input_mask = fluid.data(
-                        name='input_mask', shape=[-1, args.max_seq_len], dtype='float32')
+                name='input_mask',
+                shape=[-1, args.max_seq_len],
+                dtype='float32')
             if args.task_name == 'atis_slot':
                 labels = fluid.data(
                     name='labels', shape=[-1, args.max_seq_len], dtype='int64')
@@ -74,8 +76,7 @@ def do_save_inference_model(args):
                 labels = fluid.data(
                     name='labels', shape=[-1, num_labels], dtype='int64')
             else:
-                labels = fluid.data(
-                        name='labels', shape=[-1, 1], dtype='int64')
+                labels = fluid.data(name='labels', shape=[-1, 1], dtype='int64')
 
             input_inst = [src_ids, pos_ids, sent_ids, input_mask, labels]
             input_field = InputField(input_inst)
@@ -107,14 +108,10 @@ def do_save_inference_model(args):
     fluid.io.save_inference_model(
         args.inference_model_dir,
         feeded_var_names=[
-                input_field.src_ids.name, 
-                input_field.pos_ids.name,
-                input_field.sent_ids.name, 
-                input_field.input_mask.name
-            ],
-            target_vars=[
-                probs
+            input_field.src_ids.name, input_field.pos_ids.name,
+            input_field.sent_ids.name, input_field.input_mask.name
         ],
+        target_vars=[probs],
         executor=exe,
         main_program=test_prog,
         model_filename="model.pdmodel",

PaddleNLP/PaddleDialogue/dialogue_general_understanding/main.py → PaddleNLP/dialogue_system/dialogue_general_understanding/main.py

@@ -26,7 +26,6 @@ from inference_model import do_save_inference_model
 
 from dgu.utils.configure import PDConfig
 
-
 if __name__ == "__main__":
 
     args = PDConfig(yaml_file="./data/config/dgu.yaml")

PaddleNLP/PaddleDialogue/dialogue_general_understanding/train.py → PaddleNLP/dialogue_system/dialogue_general_understanding/train.py

@@ -66,7 +66,9 @@ def do_train(args):
             sent_ids = fluid.data(
                 name='sent_ids', shape=[-1, args.max_seq_len], dtype='int64')
             input_mask = fluid.data(
-                        name='input_mask', shape=[-1, args.max_seq_len], dtype='float32')
+                name='input_mask',
+                shape=[-1, args.max_seq_len],
+                dtype='float32')
             if args.task_name == 'atis_slot':
                 labels = fluid.data(
                     name='labels', shape=[-1, args.max_seq_len], dtype='int64')
@@ -74,13 +76,12 @@ def do_train(args):
                 labels = fluid.data(
                     name='labels', shape=[-1, num_labels], dtype='int64')
             else:
-                labels = fluid.data(
-                        name='labels', shape=[-1, 1], dtype='int64')
+                labels = fluid.data(name='labels', shape=[-1, 1], dtype='int64')
 
             input_inst = [src_ids, pos_ids, sent_ids, input_mask, labels]
             input_field = InputField(input_inst)
-            data_reader = fluid.io.PyReader(feed_list=input_inst, 
-                        capacity=4, iterable=False)
+            data_reader = fluid.io.PyReader(
+                feed_list=input_inst, capacity=4, iterable=False)
             processor = processors[task_name](data_dir=args.data_dir,
                                               vocab_path=args.vocab_path,
                                               max_seq_len=args.max_seq_len,
@@ -113,9 +114,7 @@ def do_train(args):
                 dev_count = int(os.environ.get('CPU_NUM', 1))
 
             batch_generator = processor.data_generator(
-                batch_size=args.batch_size,
-                phase='train',
-                shuffle=True)
+                batch_size=args.batch_size, phase='train', shuffle=True)
             num_train_examples = processor.get_num_examples(phase='train')
 
             if args.in_tokens:
@@ -217,37 +216,32 @@ def do_train(args):
                     current_time = time.strftime('%Y-%m-%d %H:%M:%S',
                                                  time.localtime(time.time()))
                     if accuracy is not None:
-                        print(
-                            "%s epoch: %d, step: %d, ave loss: %f, "
+                        print("%s epoch: %d, step: %d, ave loss: %f, "
                               "ave acc: %f, speed: %f steps/s" %
                               (current_time, epoch_step, steps,
-                             np.mean(np_loss),
-                             np.mean(np_acc),
+                               np.mean(np_loss), np.mean(np_acc),
                                args.print_steps / used_time))
                         ce_info.append([
-                            np.mean(np_loss),
-                            np.mean(np_acc),
+                            np.mean(np_loss), np.mean(np_acc),
                             args.print_steps / used_time
                         ])
                     else:
-                        print(
-                            "%s epoch: %d, step: %d, ave loss: %f, "
+                        print("%s epoch: %d, step: %d, ave loss: %f, "
                               "speed: %f steps/s" %
                               (current_time, epoch_step, steps,
-                             np.mean(np_loss),
-                             args.print_steps / used_time))
-                        ce_info.append([
-                            np.mean(np_loss),
-                            args.print_steps / used_time
-                        ])
+                               np.mean(np_loss), args.print_steps / used_time))
+                        ce_info.append(
+                            [np.mean(np_loss), args.print_steps / used_time])
                     time_begin = time.time()
 
                 if steps % args.save_steps == 0:
                     save_path = "step_" + str(steps)
                     if args.save_checkpoint:
-                        save_load_io.save_checkpoint(args, exe, train_prog, save_path)
+                        save_load_io.save_checkpoint(args, exe, train_prog,
+                                                     save_path)
                     if args.save_param:
-                        save_load_io.save_param(args, exe, train_prog, save_path)
+                        save_load_io.save_param(args, exe, train_prog,
+                                                save_path)
 
             except fluid.core.EOFException:
                 data_reader.reset()

PaddleNLP/emotion_detection/inference_model.py

@@ -19,8 +19,7 @@ from __future__ import print_function
 
 import os
 import sys
-sys.path.append("../")
-
+sys.path.append("../shared_modules/")
 import paddle
 import paddle.fluid as fluid
 import numpy as np

PaddleNLP/emotion_detection/run_classifier.py

@@ -11,7 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 """
 Emotion Detection Task
 """
@@ -24,7 +23,7 @@ import os
 import time
 import multiprocessing
 import sys
-sys.path.append("../")
+sys.path.append("../shared_modules/")
 
 import paddle
 import paddle.fluid as fluid
@@ -38,9 +37,7 @@ import reader
 import utils
 
 
-def create_model(args,
-                 num_labels,
-                 is_prediction=False):
+def create_model(args, num_labels, is_prediction=False):
     """
     Create Model for Emotion Detection
     """
@@ -77,10 +74,17 @@ def create_model(args,
         raise ValueError("Unknown network type!")
 
     if is_prediction:
-        probs = network(data, seq_len, None, args.vocab_size, class_dim=num_labels, is_prediction=True)
+        probs = network(
+            data,
+            seq_len,
+            None,
+            args.vocab_size,
+            class_dim=num_labels,
+            is_prediction=True)
         return loader, probs, [data.name, seq_len.name]
 
-    avg_loss, probs = network(data, seq_len, label, args.vocab_size, class_dim=num_labels)
+    avg_loss, probs = network(
+        data, seq_len, label, args.vocab_size, class_dim=num_labels)
     num_seqs = fluid.layers.create_tensor(dtype='int64')
     accuracy = fluid.layers.accuracy(input=probs, label=label, total=num_seqs)
     return loader, avg_loss, accuracy, num_seqs
@@ -142,7 +146,8 @@ def main(args):
     exe = fluid.Executor(place)
 
     task_name = args.task_name.lower()
-    processor = reader.EmoTectProcessor(data_dir=args.data_dir,
+    processor = reader.EmoTectProcessor(
+        data_dir=args.data_dir,
         vocab_path=args.vocab_path,
         random_seed=args.random_seed)
     #num_labels = len(processor.get_labels())
@@ -173,9 +178,7 @@ def main(args):
         with fluid.program_guard(train_program, startup_prog):
             with fluid.unique_name.guard():
                 train_loader, loss, accuracy, num_seqs = create_model(
-                    args,
-                    num_labels=num_labels,
-                    is_prediction=False)
+                    args, num_labels=num_labels, is_prediction=False)
 
                 sgd_optimizer = fluid.optimizer.Adagrad(learning_rate=args.lr)
                 sgd_optimizer.minimize(loss)
@@ -189,37 +192,27 @@ def main(args):
     if args.do_val:
         if args.do_train:
             test_data_generator = processor.data_generator(
-                batch_size=args.batch_size,
-                phase='dev',
-                epoch=1)
+                batch_size=args.batch_size, phase='dev', epoch=1)
         else:
             test_data_generator = processor.data_generator(
-                batch_size=args.batch_size,
-                phase='test',
-                epoch=1)
+                batch_size=args.batch_size, phase='test', epoch=1)
 
         test_prog = fluid.Program()
         with fluid.program_guard(test_prog, startup_prog):
             with fluid.unique_name.guard():
                 test_loader, loss, accuracy, num_seqs = create_model(
-                    args,
-                    num_labels=num_labels,
-                    is_prediction=False)
+                    args, num_labels=num_labels, is_prediction=False)
         test_prog = test_prog.clone(for_test=True)
 
     if args.do_infer:
         infer_data_generator = processor.data_generator(
-            batch_size=args.batch_size,
-            phase='infer',
-            epoch=1)
+            batch_size=args.batch_size, phase='infer', epoch=1)
 
         test_prog = fluid.Program()
         with fluid.program_guard(test_prog, startup_prog):
             with fluid.unique_name.guard():
                 infer_loader, probs, _ = create_model(
-                    args,
-                    num_labels=num_labels,
-                    is_prediction=True)
+                    args, num_labels=num_labels, is_prediction=True)
         test_prog = test_prog.clone(for_test=True)
 
     exe.run(startup_prog)
@@ -292,8 +285,9 @@ def main(args):
                     time_begin = time.time()
 
                 if steps % args.save_steps == 0:
-                    save_path = os.path.join(args.save_checkpoint_dir, "step_" + str(steps))
-                    fluid.io.save_persistables(exe, save_path, train_program)
+                    save_path = os.path.join(args.save_checkpoint_dir,
+                                             "step_" + str(steps))
+                    fluid.save(train_program, save_path)
 
                 if steps % args.validation_steps == 0:
                     # evaluate on dev set
@@ -306,11 +300,11 @@ def main(args):
                 print("final step: %d " % steps)
                 if args.do_val:
                     evaluate(test_exe, test_prog, test_loader,
-                        [loss.name, accuracy.name, num_seqs.name],
-                        "dev")
+                             [loss.name, accuracy.name, num_seqs.name], "dev")
 
-                save_path = os.path.join(args.save_checkpoint_dir, "step_" + str(steps))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                save_path = os.path.join(args.save_checkpoint_dir,
+                                         "step_" + str(steps))
+                fluid.save(train_program, save_path)
                 train_loader.reset()
                 break
 
@@ -334,15 +328,12 @@ def main(args):
     if not args.do_train and args.do_val:
         print("Final test result:")
         evaluate(test_exe, test_prog, test_loader,
-                 [loss.name, accuracy.name, num_seqs.name],
-                 "test")
+                 [loss.name, accuracy.name, num_seqs.name], "test")
 
     # infer
     if args.do_infer:
         print("Final infer result:")
-        infer(test_exe, test_prog, infer_loader,
-             [probs.name],
-             "infer")
+        infer(test_exe, test_prog, infer_loader, [probs.name], "infer")
 
 
 def get_cards():

PaddleNLP/emotion_detection/run_ernie_classifier.py

@@ -11,7 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 """
 Emotion Detection Task, based on ERNIE
 """
@@ -25,7 +24,7 @@ import time
 import argparse
 import multiprocessing
 import sys
-sys.path.append("../")
+sys.path.append("../shared_modules/")
 
 import paddle
 import paddle.fluid as fluid
@@ -350,7 +349,7 @@ def main(args):
 
                 if steps % args.save_steps == 0:
                     save_path = os.path.join(args.save_checkpoint_dir, "step_" + str(steps))
-                    fluid.io.save_persistables(exe, save_path, train_program)
+                    fluid.save(train_program, save_path)
 
                 if steps % args.validation_steps == 0:
                     # evaluate dev set
@@ -369,7 +368,7 @@ def main(args):
 
             except fluid.core.EOFException:
                 save_path = os.path.join(args.save_checkpoint_dir, "step_" + str(steps))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                fluid.save(train_program, save_path)
                 train_pyreader.reset()
                 break
 

PaddleNLP/emotion_detection/utils.py

@@ -11,7 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 """
 EmoTect utilities.
 """
@@ -29,27 +28,13 @@ import paddle
 import paddle.fluid as fluid
 import numpy as np
 
+
 def init_checkpoint(exe, init_checkpoint_path, main_program):
     """
     Init CheckPoint
     """
-    assert os.path.exists(
-        init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
-
-    def existed_persitables(var):
-        """
-        If existed presitabels
-        """
-        if not fluid.io.is_persistable(var):
-            return False
-        return os.path.exists(os.path.join(init_checkpoint_path, var.name))
 
-    fluid.io.load_vars(
-        exe,
-        init_checkpoint_path,
-        main_program=main_program,
-        predicate=existed_persitables)
-    print("Load model from {}".format(init_checkpoint_path))
+    fluid.load(main_program, init_checkpoint_path, exe)
 
 
 def word2id(word_dict, query):
@@ -57,8 +42,10 @@ def word2id(word_dict, query):
     Convert word sequence into id list
     """
     unk_id = len(word_dict)
-    wids = [word_dict[w] if w in word_dict else unk_id
-            for w in query.strip().split(" ")]
+    wids = [
+        word_dict[w] if w in word_dict else unk_id
+        for w in query.strip().split(" ")
+    ]
     return wids
 
 

PaddleNLP/language_model/README.md

@@ -5,7 +5,7 @@
 ## 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。**
+**目前语言模型要求使用PaddlePaddle 1.7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)
 

PaddleNLP/language_model/train.py

@@ -36,7 +36,7 @@ import sys
 if sys.version[0] == '2':
     reload(sys)
     sys.setdefaultencoding("utf-8")
-sys.path.append('../')
+sys.path.append('../shared_modules/')
 import os
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
@@ -60,7 +60,7 @@ def profile_context(profile=True, profiler_path='/tmp/paddingrnn.profile'):
 
 
 def get_current_model_para(train_prog, train_exe):
-    param_list = train_prog.block(0).all_parameters()
+    param_list = train_prog.all_parameters()
     param_name_list = [p.name for p in param_list]
 
     vals = {}
@@ -73,7 +73,7 @@ def get_current_model_para(train_prog, train_exe):
 
 def save_para_npz(train_prog, train_exe):
     print("begin to save model to model_base")
-    param_list = train_prog.block(0).all_parameters()
+    param_list = train_prog.all_parameters()
     param_name_list = [p.name for p in param_list]
 
     vals = {}

PaddleNLP/lexical_analysis/README.md

@@ -16,7 +16,7 @@ Lexical Analysis of Chinese，简称 LAC，是一个联合的词法分析模型
 
 #### 1.PaddlePaddle 安装
 
-本项目依赖 PaddlePaddle 1.6.0 及以上版本和PaddleHub 1.0.0及以上版本 ，PaddlePaddle安装请参考官网 [快速安装](http://www.paddlepaddle.org/paddle#quick-start)，PaddleHub安装参考 [PaddleHub](https://github.com/PaddlePaddle/PaddleHub)。
+本项目依赖 PaddlePaddle 1.7 及以上版本和PaddleHub 1.0.0及以上版本 ，PaddlePaddle安装请参考官网 [快速安装](http://www.paddlepaddle.org/paddle#quick-start)，PaddleHub安装参考 [PaddleHub](https://github.com/PaddlePaddle/PaddleHub)。
 
 > Warning: GPU 和 CPU 版本的 PaddlePaddle 分别是 paddlepaddle-gpu 和 paddlepaddle，请安装时注意区别。
 

PaddleNLP/lexical_analysis/creator.py

@@ -26,7 +26,7 @@ from paddle.fluid.initializer import NormalInitializer
 from reader import Dataset
 from ernie_reader import SequenceLabelReader
 
-sys.path.append("..")
+sys.path.append("../shared_modules/")
 from models.sequence_labeling import nets
 from models.representation.ernie import ernie_encoder, ernie_pyreader
 
@@ -35,9 +35,10 @@ def create_model(args, vocab_size, num_labels, mode='train'):
     """create lac model"""
 
     # model's input data
-    words = fluid.data(name='words', shape=[-1, 1], dtype='int64', lod_level=1)
+    words = fluid.data(
+        name='words', shape=[None, 1], dtype='int64', lod_level=1)
     targets = fluid.data(
-        name='targets', shape=[-1, 1], dtype='int64', lod_level=1)
+        name='targets', shape=[None, 1], dtype='int64', lod_level=1)
 
     # for inference process
     if mode == 'infer':
@@ -88,9 +89,11 @@ def create_pyreader(args,
                     return_reader=False,
                     mode='train'):
     # init reader
+    device_count = len(fluid.cuda_places()) if args.use_cuda else len(
+        fluid.cpu_places())
 
     if model == 'lac':
-        pyreader = fluid.io.PyReader(
+        pyreader = fluid.io.DataLoader.from_generator(
             feed_list=feed_list,
             capacity=50,
             use_double_buffer=True,
@@ -101,19 +104,19 @@ def create_pyreader(args,
 
         # create lac pyreader
         if mode == 'train':
-            pyreader.decorate_sample_list_generator(
+            pyreader.set_sample_list_generator(
                 fluid.io.batch(
                     fluid.io.shuffle(
                         reader.file_reader(file_name),
                         buf_size=args.traindata_shuffle_buffer),
-                    batch_size=args.batch_size),
+                    batch_size=args.batch_size / device_count),
                 places=place)
         else:
-            pyreader.decorate_sample_list_generator(
+            pyreader.set_sample_list_generator(
                 fluid.io.batch(
                     reader.file_reader(
                         file_name, mode=mode),
-                    batch_size=args.batch_size),
+                    batch_size=args.batch_size / device_count),
                 places=place)
 
     elif model == 'ernie':
@@ -162,19 +165,19 @@ def create_ernie_model(args, ernie_config):
     # ERNIE's input data
 
     src_ids = fluid.data(
-        name='src_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
+        name='src_ids', shape=[None, args.max_seq_len, 1], dtype='int64')
     sent_ids = fluid.data(
-        name='sent_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
+        name='sent_ids', shape=[None, args.max_seq_len, 1], dtype='int64')
     pos_ids = fluid.data(
-        name='pos_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
+        name='pos_ids', shape=[None, args.max_seq_len, 1], dtype='int64')
     input_mask = fluid.data(
-        name='input_mask', shape=[-1, args.max_seq_len, 1], dtype='float32')
+        name='input_mask', shape=[None, args.max_seq_len, 1], dtype='float32')
 
     padded_labels = fluid.data(
-        name='padded_labels', shape=[-1, args.max_seq_len, 1], dtype='int64')
+        name='padded_labels', shape=[None, args.max_seq_len, 1], dtype='int64')
 
     seq_lens = fluid.data(
-        name='seq_lens', shape=[-1], dtype='int64', lod_level=0)
+        name='seq_lens', shape=[None], dtype='int64', lod_level=0)
 
     squeeze_labels = fluid.layers.squeeze(padded_labels, axes=[-1])
 

PaddleNLP/lexical_analysis/ernie_reader.py

@@ -20,7 +20,7 @@ import sys
 from collections import namedtuple
 import numpy as np
 
-sys.path.append("..")
+sys.path.append("../shared_modules/")
 from preprocess.ernie.task_reader import BaseReader, tokenization
 
 

PaddleNLP/lexical_analysis/eval.py

@@ -24,7 +24,7 @@ import paddle
 import utils
 import reader
 import creator
-sys.path.append('../models/')
+sys.path.append('../shared_modules/models/')
 from model_check import check_cuda
 from model_check import check_version
 

PaddleNLP/lexical_analysis/inference_model.py

@@ -10,7 +10,7 @@ import paddle.fluid as fluid
 import creator
 import reader
 import utils
-sys.path.append('../models/')
+sys.path.append('../shared_modules/models/')
 from model_check import check_cuda
 from model_check import check_version
 

PaddleNLP/lexical_analysis/predict.py

@@ -24,7 +24,7 @@ import paddle
 import utils
 import reader
 import creator
-sys.path.append('../models/')
+sys.path.append('../shared_modules/models/')
 from model_check import check_cuda
 from model_check import check_version
 

PaddleNLP/lexical_analysis/run_ernie_sequence_labeling.py

@@ -34,7 +34,7 @@ import paddle.fluid as fluid
 
 import creator
 import utils
-sys.path.append("..")
+sys.path.append("../shared_modules/")
 from models.representation.ernie import ErnieConfig
 from models.model_check import check_cuda
 from models.model_check import check_version
@@ -188,15 +188,16 @@ def do_train(args):
 
             if steps % args.save_steps == 0:
                 save_path = os.path.join(args.model_save_dir,
-                                         "step_" + str(steps))
+                                         "step_" + str(steps), "checkpoint")
                 print("\tsaving model as %s" % (save_path))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                fluid.save(train_program, save_path)
 
             if steps % args.validation_steps == 0:
                 evaluate(exe, test_program, test_pyreader, train_ret)
 
-    save_path = os.path.join(args.model_save_dir, "step_" + str(steps))
-    fluid.io.save_persistables(exe, save_path, train_program)
+    save_path = os.path.join(args.model_save_dir, "step_" + str(steps),
+                             "checkpoint")
+    fluid.save(train_program, save_path)
 
 
 def do_eval(args):

PaddleNLP/lexical_analysis/train.py

@@ -29,7 +29,7 @@ import reader
 import utils
 import creator
 from eval import test_process
-sys.path.append('../models/')
+sys.path.append('../shared_modules/models/')
 from model_check import check_cuda
 from model_check import check_version
 
@@ -151,8 +151,8 @@ def do_train(args):
             # save checkpoints
             if step % args.save_steps == 0 and step != 0:
                 save_path = os.path.join(args.model_save_dir,
-                                         "step_" + str(step))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                                         "step_" + str(step), "checkpoint")
+                fluid.save(train_program, save_path)
             step += 1
 
     if args.enable_ce:

PaddleNLP/lexical_analysis/utils.py

@@ -200,19 +200,11 @@ def init_checkpoint(exe, init_checkpoint_path, main_program):
     assert os.path.exists(
         init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
 
-    def existed_persitables(var):
-        """
-        If existed presitabels
-        """
-        if not fluid.io.is_persistable(var):
-            return False
-        return os.path.exists(os.path.join(init_checkpoint_path, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        init_checkpoint_path,
-        main_program=main_program,
-        predicate=existed_persitables)
+    try:
+        checkpoint_path = os.path.join(init_checkpoint_path, "checkpoint")
+        fluid.load(main_program, checkpoint_path, exe)
+    except:
+        fluid.load(main_program, init_checkpoint_path, exe)
     print("Load model from {}".format(init_checkpoint_path))
 
 
@@ -224,15 +216,6 @@ def init_pretraining_params(exe,
     assert os.path.exists(pretraining_params_path
                           ), "[%s] cann't be found." % pretraining_params_path
 
-    def _existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        return os.path.exists(os.path.join(pretraining_params_path, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        pretraining_params_path,
-        main_program=main_program,
-        predicate=_existed_params)
+    fluid.load(main_program, pretraining_params_path, exe)
     print("Load pretraining parameters from {}.".format(
         pretraining_params_path))

PaddleNLP/PaddleMRC/README.md → PaddleNLP/machine_reading_comprehension/README.md

@@ -39,4 +39,3 @@ D-NET是一个以提升**阅读理解模型泛化能力**为目标的“预训
 - 在微调阶段引入多任务、多领域的学习策略 (基于[PALM](https://github.com/PaddlePaddle/PALM)多任务学习框架)，有效的提升了模型在不同领域的泛化能力
 
 百度利用D-NET框架在EMNLP 2019 [MRQA](https://mrqa.github.io/shared)国际阅读理解评测中以超过第二名近两个百分点的成绩夺得冠军，同时，在全部12个测试数据集中的10个排名第一。
-

PaddleNLP/PaddleMT/transformer/README.md → PaddleNLP/machine_translation/transformer/README.md

@@ -106,7 +106,7 @@ python -u main.py \
   --prepostprocess_dropout 0.3
 ```
 
-训练时默认使用所有 GPU，可以通过 `CUDA_VISIBLE_DEVICES` 环境变量来设置使用的 GPU 数目。也可以只使用 CPU 训练(通过参数 `--use_cuda False` 设置)，训练速度相对较慢。在执行训练时若提供了 `save_param` 和 `save_checkpoint`（默认为 trained_params 和 trained_ckpts），则每隔一定 iteration 后(通过参数 `save_step` 设置，默认为10000)将分别保存当前训练的参数值和 checkpoint 到相应目录，每隔一定数目的 iteration (通过参数 `print_step` 设置，默认为100)将打印如下的日志到标准输出：
+训练时默认使用所有 GPU，可以通过 `CUDA_VISIBLE_DEVICES` 环境变量来设置使用的 GPU 数目。也可以只使用 CPU 训练(通过参数 `--use_cuda False` 设置)，训练速度相对较慢。在执行训练时若提供了 `save_model_path`（默认为 saved_models），则每隔一定 iteration 后(通过参数 `save_step` 设置，默认为10000)将保存当前训练的 checkpoint 到相应目录（会保存分别记录了模型参数和优化器状态的 `transformer.pdparams` 和 `transformer.pdopt` 两个文件），每隔一定数目的 iteration (通过参数 `print_step` 设置，默认为100)将打印如下的日志到标准输出：
 
 ```txt
 [2019-08-02 15:30:51,656 INFO train.py:262] step_idx: 150100, epoch: 32, batch: 1364, avg loss: 2.880427, normalized loss: 1.504687, ppl: 17.821888, speed: 3.34 step/s
@@ -195,7 +195,7 @@ BLEU = 26.35, 57.7/32.1/20.0/13.0 (BP=1.000, ratio=1.013, hyp_len=63903, ref_len
 
 ### 预训练模型
 
-我们这里提供了对应有以上 BLEU 值的 [base model](https://transformer-res.bj.bcebos.com/base_model_params.tar.gz) 和 [big model](https://transformer-res.bj.bcebos.com/big_model_params.tar.gz) 的模型参数提供下载使用（注意，模型使用了提供下载的数据进行训练和测试）。
+我们这里提供了对应有以上 BLEU 值的 [base model](https://transformer-res.bj.bcebos.com/base_model_graph.tar.gz) 和 [big model](https://transformer-res.bj.bcebos.com/big_model_graph.tar.gz) 的模型参数提供下载使用（注意，模型使用了提供下载的数据进行训练和测试）。
 
 ## 进阶使用
 

PaddleNLP/PaddleMT/transformer/desc.py → PaddleNLP/machine_translation/transformer/desc.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+
 def get_input_descs(args):
     """
     Generate a dict mapping data fields to the corresponding data shapes and
@@ -42,7 +43,8 @@ def get_input_descs(args):
         # encoder.
         # The actual data shape of src_slf_attn_bias is:
         # [batch_size, n_head, max_src_len_in_batch, max_src_len_in_batch]
-        "src_slf_attn_bias": [(batch_size, n_head, seq_len, seq_len), "float32"],
+        "src_slf_attn_bias":
+        [(batch_size, n_head, seq_len, seq_len), "float32"],
         # The actual data shape of trg_word is:
         # [batch_size, max_trg_len_in_batch, 1]
         "trg_word": [(batch_size, seq_len), "int64",
@@ -54,12 +56,14 @@ def get_input_descs(args):
         # subsequent words in the decoder.
         # The actual data shape of trg_slf_attn_bias is:
         # [batch_size, n_head, max_trg_len_in_batch, max_trg_len_in_batch]
-        "trg_slf_attn_bias": [(batch_size, n_head, seq_len, seq_len), "float32"],
+        "trg_slf_attn_bias":
+        [(batch_size, n_head, seq_len, seq_len), "float32"],
         # This input is used to remove attention weights on paddings of the source
         # input in the encoder-decoder attention.
         # The actual data shape of trg_src_attn_bias is:
         # [batch_size, n_head, max_trg_len_in_batch, max_src_len_in_batch]
-        "trg_src_attn_bias": [(batch_size, n_head, seq_len, seq_len), "float32"],
+        "trg_src_attn_bias":
+        [(batch_size, n_head, seq_len, seq_len), "float32"],
         # This input is used in independent decoder program for inference.
         # The actual data shape of enc_output is:
         # [batch_size, max_src_len_in_batch, d_model]
@@ -80,6 +84,7 @@ def get_input_descs(args):
 
     return input_descs
 
+
 # Names of word embedding table which might be reused for weight sharing.
 word_emb_param_names = (
     "src_word_emb_table",

PaddleNLP/PaddleMT/transformer/inference_model.py → PaddleNLP/machine_translation/transformer/inference_model.py

@@ -24,6 +24,7 @@ import paddle.fluid as fluid
 
 from utils.input_field import InputField
 from utils.configure import PDConfig
+from utils.load import load
 
 # include task-specific libs
 import desc
@@ -31,51 +32,6 @@ import reader
 from transformer import create_net
 
 
-def init_from_pretrain_model(args, exe, program):
-
-    assert isinstance(args.init_from_pretrain_model, str)
-
-    if not os.path.exists(args.init_from_pretrain_model):
-        raise Warning("The pretrained params do not exist.")
-        return False
-
-    def existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        return os.path.exists(
-            os.path.join(args.init_from_pretrain_model, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        args.init_from_pretrain_model,
-        main_program=program,
-        predicate=existed_params)
-
-    print("finish initing model from pretrained params from %s" %
-          (args.init_from_pretrain_model))
-
-    return True
-
-
-def init_from_params(args, exe, program):
-
-    assert isinstance(args.init_from_params, str)
-
-    if not os.path.exists(args.init_from_params):
-        raise Warning("the params path does not exist.")
-        return False
-
-    fluid.io.load_params(
-        executor=exe,
-        dirname=args.init_from_params,
-        main_program=program,
-        filename="params.pdparams")
-
-    print("finish init model from params from %s" % (args.init_from_params))
-
-    return True
-
-
 def do_save_inference_model(args):
     if args.use_cuda:
         dev_count = fluid.core.get_cuda_device_count()
@@ -84,6 +40,11 @@ def do_save_inference_model(args):
         dev_count = int(os.environ.get('CPU_NUM', 1))
         place = fluid.CPUPlace()
 
+    src_vocab = reader.DataProcessor.load_dict(args.src_vocab_fpath)
+    trg_vocab = reader.DataProcessor.load_dict(args.trg_vocab_fpath)
+    args.src_vocab_size = len(src_vocab)
+    args.trg_vocab_size = len(trg_vocab)
+
     test_prog = fluid.default_main_program()
     startup_prog = fluid.default_startup_program()
 
@@ -119,13 +80,10 @@ def do_save_inference_model(args):
     exe = fluid.Executor(place)
 
     exe.run(startup_prog)
-    assert (args.init_from_params) or (args.init_from_pretrain_model)
-
-    if args.init_from_params:
-        init_from_params(args, exe, test_prog)
-
-    elif args.init_from_pretrain_model:
-        init_from_pretrain_model(args, exe, test_prog)
+    assert (
+        args.init_from_params), "must set init_from_params to load parameters"
+    load(test_prog, os.path.join(args.init_from_params, "transformer"), exe)
+    print("finish initing model from params from %s" % (args.init_from_params))
 
     # saving inference model
 

PaddleNLP/PaddleMT/transformer/main.py → PaddleNLP/machine_translation/transformer/main.py

@@ -25,7 +25,6 @@ from train import do_train
 from predict import do_predict
 from inference_model import do_save_inference_model
 
-
 if __name__ == "__main__":
     LOG_FORMAT = "[%(asctime)s %(levelname)s %(filename)s:%(lineno)d] %(message)s"
     logging.basicConfig(

PaddleNLP/PaddleMT/transformer/predict.py → PaddleNLP/machine_translation/transformer/predict.py

@@ -25,6 +25,7 @@ import paddle.fluid as fluid
 from utils.input_field import InputField
 from utils.configure import PDConfig
 from utils.check import check_gpu, check_version
+from utils.load import load
 
 # include task-specific libs
 import desc
@@ -32,51 +33,6 @@ import reader
 from transformer import create_net, position_encoding_init
 
 
-def init_from_pretrain_model(args, exe, program):
-
-    assert isinstance(args.init_from_pretrain_model, str)
-
-    if not os.path.exists(args.init_from_pretrain_model):
-        raise Warning("The pretrained params do not exist.")
-        return False
-
-    def existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        return os.path.exists(
-            os.path.join(args.init_from_pretrain_model, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        args.init_from_pretrain_model,
-        main_program=program,
-        predicate=existed_params)
-
-    print("finish initing model from pretrained params from %s" %
-          (args.init_from_pretrain_model))
-
-    return True
-
-
-def init_from_params(args, exe, program):
-
-    assert isinstance(args.init_from_params, str)
-
-    if not os.path.exists(args.init_from_params):
-        raise Warning("the params path does not exist.")
-        return False
-
-    fluid.io.load_params(
-        executor=exe,
-        dirname=args.init_from_params,
-        main_program=program,
-        filename="params.pdparams")
-
-    print("finish init model from params from %s" % (args.init_from_params))
-
-    return True
-
-
 def post_process_seq(seq, bos_idx, eos_idx, output_bos=False, output_eos=False):
     """
     Post-process the beam-search decoded sequence. Truncate from the first
@@ -160,13 +116,10 @@ def do_predict(args):
     exe = fluid.Executor(place)
 
     exe.run(startup_prog)
-    assert (args.init_from_params) or (args.init_from_pretrain_model)
-
-    if args.init_from_params:
-        init_from_params(args, exe, test_prog)
-
-    elif args.init_from_pretrain_model:
-        init_from_pretrain_model(args, exe, test_prog)
+    assert (
+        args.init_from_params), "must set init_from_params to load parameters"
+    load(test_prog, os.path.join(args.init_from_params, "transformer"), exe)
+    print("finish initing model from params from %s" % (args.init_from_params))
 
     # to avoid a longer length than training, reset the size of position encoding to max_length
     for pos_enc_param_name in desc.pos_enc_param_names:

PaddleNLP/PaddleMT/transformer/train.py → PaddleNLP/machine_translation/transformer/train.py

@@ -27,6 +27,7 @@ import utils.dist_utils as dist_utils
 from utils.input_field import InputField
 from utils.configure import PDConfig
 from utils.check import check_gpu, check_version
+from utils.load import load
 
 # include task-specific libs
 import desc
@@ -39,91 +40,6 @@ if os.environ.get('FLAGS_eager_delete_tensor_gb', None) is None:
 num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
 
 
-def init_from_pretrain_model(args, exe, program):
-
-    assert isinstance(args.init_from_pretrain_model, str)
-
-    if not os.path.exists(args.init_from_pretrain_model):
-        raise Warning("The pretrained params do not exist.")
-        return False
-
-    def existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        return os.path.exists(
-            os.path.join(args.init_from_pretrain_model, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        args.init_from_pretrain_model,
-        main_program=program,
-        predicate=existed_params)
-
-    print("finish initing model from pretrained params from %s" %
-          (args.init_from_pretrain_model))
-
-    return True
-
-
-def init_from_checkpoint(args, exe, program):
-
-    assert isinstance(args.init_from_checkpoint, str)
-
-    if not os.path.exists(args.init_from_checkpoint):
-        raise Warning("the checkpoint path does not exist.")
-        return False
-
-    fluid.io.load_persistables(
-        executor=exe,
-        dirname=args.init_from_checkpoint,
-        main_program=program,
-        filename="checkpoint.pdckpt")
-
-    print("finish initing model from checkpoint from %s" %
-          (args.init_from_checkpoint))
-
-    return True
-
-
-def save_checkpoint(args, exe, program, dirname):
-
-    assert isinstance(args.save_model_path, str)
-
-    checkpoint_dir = os.path.join(args.save_model_path, args.save_checkpoint)
-
-    if not os.path.exists(checkpoint_dir):
-        os.mkdir(checkpoint_dir)
-
-    fluid.io.save_persistables(
-        exe,
-        os.path.join(checkpoint_dir, dirname),
-        main_program=program,
-        filename="checkpoint.pdparams")
-
-    print("save checkpoint at %s" % (os.path.join(checkpoint_dir, dirname)))
-
-    return True
-
-
-def save_param(args, exe, program, dirname):
-
-    assert isinstance(args.save_model_path, str)
-
-    param_dir = os.path.join(args.save_model_path, args.save_param)
-
-    if not os.path.exists(param_dir):
-        os.mkdir(param_dir)
-
-    fluid.io.save_params(
-        exe,
-        os.path.join(param_dir, dirname),
-        main_program=program,
-        filename="params.pdparams")
-    print("save parameters at %s" % (os.path.join(param_dir, dirname)))
-
-    return True
-
-
 def do_train(args):
     if args.use_cuda:
         if num_trainers > 1:  # for multi-process gpu training
@@ -226,11 +142,17 @@ def do_train(args):
 
     ## init from some checkpoint, to resume the previous training
     if args.init_from_checkpoint:
-        init_from_checkpoint(args, exe, train_prog)
+        load(train_prog,
+             os.path.join(args.init_from_checkpoint, "transformer"), exe)
+        print("finish initing model from checkpoint from %s" %
+              (args.init_from_checkpoint))
 
     ## init from some pretrain models, to better solve the current task
     if args.init_from_pretrain_model:
-        init_from_pretrain_model(args, exe, train_prog)
+        load(train_prog,
+             os.path.join(args.init_from_pretrain_model, "transformer"), exe)
+        print("finish initing model from pretrained params from %s" %
+              (args.init_from_pretrain_model))
 
     build_strategy = fluid.compiler.BuildStrategy()
     build_strategy.enable_inplace = True
@@ -293,14 +215,11 @@ def do_train(args):
                         avg_batch_time = time.time()
 
                 if step_idx % args.save_step == 0 and step_idx != 0:
-
-                    if args.save_checkpoint:
-                        save_checkpoint(args, exe, train_prog,
-                                        "step_" + str(step_idx))
-
-                    if args.save_param:
-                        save_param(args, exe, train_prog,
-                                   "step_" + str(step_idx))
+                    if args.save_model_path:
+                        model_path = os.path.join(args.save_model_path,
+                                                  "step_" + str(step_idx),
+                                                  "transformer")
+                        fluid.save(train_prog, model_path)
 
                 batch_id += 1
                 step_idx += 1
@@ -319,11 +238,10 @@ def do_train(args):
 
         time_consumed = time.time() - pass_start_time
 
-    if args.save_checkpoint:
-        save_checkpoint(args, exe, train_prog, "step_final")
-
-    if args.save_param:
-        save_param(args, exe, train_prog, "step_final")
+    if args.save_model_path:
+        model_path = os.path.join(args.save_model_path, "step_final",
+                                  "transformer")
+        fluid.save(train_prog, model_path)
 
     if args.enable_ce:  # For CE
         print("kpis\ttrain_cost_card%d\t%f" % (dev_count, total_avg_cost))

PaddleNLP/PaddleMT/transformer/transformer.py → PaddleNLP/machine_translation/transformer/transformer.py

@@ -17,6 +17,7 @@ import numpy as np
 
 import paddle.fluid as fluid
 import paddle.fluid.layers as layers
+from paddle.fluid.layers.utils import map_structure
 
 from desc import *
 
@@ -90,7 +91,6 @@ def multi_head_attention(queries,
                          n_head=1,
                          dropout_rate=0.,
                          cache=None,
-                         gather_idx=None,
                          static_kv=False):
     """
     Multi-Head Attention. Note that attn_bias is added to the logit before
@@ -161,30 +161,28 @@ def multi_head_attention(queries,
         v = transpose_layer(x=reshaped_v, perm=[0, 2, 1, 3])
 
         if cache is not None:  # only for faster inference
+            cache_, i = cache
             if static_kv:  # For encoder-decoder attention in inference
-                cache_k, cache_v = cache["static_k"], cache["static_v"]
-                # To init the static_k and static_v in cache.
-                # Maybe we can use condition_op(if_else) to do these at the first
-                # step in while loop to replace these, however it might be less
-                # efficient.
+                cache_k, cache_v = cache_["static_k"], cache_["static_v"]
+                # To init the static_k and static_v in global block.
                 static_cache_init = wrap_layer_with_block(
                     layers.assign,
                     fluid.default_main_program().current_block().parent_idx)
-                static_cache_init(k, cache_k)
-                static_cache_init(v, cache_v)
+                static_cache_init(
+                    k,
+                    fluid.default_main_program().global_block().var(
+                        "static_k_%d" % i))
+                static_cache_init(
+                    v,
+                    fluid.default_main_program().global_block().var(
+                        "static_v_%d" % i))
+                k, v = cache_k, cache_v
             else:  # For decoder self-attention in inference
-                cache_k, cache_v = cache["k"], cache["v"]
-            # gather cell states corresponding to selected parent
-            select_k = layers.gather(cache_k, index=gather_idx)
-            select_v = layers.gather(cache_v, index=gather_idx)
-            if not static_kv:
-                # For self attention in inference, use cache and concat time steps.
-                select_k = layers.concat([select_k, k], axis=2)
-                select_v = layers.concat([select_v, v], axis=2)
-            # update cell states(caches) cached in global block
-            layers.assign(select_k, cache_k)
-            layers.assign(select_v, cache_v)
-            return q, select_k, select_v
+                # use cache and concat time steps.
+                cache_k, cache_v = cache_["k"], cache_["v"]
+                k = layers.concat([cache_k, k], axis=2)
+                v = layers.concat([cache_v, v], axis=2)
+                cache_["k"], cache_["v"] = (k, v)
         return q, k, v
 
     def __combine_heads(x):
@@ -301,12 +299,13 @@ def prepare_encoder_decoder(src_word,
         src_word,
         size=[src_vocab_size, src_emb_dim],
         padding_idx=bos_idx,  # set embedding of bos to 0
-        param_attr=fluid.ParamAttr(name=word_emb_param_name,
-                                   initializer=fluid.initializer.Normal(
-                                       0., src_emb_dim**-0.5)))
+        param_attr=fluid.ParamAttr(
+            name=word_emb_param_name,
+            initializer=fluid.initializer.Normal(0., src_emb_dim**-0.5)))
 
     src_word_emb = layers.scale(x=src_word_emb, scale=src_emb_dim**0.5)
-    src_pos_enc = fluid.embedding(src_pos,
+    src_pos_enc = fluid.embedding(
+        src_pos,
         size=[src_max_len, src_emb_dim],
         param_attr=fluid.ParamAttr(
             name=pos_enc_param_name, trainable=False))
@@ -405,8 +404,7 @@ def decoder_layer(dec_input,
                   relu_dropout,
                   preprocess_cmd,
                   postprocess_cmd,
-                  cache=None,
-                  gather_idx=None):
+                  cache=None):
     """ The layer to be stacked in decoder part.
     The structure of this module is similar to that in the encoder part except
     a multi-head attention is added to implement encoder-decoder attention.
@@ -421,8 +419,7 @@ def decoder_layer(dec_input,
         d_model,
         n_head,
         attention_dropout,
-        cache=cache,
-        gather_idx=gather_idx)
+        cache=cache)
     slf_attn_output = post_process_layer(
         dec_input,
         slf_attn_output,
@@ -440,7 +437,6 @@ def decoder_layer(dec_input,
         n_head,
         attention_dropout,
         cache=cache,
-        gather_idx=gather_idx,
         static_kv=True)
     enc_attn_output = post_process_layer(
         slf_attn_output,
@@ -476,8 +472,7 @@ def decoder(dec_input,
             relu_dropout,
             preprocess_cmd,
             postprocess_cmd,
-            caches=None,
-            gather_idx=None):
+            caches=None):
     """
     The decoder is composed of a stack of identical decoder_layer layers.
     """
@@ -497,8 +492,7 @@ def decoder(dec_input,
             relu_dropout,
             preprocess_cmd,
             postprocess_cmd,
-            cache=None if caches is None else caches[i],
-            gather_idx=gather_idx)
+            cache=None if caches is None else (caches[i], i))
         dec_input = dec_output
     dec_output = pre_process_layer(dec_output, preprocess_cmd,
                                    prepostprocess_dropout)
@@ -536,7 +530,8 @@ def transformer(model_input,
     label = model_input.lbl_word
     weights = model_input.lbl_weight
 
-    enc_output = wrap_encoder(enc_inputs,
+    enc_output = wrap_encoder(
+        enc_inputs,
         src_vocab_size,
         max_length,
         n_layer,
@@ -553,7 +548,8 @@ def transformer(model_input,
         weight_sharing,
         bos_idx=bos_idx)
 
-    predict = wrap_decoder(dec_inputs,
+    predict = wrap_decoder(
+        dec_inputs,
         trg_vocab_size,
         max_length,
         n_layer,
@@ -575,8 +571,9 @@ def transformer(model_input,
     if label_smooth_eps:
         # TODO: use fluid.input.one_hot after softmax_with_cross_entropy removing
         # the enforcement that the last dimension of label must be 1.
-        label = layers.label_smooth(label=layers.one_hot(input=label,
-                                                         depth=trg_vocab_size),
+        label = layers.label_smooth(
+            label=layers.one_hot(
+                input=label, depth=trg_vocab_size),
             epsilon=label_smooth_eps)
 
     cost = layers.softmax_with_cross_entropy(
@@ -654,7 +651,6 @@ def wrap_decoder(dec_inputs,
                  weight_sharing,
                  enc_output=None,
                  caches=None,
-                 gather_idx=None,
                  bos_idx=0):
     """
     The wrapper assembles together all needed layers for the decoder.
@@ -687,8 +683,7 @@ def wrap_decoder(dec_inputs,
         relu_dropout,
         preprocess_cmd,
         postprocess_cmd,
-        caches=caches,
-        gather_idx=gather_idx)
+        caches=caches)
     # Reshape to 2D tensor to use GEMM instead of BatchedGEMM
     dec_output = layers.reshape(
         dec_output, shape=[-1, dec_output.shape[-1]], inplace=True)
@@ -722,7 +717,8 @@ def fast_decode(model_input, src_vocab_size, trg_vocab_size, max_in_len,
     dec_inputs = (model_input.trg_word, model_input.init_score,
                   model_input.init_idx, model_input.trg_src_attn_bias)
 
-    enc_output = wrap_encoder(enc_inputs,
+    enc_output = wrap_encoder(
+        enc_inputs,
         src_vocab_size,
         max_in_len,
         n_layer,
@@ -748,8 +744,6 @@ def fast_decode(model_input, src_vocab_size, trg_vocab_size, max_in_len,
             force_cpu=True)
         step_idx = layers.fill_constant(
             shape=[1], dtype=start_tokens.dtype, value=0, force_cpu=True)
-        cond = layers.less_than(x=step_idx, y=max_len)  # default force_cpu=True
-        while_op = layers.While(cond)
         # array states will be stored for each step.
         ids = layers.array_write(
             layers.reshape(start_tokens, (-1, 1)), step_idx)
@@ -773,21 +767,31 @@ def fast_decode(model_input, src_vocab_size, trg_vocab_size, max_in_len,
                     dtype=enc_output.dtype,
                     value=0),
                 "static_k":  # for encoder-decoder attention
-                layers.create_tensor(dtype=enc_output.dtype),
+                fluid.data(
+                    shape=[None, n_head, 0, d_key],
+                    dtype=enc_output.dtype,
+                    name=("static_k_%d" % i)),
                 "static_v":  # for encoder-decoder attention
-                layers.create_tensor(dtype=enc_output.dtype)
+                fluid.data(
+                    shape=[None, n_head, 0, d_value],
+                    dtype=enc_output.dtype,
+                    name=("static_v_%d" % i)),
             } for i in range(n_layer)
         ]
 
-        with while_op.block():
-            pre_ids = layers.array_read(array=ids, i=step_idx)
-            # Since beam_search_op dosen't enforce pre_ids' shape, we can do
-            # inplace reshape here which actually change the shape of pre_ids.
-            # pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
-            pre_scores = layers.array_read(array=scores, i=step_idx)
+        def cond_func(step_idx, selected_ids, selected_scores, gather_idx,
+                      caches, trg_src_attn_bias):
+            length_cond = layers.less_than(x=step_idx, y=max_len)
+            finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
+            return layers.logical_and(x=length_cond, y=finish_cond)
+
+        def body_func(step_idx, pre_ids, pre_scores, gather_idx, caches,
+                      trg_src_attn_bias):
             # gather cell states corresponding to selected parent
+            pre_caches = map_structure(
+                lambda x: layers.gather(x, index=gather_idx), caches)
             pre_src_attn_bias = layers.gather(
-                trg_src_attn_bias, index=parent_idx)
+                trg_src_attn_bias, index=gather_idx)
             pre_pos = layers.elementwise_mul(
                 x=layers.fill_constant_batch_size_like(
                     input=pre_src_attn_bias,  # cann't use lod tensor here
@@ -796,7 +800,8 @@ def fast_decode(model_input, src_vocab_size, trg_vocab_size, max_in_len,
                     dtype=pre_ids.dtype),
                 y=step_idx,
                 axis=0)
-            logits = wrap_decoder((pre_ids, pre_pos, None, pre_src_attn_bias),
+            logits = wrap_decoder(
+                (pre_ids, pre_pos, None, pre_src_attn_bias),
                 trg_vocab_size,
                 max_in_len,
                 n_layer,
@@ -812,8 +817,7 @@ def fast_decode(model_input, src_vocab_size, trg_vocab_size, max_in_len,
                 postprocess_cmd,
                 weight_sharing,
                 enc_output=enc_output,
-                                  caches=caches,
-                                  gather_idx=parent_idx,
+                caches=pre_caches,
                 bos_idx=bos_idx)
             # intra-beam topK
             topk_scores, topk_indices = layers.topk(
@@ -832,16 +836,20 @@ def fast_decode(model_input, src_vocab_size, trg_vocab_size, max_in_len,
                 beam_size=beam_size,
                 end_id=eos_idx,
                 return_parent_idx=True)
-            layers.increment(x=step_idx, value=1.0, in_place=True)
-            # cell states(caches) have been updated in wrap_decoder,
-            # only need to update beam search states here.
+            step_idx = layers.increment(x=step_idx, value=1.0, in_place=False)
             layers.array_write(selected_ids, i=step_idx, array=ids)
             layers.array_write(selected_scores, i=step_idx, array=scores)
-            layers.assign(gather_idx, parent_idx)
-            layers.assign(pre_src_attn_bias, trg_src_attn_bias)
-            length_cond = layers.less_than(x=step_idx, y=max_len)
-            finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
-            layers.logical_and(x=length_cond, y=finish_cond, out=cond)
+            return (step_idx, selected_ids, selected_scores, gather_idx,
+                    pre_caches, pre_src_attn_bias)
+
+        _ = layers.while_loop(
+            cond=cond_func,
+            body=body_func,
+            loop_vars=[
+                step_idx, start_tokens, init_scores, parent_idx, caches,
+                trg_src_attn_bias
+            ],
+            is_test=True)
 
         finished_ids, finished_scores = layers.beam_search_decode(
             ids, scores, beam_size=beam_size, end_id=eos_idx)

PaddleNLP/PaddleMT/transformer/transformer.yaml → PaddleNLP/machine_translation/transformer/transformer.yaml

@@ -11,10 +11,11 @@ init_from_checkpoint: ""
 init_from_pretrain_model: ""
 # path of trained parameter, to make prediction
 init_from_params: "trained_params/step_100000"
-save_model_path: ""
-# the directory for saving checkpoints.
+# the directory for saving models.
+save_model_path: "saved_models"
+# deprecated, the directory for saving checkpoints.
 save_checkpoint: "trained_ckpts"
-# the directory for saving trained parameters.
+# deprecated, the directory for saving trained parameters.
 save_param: "trained_params"
 # the directory for saving inference model.
 inference_model_dir: "infer_model"

PaddleNLP/PaddleMT/transformer/utils/configure.py → PaddleNLP/machine_translation/transformer/utils/configure.py

@@ -199,9 +199,14 @@ class PDConfig(object):
                                "Whether to perform model saving for inference.")
 
         # NOTE: args for profiler
-        self.default_g.add_arg("is_profiler", int, 0, "the switch of profiler tools. (used for benchmark)")
-        self.default_g.add_arg("profiler_path", str, './', "the profiler output file path. (used for benchmark)")
-        self.default_g.add_arg("max_iter", int, 0, "the max train batch num.(used for benchmark)")
+        self.default_g.add_arg(
+            "is_profiler", int, 0,
+            "the switch of profiler tools. (used for benchmark)")
+        self.default_g.add_arg(
+            "profiler_path", str, './',
+            "the profiler output file path. (used for benchmark)")
+        self.default_g.add_arg("max_iter", int, 0,
+                               "the max train batch num.(used for benchmark)")
 
         self.parser = parser
 

PaddleNLP/machine_translation/transformer/utils/load.py

+import pickle
+import six
+import warnings
+from functools import partial
+
+import paddle.fluid as fluid
+
+
+def load(program, model_path, executor=None, var_list=None):
+    """
+    To load python2 saved models in python3.
+    """
+    try:
+        fluid.load(program, model_path, executor, var_list)
+    except UnicodeDecodeError:
+        warnings.warn(
+            "An UnicodeDecodeError is catched, which might be caused by loading "
+            "a python2 saved model. Encoding of pickle.load would be set and "
+            "load again automatically.")
+        if six.PY3:
+            load_bak = pickle.load
+            pickle.load = partial(load_bak, encoding="latin1")
+            fluid.load(program, model_path, executor, var_list)
+            pickle.load = load_bak

PaddleNLP/PaddleLARK/BERT/README.md → PaddleNLP/pretrain_langauge_models/BERT/README.md

@@ -22,6 +22,8 @@
 | :------| :------: | :------: |:------: |:------: |
 | [BERT-Large, Uncased (Whole Word Masking)](https://bert-models.bj.bcebos.com/wwm_uncased_L-24_H-1024_A-16.tar.gz)| 24 | 1024 | 16 | 340M |
 | [BERT-Large, Cased (Whole Word Masking)](https://bert-models.bj.bcebos.com/wwm_cased_L-24_H-1024_A-16.tar.gz)| 24 | 1024 | 16 | 340M |
+| [RoBERTa-Base, Chinese](https://bert-models.bj.bcebos.com/chinese_roberta_wwm_ext_L-12_H-768_A-12.tar.gz) | 12 | 768 |12 |110M |
+| [RoBERTa-Large, Chinese](https://bert-models.bj.bcebos.com/chinese_roberta_wwm_large_ext_L-24_H-1024_A-16.tar.gz) | 24 | 1024 |16 |340M |
 | [BERT-Base, Uncased](https://bert-models.bj.bcebos.com/uncased_L-12_H-768_A-12.tar.gz) | 12 | 768 |12 |110M |
 | [BERT-Large, Uncased](https://bert-models.bj.bcebos.com/uncased_L-24_H-1024_A-16.tar.gz) | 24 | 1024 |16 |340M |
 |[BERT-Base, Cased](https://bert-models.bj.bcebos.com/cased_L-12_H-768_A-12.tar.gz)|12|768|12|110M|
@@ -415,5 +417,3 @@ for (size_t i = 0; i < output.front().data.length() / sizeof(float); i += 3) {
             << static_cast<float *>(output.front().data.data())[i + 2] << std::endl;
 }
 ```
-
-

PaddleNLP/PaddleLARK/BERT/optimization.py → PaddleNLP/pretrain_langauge_models/BERT/optimization.py

@@ -158,7 +158,7 @@ def optimization(loss,
 
     else:
         if weight_decay > 0:
-            for param in train_program.global_block().all_parameters():
+            for param in train_program.all_parameters():
                 param_list[param.name] = param * 1.0
                 param_list[param.name].stop_gradient = True
 

PaddleNLP/PaddleLARK/BERT/run_classifier.py → PaddleNLP/pretrain_langauge_models/BERT/run_classifier.py

@@ -392,7 +392,7 @@ def main(args):
                 if steps % args.save_steps == 0:
                     save_path = os.path.join(args.checkpoints,
                                              "step_" + str(steps))
-                    fluid.io.save_persistables(exe, save_path, train_program)
+                    fluid.save(program=train_program, model_path=save_path)
 
                 if steps % args.validation_steps == 0:
                     print("Average throughtput: %s" % (np.average(throughput)))
@@ -409,7 +409,7 @@ def main(args):
                                  "test")
             except fluid.core.EOFException:
                 save_path = os.path.join(args.checkpoints, "step_" + str(steps))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                fluid.save(program=train_program, model_path=save_path)
                 train_data_loader.reset()
                 break
         if args.enable_ce:

PaddleNLP/PaddleLARK/BERT/run_squad.py → PaddleNLP/pretrain_langauge_models/BERT/run_squad.py

@@ -398,11 +398,11 @@ def train(args):
                 if steps % args.save_steps == 0 or steps == max_train_steps:
                     save_path = os.path.join(args.checkpoints,
                                              "step_" + str(steps))
-                    fluid.io.save_persistables(exe, save_path, train_program)
+                    fluid.save(program=train_program, model_path=save_path)
             except fluid.core.EOFException:
                 save_path = os.path.join(args.checkpoints,
                                          "step_" + str(steps) + "_final")
-                fluid.io.save_persistables(exe, save_path, train_program)
+                fluid.save(program=train_program, model_path=save_path)
                 train_data_loader.reset()
                 break
 

PaddleNLP/PaddleLARK/BERT/train.py → PaddleNLP/pretrain_langauge_models/BERT/train.py

@@ -412,7 +412,7 @@ def train(args):
 
             if steps % args.save_steps == 0:
                 save_path = os.path.join(args.checkpoints, "step_" + str(steps))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                fluid.save(program=train_program, model_path=save_path)
 
             if args.validation_set_dir and steps % args.validation_steps == 0:
                 vali_cost, vali_lm_cost, vali_acc, vali_steps, vali_speed = predict(

PaddleNLP/PaddleLARK/BERT/utils/init.py → PaddleNLP/pretrain_langauge_models/BERT/utils/init.py

@@ -25,7 +25,7 @@ import paddle.fluid as fluid
 
 def cast_fp32_to_fp16(exe, main_program):
     print("Cast parameters to float16 data format.")
-    for param in main_program.global_block().all_parameters():
+    for param in main_program.all_parameters():
         if not param.name.endswith(".master"):
             param_t = fluid.global_scope().find_var(param.name).get_tensor()
             data = np.array(param_t)
@@ -38,21 +38,9 @@ def cast_fp32_to_fp16(exe, main_program):
 
 
 def init_checkpoint(exe, init_checkpoint_path, main_program, use_fp16=False):
-    assert os.path.exists(
-        init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
+    fluid.load(
+        program=main_program, model_path=init_checkpoint_path, executor=exe)
 
-    def existed_persitables(var):
-        if not fluid.io.is_persistable(var):
-            return False
-        if os.path.exists(os.path.join(init_checkpoint_path, var.name)):
-            print("INIT {}".format(var.name))
-            return True
-
-    fluid.io.load_vars(
-        exe,
-        init_checkpoint_path,
-        main_program=main_program,
-        predicate=existed_persitables)
     print("Load model from {}".format(init_checkpoint_path))
 
     if use_fp16:
@@ -63,24 +51,8 @@ def init_pretraining_params(exe,
                             pretraining_params_path,
                             main_program,
                             use_fp16=False):
-    assert os.path.exists(pretraining_params_path
-                          ), "[%s] cann't be found." % pretraining_params_path
-
-    def existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        if os.path.exists(os.path.join(pretraining_params_path, var.name)):
-            print("INIT {}".format(var.name))
-            return True
-        else:
-            print("SKIP {}".format(var.name))
-            return False
-
-    fluid.io.load_vars(
-        exe,
-        pretraining_params_path,
-        main_program=main_program,
-        predicate=existed_params)
+    fluid.load(
+        program=main_program, model_path=pretraining_params_path, executor=exe)
     print("Load pretraining parameters from {}.".format(
         pretraining_params_path))
 

PaddleNLP/PaddleLARK/ELMo/README.md → PaddleNLP/pretrain_langauge_models/ELMo/README.md

@@ -90,5 +90,3 @@ word_embedding=fluid.layers.concat(input=[elmo_embedding, word_embedding], axis=
 
 ### 参考论文
 [Deep contextualized word representations](https://arxiv.org/abs/1802.05365)
-
-

PaddleNLP/PaddleLARK/XLNet/_ce.py → PaddleNLP/pretrain_langauge_models/XLNet/_ce.py

@@ -7,7 +7,6 @@ from kpi import CostKpi, DurationKpi, AccKpi
 
 #### NOTE kpi.py should shared in models in some way!!!!
 
-
 train_duration_sts_b_card1 = DurationKpi(
     'train_duration_sts_b_card1', 0.01, 0, actived=True)
 train_cost_sts_b_card1 = CostKpi(

PaddleNLP/sentiment_classification/README.md

@@ -29,7 +29,7 @@
 
 1. PaddlePaddle 安装
 
-   本项目依赖于 PaddlePaddle Fluid 1.6 及以上版本，请参考 [安装指南](http://www.paddlepaddle.org/#quick-start) 进行安装
+   本项目依赖于 PaddlePaddle Fluid 1.7 及以上版本，请参考 [安装指南](http://www.paddlepaddle.org/#quick-start) 进行安装
 
 2. 代码安装
 

PaddleNLP/sentiment_classification/inference_model.py

@@ -13,6 +13,7 @@ from run_classifier import create_model
 import utils
 import reader
 
+
 def do_save_inference_model(args):
     if args.use_cuda:
         dev_count = fluid.core.get_cuda_device_count()
@@ -53,6 +54,7 @@ def do_save_inference_model(args):
 
     print("save inference model at %s" % (args.inference_model_dir))
 
+
 def inference(exe, test_program, test_pyreader, fetch_list, infer_phrase):
     """
     Inference Function
@@ -61,13 +63,16 @@ def inference(exe, test_program, test_pyreader, fetch_list, infer_phrase):
     test_pyreader.start()
     while True:
         try:
-            np_props = exe.run(program=test_program, fetch_list=fetch_list, return_numpy=True)
+            np_props = exe.run(program=test_program,
+                               fetch_list=fetch_list,
+                               return_numpy=True)
             for probs in np_props[0]:
                 print("%d\t%f\t%f" % (np.argmax(probs), probs[0], probs[1]))
         except fluid.core.EOFException:
             test_pyreader.reset()
             break
 
+
 def test_inference_model(args):
     if args.use_cuda:
         dev_count = fluid.core.get_cuda_device_count()
@@ -92,7 +97,8 @@ def test_inference_model(args):
     exe = fluid.Executor(place)
     exe.run(startup_prog)
 
-    processor = reader.SentaProcessor(data_dir=args.data_dir,
+    processor = reader.SentaProcessor(
+        data_dir=args.data_dir,
         vocab_path=args.vocab_path,
         random_seed=args.random_seed,
         max_seq_len=args.max_seq_len)
@@ -107,14 +113,14 @@ def test_inference_model(args):
         params_filename="params.pdparams")
 
     infer_data_generator = processor.data_generator(
-        batch_size=args.batch_size,
+        batch_size=args.batch_size / dev_count,
         phase="infer",
         epoch=1,
         shuffle=False)
 
-    infer_pyreader.decorate_sample_list_generator(infer_data_generator)
-    inference(exe, test_prog, infer_pyreader,
-        [probs.name], "infer")
+    infer_pyreader.set_sample_list_generator(infer_data_generator)
+    inference(exe, test_prog, infer_pyreader, [probs.name], "infer")
+
 
 if __name__ == "__main__":
     args = PDConfig('senta_config.json')

PaddleNLP/sentiment_classification/inference_model_ernie.py

 # -*- coding: utf_8 -*-
 import os
 import sys
-sys.path.append("../")
-sys.path.append("../models/classification")
+sys.path.append("../shared_modules/")
+sys.path.append("../shared_modules/models/classification")
 import paddle
 import paddle.fluid as fluid
 import numpy as np
@@ -17,6 +17,7 @@ from models.representation.ernie import ErnieConfig
 from models.representation.ernie import ernie_encoder, ernie_encoder_with_paddle_hub
 from preprocess.ernie import task_reader
 
+
 def do_save_inference_model(args):
 
     ernie_config = ErnieConfig(args.ernie_config_path)
@@ -37,18 +38,17 @@ def do_save_inference_model(args):
     with fluid.program_guard(test_prog, startup_prog):
         with fluid.unique_name.guard():
             infer_pyreader, ernie_inputs, labels = ernie_pyreader(
-                args,
-                pyreader_name="infer_reader")
+                args, pyreader_name="infer_reader")
 
             if args.use_paddle_hub:
-                embeddings = ernie_encoder_with_paddle_hub(ernie_inputs, args.max_seq_len)
+                embeddings = ernie_encoder_with_paddle_hub(ernie_inputs,
+                                                           args.max_seq_len)
             else:
-                embeddings = ernie_encoder(ernie_inputs, ernie_config=ernie_config)
+                embeddings = ernie_encoder(
+                    ernie_inputs, ernie_config=ernie_config)
 
-            probs = create_model(args,
-                    embeddings,
-                    labels=labels,
-                    is_prediction=True)
+            probs = create_model(
+                args, embeddings, labels=labels, is_prediction=True)
     test_prog = test_prog.clone(for_test=True)
     exe.run(startup_prog)
 
@@ -59,11 +59,11 @@ def do_save_inference_model(args):
 
     fluid.io.save_inference_model(
         args.inference_model_dir,
-        feeded_var_names=[ernie_inputs["src_ids"].name,
-                          ernie_inputs["sent_ids"].name,
-                          ernie_inputs["pos_ids"].name,
-                          ernie_inputs["input_mask"].name,
-                          ernie_inputs["seq_lens"].name],
+        feeded_var_names=[
+            ernie_inputs["src_ids"].name, ernie_inputs["sent_ids"].name,
+            ernie_inputs["pos_ids"].name, ernie_inputs["input_mask"].name,
+            ernie_inputs["seq_lens"].name
+        ],
         target_vars=[probs],
         executor=exe,
         main_program=test_prog,
@@ -72,6 +72,7 @@ def do_save_inference_model(args):
 
     print("save inference model at %s" % (args.inference_model_dir))
 
+
 def inference(exe, test_program, test_pyreader, fetch_list, infer_phrase):
     """
     Inference Function
@@ -80,13 +81,16 @@ def inference(exe, test_program, test_pyreader, fetch_list, infer_phrase):
     test_pyreader.start()
     while True:
         try:
-            np_props = exe.run(program=test_program, fetch_list=fetch_list, return_numpy=True)
+            np_props = exe.run(program=test_program,
+                               fetch_list=fetch_list,
+                               return_numpy=True)
             for probs in np_props[0]:
                 print("%d\t%f\t%f" % (np.argmax(probs), probs[0], probs[1]))
         except fluid.core.EOFException:
             test_pyreader.reset()
             break
 
+
 def test_inference_model(args):
     ernie_config = ErnieConfig(args.ernie_config_path)
     ernie_config.print_config()
@@ -113,15 +117,11 @@ def test_inference_model(args):
     with fluid.program_guard(test_prog, startup_prog):
         with fluid.unique_name.guard():
             infer_pyreader, ernie_inputs, labels = ernie_pyreader(
-                args,
-                pyreader_name="infer_pyreader")
+                args, pyreader_name="infer_pyreader")
 
             embeddings = ernie_encoder(ernie_inputs, ernie_config=ernie_config)
             probs = create_model(
-                args,
-                embeddings,
-                labels=labels,
-                is_prediction=True)
+                args, embeddings, labels=labels, is_prediction=True)
 
     test_prog = test_prog.clone(for_test=True)
     exe.run(startup_prog)
@@ -129,7 +129,7 @@ def test_inference_model(args):
     assert (args.inference_model_dir)
     infer_data_generator = reader.data_generator(
         input_file=args.test_set,
-        batch_size=args.batch_size,
+        batch_size=args.batch_size / dev_count,
         phase="infer",
         epoch=1,
         shuffle=False)
@@ -140,9 +140,9 @@ def test_inference_model(args):
         model_filename="model.pdmodel",
         params_filename="params.pdparams")
 
-    infer_pyreader.decorate_batch_generator(infer_data_generator)
-    inference(exe, test_prog, infer_pyreader,
-        [probs.name], "infer")
+    infer_pyreader.set_batch_generator(infer_data_generator)
+    inference(exe, test_prog, infer_pyreader, [probs.name], "infer")
+
 
 if __name__ == "__main__":
     args = PDConfig()

PaddleNLP/sentiment_classification/run_classifier.py

@@ -12,8 +12,8 @@ import argparse
 import numpy as np
 import multiprocessing
 import sys
-sys.path.append("../models/classification/")
-sys.path.append("../")
+sys.path.append("../shared_modules/models/classification/")
+sys.path.append("../shared_modules/")
 
 from nets import bow_net
 from nets import lstm_net
@@ -30,24 +30,19 @@ import paddle.fluid as fluid
 import reader
 from utils import init_checkpoint
 
-def create_model(args,
-                 pyreader_name,
-                 num_labels,
-                 is_prediction=False):
 
+def create_model(args, pyreader_name, num_labels, is_prediction=False):
     """
     Create Model for sentiment classification
     """
 
-    data = fluid.layers.data(
-            name="src_ids", shape=[-1, args.max_seq_len], dtype='int64')
-    label = fluid.layers.data(
-            name="label", shape=[-1, 1], dtype="int64")
-    seq_len = fluid.layers.data(
-            name="seq_len", shape=[-1], dtype="int64")
+    data = fluid.data(
+        name="src_ids", shape=[None, args.max_seq_len], dtype='int64')
+    label = fluid.data(name="label", shape=[None, 1], dtype="int64")
+    seq_len = fluid.data(name="seq_len", shape=[None], dtype="int64")
 
-    data_reader = fluid.io.PyReader(feed_list=[data, label, seq_len], 
-        capacity=4, iterable=False)
+    data_reader = fluid.io.DataLoader.from_generator(
+        feed_list=[data, label, seq_len], capacity=4, iterable=False)
 
     if args.model_type == "bilstm_net":
         network = bilstm_net
@@ -63,18 +58,19 @@ def create_model(args,
         raise ValueError("Unknown network type!")
 
     if is_prediction:
-        probs = network(data, seq_len, None, args.vocab_size, is_prediction=is_prediction)
+        probs = network(
+            data, seq_len, None, args.vocab_size, is_prediction=is_prediction)
         print("create inference model...")
         return data_reader, probs, [data.name, seq_len.name]
 
-    ce_loss, probs = network(data, seq_len, label, args.vocab_size, is_prediction=is_prediction)
+    ce_loss, probs = network(
+        data, seq_len, label, args.vocab_size, is_prediction=is_prediction)
     loss = fluid.layers.mean(x=ce_loss)
     num_seqs = fluid.layers.create_tensor(dtype='int64')
     accuracy = fluid.layers.accuracy(input=probs, label=label, total=num_seqs)
     return data_reader, loss, accuracy, num_seqs
 
 
-
 def evaluate(exe, test_program, test_pyreader, fetch_list, eval_phase):
     """
     Evaluation Function
@@ -111,7 +107,8 @@ def inference(exe, test_program, test_pyreader, fetch_list, infer_phrase):
     time_begin = time.time()
     while True:
         try:
-            np_props = exe.run(program=test_program, fetch_list=fetch_list,
+            np_props = exe.run(program=test_program,
+                               fetch_list=fetch_list,
                                return_numpy=True)
             for probs in np_props[0]:
                 print("%d\t%f\t%f" % (np.argmax(probs), probs[0], probs[1]))
@@ -135,7 +132,8 @@ def main(args):
     exe = fluid.Executor(place)
 
     task_name = args.task_name.lower()
-    processor = reader.SentaProcessor(data_dir=args.data_dir,
+    processor = reader.SentaProcessor(
+        data_dir=args.data_dir,
         vocab_path=args.vocab_path,
         random_seed=args.random_seed,
         max_seq_len=args.max_seq_len)
@@ -151,7 +149,7 @@ def main(args):
 
     if args.do_train:
         train_data_generator = processor.data_generator(
-            batch_size=args.batch_size,
+            batch_size=args.batch_size / dev_count,
             phase='train',
             epoch=args.epoch,
             shuffle=True)
@@ -187,7 +185,7 @@ def main(args):
 
     if args.do_val:
         test_data_generator = processor.data_generator(
-            batch_size=args.batch_size,
+            batch_size=args.batch_size / dev_count,
             phase='dev',
             epoch=1,
             shuffle=False)
@@ -204,7 +202,7 @@ def main(args):
 
     if args.do_infer:
         infer_data_generator = processor.data_generator(
-            batch_size=args.batch_size,
+            batch_size=args.batch_size / dev_count,
             phase='infer',
             epoch=1,
             shuffle=False)
@@ -223,30 +221,25 @@ def main(args):
     if args.do_train:
         if args.init_checkpoint:
             init_checkpoint(
-                exe,
-                args.init_checkpoint,
-                main_program=startup_prog)
+                exe, args.init_checkpoint, main_program=startup_prog)
 
     elif args.do_val or args.do_infer:
         if not args.init_checkpoint:
             raise ValueError("args 'init_checkpoint' should be set if"
                              "only doing validation or testing!")
-        init_checkpoint(
-            exe,
-            args.init_checkpoint,
-            main_program=startup_prog)
+        init_checkpoint(exe, args.init_checkpoint, main_program=startup_prog)
 
     if args.do_train:
         train_exe = exe
-        train_reader.decorate_sample_list_generator(train_data_generator)
+        train_reader.set_sample_list_generator(train_data_generator)
     else:
         train_exe = None
     if args.do_val:
         test_exe = exe
-        test_reader.decorate_sample_list_generator(test_data_generator)
+        test_reader.set_sample_list_generator(test_data_generator)
     if args.do_infer:
         test_exe = exe
-        infer_reader.decorate_sample_list_generator(infer_data_generator)
+        infer_reader.set_sample_list_generator(infer_data_generator)
 
     if args.do_train:
         train_reader.start()
@@ -262,7 +255,9 @@ def main(args):
                 else:
                     fetch_list = []
 
-                outputs = train_exe.run(program=train_program, fetch_list=fetch_list, return_numpy=False)
+                outputs = train_exe.run(program=train_program,
+                                        fetch_list=fetch_list,
+                                        return_numpy=False)
                 #print("finished one step")
                 if steps % args.skip_steps == 0:
                     np_loss, np_acc, np_num_seqs = outputs
@@ -274,7 +269,8 @@ def main(args):
                     total_num_seqs.extend(np_num_seqs)
 
                     if args.verbose:
-                        verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size()
+                        verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
+                        )
                         print(verbose)
 
                     time_end = time.time()
@@ -289,8 +285,8 @@ def main(args):
 
                 if steps % args.save_steps == 0:
                     save_path = os.path.join(args.checkpoints,
-                                         "step_" + str(steps))
-                    fluid.io.save_persistables(exe, save_path, train_program)
+                                             "step_" + str(steps), "checkpoint")
+                    fluid.save(train_program, save_path)
 
                 if steps % args.validation_steps == 0:
                     # evaluate dev set
@@ -301,8 +297,9 @@ def main(args):
                                  "dev")
 
             except fluid.core.EOFException:
-                save_path = os.path.join(args.checkpoints, "step_" + str(steps))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                save_path = os.path.join(args.checkpoints, "step_" + str(steps),
+                                         "checkpoint")
+                fluid.save(train_program, save_path)
                 train_reader.reset()
                 break
 
@@ -315,8 +312,7 @@ def main(args):
     # final eval on test set
     if args.do_infer:
         print("Final test result:")
-        inference(exe, infer_prog, infer_reader,
-            [prop.name], "infer")
+        inference(exe, infer_prog, infer_reader, [prop.name], "infer")
 
 
 if __name__ == "__main__":

PaddleNLP/sentiment_classification/run_ernie_classifier.py

@@ -16,8 +16,8 @@ import sys
 import paddle
 import paddle.fluid as fluid
 
-sys.path.append("../models/classification/")
-sys.path.append("..")
+sys.path.append("../shared_modules/models/classification/")
+sys.path.append("../shared_modules/")
 print(sys.path)
 
 from nets import bow_net
@@ -36,19 +36,18 @@ from config import PDConfig
 
 from utils import init_checkpoint
 
+
 def ernie_pyreader(args, pyreader_name):
-    src_ids = fluid.layers.data(
-        name="src_ids", shape=[-1, args.max_seq_len, 1], dtype="int64")
-    sent_ids = fluid.layers.data(
-        name="sent_ids", shape=[-1, args.max_seq_len, 1], dtype="int64")
-    pos_ids = fluid.layers.data(
-        name="pos_ids", shape=[-1, args.max_seq_len, 1], dtype="int64")
-    input_mask = fluid.layers.data(
-        name="input_mask", shape=[-1, args.max_seq_len, 1], dtype="float32")
-    labels = fluid.layers.data(
-        name="labels", shape=[-1, 1], dtype="int64")
-    seq_lens = fluid.layers.data(
-        name="seq_lens", shape=[-1], dtype="int64")
+    src_ids = fluid.data(
+        name="src_ids", shape=[None, args.max_seq_len, 1], dtype="int64")
+    sent_ids = fluid.data(
+        name="sent_ids", shape=[None, args.max_seq_len, 1], dtype="int64")
+    pos_ids = fluid.data(
+        name="pos_ids", shape=[None, args.max_seq_len, 1], dtype="int64")
+    input_mask = fluid.data(
+        name="input_mask", shape=[None, args.max_seq_len, 1], dtype="float32")
+    labels = fluid.data(name="labels", shape=[None, 1], dtype="int64")
+    seq_lens = fluid.data(name="seq_lens", shape=[None], dtype="int64")
 
     pyreader = fluid.io.DataLoader.from_generator(
         feed_list=[src_ids, sent_ids, pos_ids, input_mask, labels, seq_lens],
@@ -61,15 +60,13 @@ def ernie_pyreader(args, pyreader_name):
         "sent_ids": sent_ids,
         "pos_ids": pos_ids,
         "input_mask": input_mask,
-        "seq_lens": seq_lens}
+        "seq_lens": seq_lens
+    }
 
     return pyreader, ernie_inputs, labels
 
-def create_model(args,
-                 embeddings,
-                 labels,
-                 is_prediction=False):
 
+def create_model(args, embeddings, labels, is_prediction=False):
     """
     Create Model for sentiment classification based on ERNIE encoder
     """
@@ -132,7 +129,8 @@ def infer(exe, infer_program, infer_pyreader, fetch_list, infer_phase):
     time_begin = time.time()
     while True:
         try:
-            batch_probs = exe.run(program=infer_program, fetch_list=fetch_list,
+            batch_probs = exe.run(program=infer_program,
+                                  fetch_list=fetch_list,
                                   return_numpy=True)
             for probs in batch_probs[0]:
                 print("%d\t%f\t%f" % (np.argmax(probs), probs[0], probs[1]))
@@ -195,21 +193,19 @@ def main(args):
             with fluid.unique_name.guard():
                 # create ernie_pyreader
                 train_pyreader, ernie_inputs, labels = ernie_pyreader(
-                    args,
-                    pyreader_name='train_pyreader')
+                    args, pyreader_name='train_pyreader')
 
                 # get ernie_embeddings
                 if args.use_paddle_hub:
-                    embeddings = ernie_encoder_with_paddle_hub(ernie_inputs, args.max_seq_len)
+                    embeddings = ernie_encoder_with_paddle_hub(ernie_inputs,
+                                                               args.max_seq_len)
                 else:
-                    embeddings = ernie_encoder(ernie_inputs, ernie_config=ernie_config)
+                    embeddings = ernie_encoder(
+                        ernie_inputs, ernie_config=ernie_config)
 
                 # user defined model based on ernie embeddings
                 loss, accuracy, num_seqs = create_model(
-                args,
-                embeddings,
-                labels=labels,
-                is_prediction=False)
+                    args, embeddings, labels=labels, is_prediction=False)
 
                 optimizer = fluid.optimizer.Adam(learning_rate=args.lr)
                 optimizer.minimize(loss)
@@ -232,21 +228,19 @@ def main(args):
             with fluid.unique_name.guard():
                 # create ernie_pyreader
                 test_pyreader, ernie_inputs, labels = ernie_pyreader(
-                    args,
-                    pyreader_name='eval_reader')
+                    args, pyreader_name='eval_reader')
 
                 # get ernie_embeddings
                 if args.use_paddle_hub:
-                    embeddings = ernie_encoder_with_paddle_hub(ernie_inputs, args.max_seq_len)
+                    embeddings = ernie_encoder_with_paddle_hub(ernie_inputs,
+                                                               args.max_seq_len)
                 else:
-                    embeddings = ernie_encoder(ernie_inputs, ernie_config=ernie_config)
+                    embeddings = ernie_encoder(
+                        ernie_inputs, ernie_config=ernie_config)
 
                 # user defined model based on ernie embeddings
                 loss, accuracy, num_seqs = create_model(
-                args,
-                embeddings,
-                labels=labels,
-                is_prediction=False)
+                    args, embeddings, labels=labels, is_prediction=False)
 
         test_prog = test_prog.clone(for_test=True)
 
@@ -261,19 +255,18 @@ def main(args):
         with fluid.program_guard(infer_prog, startup_prog):
             with fluid.unique_name.guard():
                 infer_pyreader, ernie_inputs, labels = ernie_pyreader(
-                    args,
-                    pyreader_name="infer_pyreader")
+                    args, pyreader_name="infer_pyreader")
 
                 # get ernie_embeddings
                 if args.use_paddle_hub:
-                    embeddings = ernie_encoder_with_paddle_hub(ernie_inputs, args.max_seq_len)
+                    embeddings = ernie_encoder_with_paddle_hub(ernie_inputs,
+                                                               args.max_seq_len)
                 else:
-                    embeddings = ernie_encoder(ernie_inputs, ernie_config=ernie_config)
+                    embeddings = ernie_encoder(
+                        ernie_inputs, ernie_config=ernie_config)
 
-                probs = create_model(args,
-                                    embeddings,
-                                    labels=labels,
-                                    is_prediction=True)
+                probs = create_model(
+                    args, embeddings, labels=labels, is_prediction=True)
 
         infer_prog = infer_prog.clone(for_test=True)
 
@@ -282,25 +275,17 @@ def main(args):
     if args.do_train:
         if args.init_checkpoint:
             init_checkpoint(
-                exe,
-                args.init_checkpoint,
-                main_program=train_program)
+                exe, args.init_checkpoint, main_program=train_program)
     elif args.do_val:
         if not args.init_checkpoint:
             raise ValueError("args 'init_checkpoint' should be set if"
                              "only doing validation or testing!")
-        init_checkpoint(
-            exe,
-            args.init_checkpoint,
-            main_program=test_prog)
+        init_checkpoint(exe, args.init_checkpoint, main_program=test_prog)
     elif args.do_infer:
         if not args.init_checkpoint:
             raise ValueError("args 'init_checkpoint' should be set if"
                              "only doing validation or testing!")
-        init_checkpoint(
-            exe,
-            args.init_checkpoint,
-            main_program=infer_prog)
+        init_checkpoint(exe, args.init_checkpoint, main_program=infer_prog)
 
     if args.do_train:
         train_exe = exe
@@ -327,7 +312,9 @@ def main(args):
                 else:
                     fetch_list = []
 
-                outputs = train_exe.run(program=train_program, fetch_list=fetch_list, return_numpy=False)
+                outputs = train_exe.run(program=train_program,
+                                        fetch_list=fetch_list,
+                                        return_numpy=False)
                 if steps % args.skip_steps == 0:
                     np_loss, np_acc, np_num_seqs = outputs
                     np_loss = np.array(np_loss)
@@ -338,7 +325,8 @@ def main(args):
                     total_num_seqs.extend(np_num_seqs)
 
                     if args.verbose:
-                        verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size()
+                        verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
+                        )
                         print(verbose)
 
                     time_end = time.time()
@@ -353,8 +341,8 @@ def main(args):
 
                 if steps % args.save_steps == 0:
                     save_path = os.path.join(args.checkpoints,
-                                         "step_" + str(steps))
-                    fluid.io.save_persistables(exe, save_path, train_program)
+                                             "step_" + str(steps), "checkpoint")
+                    fluid.save(train_program, save_path)
 
                 if steps % args.validation_steps == 0:
                     # evaluate dev set
@@ -364,8 +352,9 @@ def main(args):
                                  "dev")
 
             except fluid.core.EOFException:
-                save_path = os.path.join(args.checkpoints, "step_" + str(steps))
-                fluid.io.save_persistables(exe, save_path, train_program)
+                save_path = os.path.join(args.checkpoints, "step_" + str(steps),
+                                         "checkpoint")
+                fluid.save(train_program, save_path)
                 train_pyreader.reset()
                 break
 
@@ -378,8 +367,8 @@ def main(args):
     # final eval on test set
     if args.do_infer:
         print("Final test result:")
-        infer(exe, infer_prog, infer_pyreader,
-            [probs.name], "infer")
+        infer(exe, infer_prog, infer_pyreader, [probs.name], "infer")
+
 
 if __name__ == "__main__":
     args = PDConfig()

PaddleNLP/sentiment_classification/utils.py

@@ -31,6 +31,7 @@ class ArgumentGroup(object):
     """
     Argument Class
     """
+
     def __init__(self, parser, title, des):
         self._group = parser.add_argument_group(title=title, description=des)
 
@@ -63,20 +64,11 @@ def init_checkpoint(exe, init_checkpoint_path, main_program):
     """
     assert os.path.exists(
         init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
-
-    def existed_persitables(var):
-        """
-        If existed presitabels
-        """
-        if not fluid.io.is_persistable(var):
-            return False
-        return os.path.exists(os.path.join(init_checkpoint_path, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        init_checkpoint_path,
-        main_program=main_program,
-        predicate=existed_persitables)
+    try:
+        checkpoint_path = os.path.join(init_checkpoint_path, "checkpoint")
+        fluid.load(main_program, checkpoint_path, exe)
+    except:
+        fluid.load(main_program, init_checkpoint_path, exe)
     print("Load model from {}".format(init_checkpoint_path))
 
 
@@ -96,8 +88,10 @@ def data_reader(file_path, word_dict, num_examples, phrase, epoch, max_seq_len):
                 sys.stderr.write("[NOTICE] Error Format Line!")
                 continue
             label = int(cols[1])
-            wids = [word_dict[x] if x in word_dict else unk_id
-                    for x in cols[0].split(" ")]
+            wids = [
+                word_dict[x] if x in word_dict else unk_id
+                for x in cols[0].split(" ")
+            ]
             seq_len = len(wids)
             if seq_len < max_seq_len:
                 for i in range(max_seq_len - seq_len):
@@ -119,8 +113,10 @@ def data_reader(file_path, word_dict, num_examples, phrase, epoch, max_seq_len):
         for epoch_index in range(epoch):
             for doc, label, seq_len in all_data:
                 yield doc, label, seq_len
+
     return reader
 
+
 def load_vocab(file_path):
     """
     load the given vocabulary
@@ -144,15 +140,6 @@ def init_pretraining_params(exe,
     assert os.path.exists(pretraining_params_path
                           ), "[%s] cann't be found." % pretraining_params_path
 
-    def _existed_params(var):
-        if not isinstance(var, fluid.framework.Parameter):
-            return False
-        return os.path.exists(os.path.join(pretraining_params_path, var.name))
-
-    fluid.io.load_vars(
-        exe,
-        pretraining_params_path,
-        main_program=main_program,
-        predicate=_existed_params)
+    fluid.load(main_program, pretraining_params_path, exe)
     print("Load pretraining parameters from {}.".format(
         pretraining_params_path))

PaddleNLP/PaddleTextGEN/seq2seq/README.md → PaddleNLP/seq2seq/seq2seq/README.md

-运行本目录下的范例模型需要安装PaddlePaddle Fluid 1.6版。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](https://www.paddlepaddle.org.cn/#quick-start)中的说明更新 PaddlePaddle 安装版本。
+运行本目录下的范例模型需要安装PaddlePaddle Fluid 1.7版本。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](https://www.paddlepaddle.org.cn/#quick-start)中的说明更新 PaddlePaddle 安装版本。
 
 # Sequence to Sequence (Seq2Seq)
 

PaddleNLP/PaddleTextGEN/seq2seq/infer.py → PaddleNLP/seq2seq/seq2seq/infer.py

@@ -93,7 +93,7 @@ def infer():
     # clone from default main program and use it as the validation program
     main_program = fluid.default_main_program()
     main_program = main_program.clone(for_test=True)
-    print([param.name for param in main_program.blocks[0].all_parameters()])
+    print([param.name for param in main_program.all_parameters()])
 
     place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
     exe = Executor(place)
@@ -127,7 +127,8 @@ def infer():
 
     dir_name = args.reload_model
     print("dir name", dir_name)
-    fluid.io.load_params(exe, dir_name)
+    dir_name = os.path.join(dir_name, "checkpoint")
+    fluid.load(main_program, dir_name, exe)
 
     train_data_iter = reader.get_data_iter(infer_data, 1, mode='eval')
 

PaddleNLP/PaddleTextGEN/seq2seq/train.py → PaddleNLP/seq2seq/seq2seq/train.py

@@ -229,10 +229,10 @@ def main():
                 % (epoch_id, epoch_time, sum(batch_times) / len(batch_times)))
 
             if not args.profile:
-                dir_name = os.path.join(args.model_path,
-                                        "epoch_" + str(epoch_id))
-                print("begin to save", dir_name)
-                fluid.io.save_params(exe, dir_name, main_program=train_program)
+                save_path = os.path.join(args.model_path,
+                                         "epoch_" + str(epoch_id), "checkpoint")
+                print("begin to save", save_path)
+                fluid.save(train_program, save_path)
                 print("save finished")
                 dev_ppl = eval(valid_data)
                 print("dev ppl", dev_ppl)

PaddleNLP/PaddleTextGEN/variational_seq2seq/infer.py → PaddleNLP/seq2seq/variational_seq2seq/infer.py

@@ -88,7 +88,8 @@ def infer():
 
     dir_name = args.reload_model
     print("dir name", dir_name)
-    fluid.io.load_params(exe, dir_name)
+    dir_name = os.path.join(dir_name, "checkpoint")
+    fluid.load(main_program, dir_name, exe)
     vocab, tar_id2vocab = get_vocab(args.dataset_prefix)
     infer_output = np.ones((batch_size, 1), dtype='int64') * BOS_ID
 

PaddleNLP/PaddleTextGEN/variational_seq2seq/train.py → PaddleNLP/seq2seq/variational_seq2seq/train.py

@@ -255,10 +255,11 @@ def main():
                 best_nll = test_nll
                 best_ppl = test_ppl
                 best_epoch_id = epoch_id
-                dir_name = os.path.join(args.model_path,
-                                        "epoch_" + str(best_epoch_id))
-                print("save model {}".format(dir_name))
-                fluid.io.save_params(exe, dir_name, main_program)
+                save_path = os.path.join(args.model_path,
+                                         "epoch_" + str(best_epoch_id),
+                                         "checkpoint")
+                print("save model {}".format(save_path))
+                fluid.save(main_program, save_path)
             else:
                 steps_not_improved += 1
                 if steps_not_improved == decay_ts:

PaddleNLP/models/classification/nets.py → PaddleNLP/shared_modules/models/classification/nets.py

@@ -4,6 +4,7 @@ This module provide nets for text classification
 
 import paddle.fluid as fluid
 
+
 def bow_net(data,
             seq_len,
             label,

PaddleNLP/models/matching/cnn.py → PaddleNLP/shared_modules/models/matching/cnn.py

@@ -43,8 +43,8 @@ class CNN(object):
         left_emb = emb_layer.ops(left)
         right_emb = emb_layer.ops(right)
         # Presentation context
-        cnn_layer = layers.SequenceConvPoolLayer(
-            self.filter_size, self.num_filters, "conv")
+        cnn_layer = layers.SequenceConvPoolLayer(self.filter_size,
+                                                 self.num_filters, "conv")
         left_cnn = cnn_layer.ops(left_emb)
         right_cnn = cnn_layer.ops(right_emb)
         # matching layer

PaddleNLP/models/model_check.py → PaddleNLP/shared_modules/models/model_check.py

@@ -33,6 +33,7 @@ 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

PaddleNLP/models/representation/ernie.py → PaddleNLP/shared_modules/models/representation/ernie.py

@@ -30,10 +30,14 @@ from models.transformer_encoder import encoder, pre_process_layer
 
 def ernie_pyreader(args, pyreader_name):
     """define standard ernie pyreader"""
-    src_ids = fluid.data(name='1', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    sent_ids = fluid.data(name='2', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    pos_ids = fluid.data(name='3', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    input_mask = fluid.data(name='4', shape=[-1, args.max_seq_len, 1], dtype='float32')
+    src_ids = fluid.data(
+        name='1', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    sent_ids = fluid.data(
+        name='2', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    pos_ids = fluid.data(
+        name='3', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    input_mask = fluid.data(
+        name='4', shape=[-1, args.max_seq_len, 1], dtype='float32')
     labels = fluid.data(name='5', shape=[-1, 1], dtype='int64')
     seq_lens = fluid.data(name='6', shape=[-1], dtype='int64')
 

PaddleNLP/preprocess/ernie/task_reader.py → PaddleNLP/shared_modules/preprocess/ernie/task_reader.py

@@ -29,6 +29,7 @@ from preprocess.ernie import tokenization
 from preprocess.padding import pad_batch_data
 import io
 
+
 def csv_reader(fd, delimiter='\t'):
     def gen():
         for i in fd:
@@ -37,8 +38,10 @@ def csv_reader(fd, delimiter='\t'):
                 yield slots,
             else:
                 yield slots
+
     return gen()
 
+
 class BaseReader(object):
     """BaseReader for classify and sequence labeling task"""
 

PaddleNLP/preprocess/ernie/tokenization.py → PaddleNLP/shared_modules/preprocess/ernie/tokenization.py

@@ -23,6 +23,7 @@ import unicodedata
 import six
 import io
 
+
 def convert_to_unicode(text):
     """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
     if six.PY3:

PaddleNLP/similarity_net/run_classifier.py

@@ -30,7 +30,7 @@ if sys.getdefaultencoding() != defaultencoding:
     reload(sys)
     sys.setdefaultencoding(defaultencoding)
 
-sys.path.append("..")
+sys.path.append("../shared_modules/")
 
 import paddle
 import paddle.fluid as fluid
@@ -47,14 +47,14 @@ from models.model_check import check_version
 from models.model_check import check_cuda
 
 
-def create_model(args, pyreader_name, is_inference = False, is_pointwise = False):
+def create_model(args, pyreader_name, is_inference=False, is_pointwise=False):
     """
     Create Model for simnet
     """
     if is_inference:
         inf_pyreader = fluid.layers.py_reader(
             capacity=16,
-        shapes=([-1,1], [-1,1]),
+            shapes=([-1], [-1]),
             dtypes=('int64', 'int64'),
             lod_levels=(1, 1),
             name=pyreader_name,
@@ -67,7 +67,7 @@ def create_model(args, pyreader_name, is_inference = False, is_pointwise = False
         if is_pointwise:
             pointwise_pyreader = fluid.layers.py_reader(
                 capacity=16,
-            shapes=([-1,1], [-1,1], [-1,1]),
+                shapes=([-1], [-1], [-1]),
                 dtypes=('int64', 'int64', 'int64'),
                 lod_levels=(1, 1, 0),
                 name=pyreader_name,
@@ -79,15 +79,17 @@ def create_model(args, pyreader_name, is_inference = False, is_pointwise = False
         else:
             pairwise_pyreader = fluid.layers.py_reader(
                 capacity=16,
-            shapes=([-1,1], [-1,1], [-1,1]),
+                shapes=([-1], [-1], [-1]),
                 dtypes=('int64', 'int64', 'int64'),
                 lod_levels=(1, 1, 1),
                 name=pyreader_name,
                 use_double_buffer=False)
 
-            left, pos_right, neg_right = fluid.layers.read_file(pairwise_pyreader)
+            left, pos_right, neg_right = fluid.layers.read_file(
+                pairwise_pyreader)
             return pairwise_pyreader, left, pos_right, neg_right
 
+
 def train(conf_dict, args):
     """
     train processic
@@ -97,16 +99,16 @@ def train(conf_dict, args):
     # get vocab size
     conf_dict['dict_size'] = len(vocab)
     # Load network structure dynamically
-    net = utils.import_class("../models/matching",
+    net = utils.import_class("../shared_modules/models/matching",
                              conf_dict["net"]["module_name"],
                              conf_dict["net"]["class_name"])(conf_dict)
     # Load loss function dynamically
-    loss = utils.import_class("../models/matching/losses",
+    loss = utils.import_class("../shared_modules/models/matching/losses",
                               conf_dict["loss"]["module_name"],
                               conf_dict["loss"]["class_name"])(conf_dict)
     # Load Optimization method
     optimizer = utils.import_class(
-        "../models/matching/optimizers", "paddle_optimizers",
+        "../shared_modules/models/matching/optimizers", "paddle_optimizers",
         conf_dict["optimizer"]["class_name"])(conf_dict)
     # load auc method
     metric = fluid.metrics.Auc(name="auc")
@@ -131,8 +133,7 @@ def train(conf_dict, args):
         with fluid.program_guard(train_program, startup_prog):
             with fluid.unique_name.guard():
                 train_pyreader, left, pos_right, neg_right = create_model(
-                    args, 
-                    pyreader_name='train_reader')
+                    args, pyreader_name='train_reader')
                 left_feat, pos_score = net.predict(left, pos_right)
                 pred = pos_score
                 _, neg_score = net.predict(left, neg_right)
@@ -141,12 +142,14 @@ def train(conf_dict, args):
                 optimizer.ops(avg_cost)
 
         # Get Reader
-        get_train_examples = simnet_process.get_reader("train",epoch=args.epoch)
+        get_train_examples = simnet_process.get_reader(
+            "train", epoch=args.epoch)
         if args.do_valid:
             test_prog = fluid.Program()
             with fluid.program_guard(test_prog, startup_prog):
                 with fluid.unique_name.guard():
-                    test_pyreader, left, pos_right= create_model(args, pyreader_name = 'test_reader',is_inference=True)
+                    test_pyreader, left, pos_right = create_model(
+                        args, pyreader_name='test_reader', is_inference=True)
                     left_feat, pos_score = net.predict(left, pos_right)
                     pred = pos_score
             test_prog = test_prog.clone(for_test=True)
@@ -156,40 +159,41 @@ def train(conf_dict, args):
         with fluid.program_guard(train_program, startup_prog):
             with fluid.unique_name.guard():
                 train_pyreader, left, right, label = create_model(
-                    args, 
-                    pyreader_name='train_reader',
-                    is_pointwise=True)
+                    args, pyreader_name='train_reader', is_pointwise=True)
                 left_feat, pred = net.predict(left, right)
                 avg_cost = loss.compute(pred, label)
                 avg_cost.persistable = True
                 optimizer.ops(avg_cost)
 
         # Get Feeder and Reader
-        get_train_examples = simnet_process.get_reader("train",epoch=args.epoch)
+        get_train_examples = simnet_process.get_reader(
+            "train", epoch=args.epoch)
         if args.do_valid:
             test_prog = fluid.Program()
             with fluid.program_guard(test_prog, startup_prog):
                 with fluid.unique_name.guard():
-                    test_pyreader, left, right= create_model(args, pyreader_name = 'test_reader',is_inference=True)
+                    test_pyreader, left, right = create_model(
+                        args, pyreader_name='test_reader', is_inference=True)
                     left_feat, pred = net.predict(left, right)
             test_prog = test_prog.clone(for_test=True)
 
     if args.init_checkpoint is not "":
-        utils.init_checkpoint(exe, args.init_checkpoint, 
-                              startup_prog)
+        utils.init_checkpoint(exe, args.init_checkpoint, startup_prog)
 
-    def valid_and_test(test_program, test_pyreader, get_valid_examples, process, mode, exe, fetch_list):
+    def valid_and_test(test_program, test_pyreader, get_valid_examples, process,
+                       mode, exe, fetch_list):
         """
         return auc and acc
         """
         # Get Batch Data
-        batch_data = fluid.io.batch(get_valid_examples, args.batch_size, drop_last=False)
+        batch_data = fluid.io.batch(
+            get_valid_examples, args.batch_size, drop_last=False)
         test_pyreader.decorate_paddle_reader(batch_data)
         test_pyreader.start()
         pred_list = []
         while True:
             try:
-                _pred = exe.run(program=test_program,fetch_list=[pred.name])
+                _pred = exe.run(program=test_program, fetch_list=[pred.name])
                 pred_list += list(_pred)
             except fluid.core.EOFException:
                 test_pyreader.reset()
@@ -222,7 +226,8 @@ def train(conf_dict, args):
     #for epoch_id in range(args.epoch):
     # used for continuous evaluation
     if args.enable_ce:
-        train_batch_data = fluid.io.batch(get_train_examples, args.batch_size, drop_last=False)
+        train_batch_data = fluid.io.batch(
+            get_train_examples, args.batch_size, drop_last=False)
     else:
         train_batch_data = fluid.io.batch(
             fluid.io.shuffle(
@@ -238,19 +243,23 @@ def train(conf_dict, args):
         try:
             global_step += 1
             fetch_list = [avg_cost.name]
-            avg_loss = train_exe.run(program=train_program, fetch_list = fetch_list)
+            avg_loss = train_exe.run(program=train_program,
+                                     fetch_list=fetch_list)
             losses.append(np.mean(avg_loss[0]))
             if args.do_valid and global_step % args.validation_steps == 0:
                 get_valid_examples = simnet_process.get_reader("valid")
-                valid_result = valid_and_test(test_prog,test_pyreader,get_valid_examples,simnet_process,"valid",exe,[pred.name])
+                valid_result = valid_and_test(
+                    test_prog, test_pyreader, get_valid_examples,
+                    simnet_process, "valid", exe, [pred.name])
                 if args.compute_accuracy:
                     valid_auc, valid_acc = valid_result
                     logging.info(
-                        "global_steps: %d, valid_auc: %f, valid_acc: %f, valid_loss: %f" %
-                        (global_step, valid_auc, valid_acc, np.mean(losses)))
+                        "global_steps: %d, valid_auc: %f, valid_acc: %f, valid_loss: %f"
+                        % (global_step, valid_auc, valid_acc, np.mean(losses)))
                 else:
                     valid_auc = valid_result
-                    logging.info("global_steps: %d, valid_auc: %f, valid_loss: %f" %
+                    logging.info(
+                        "global_steps: %d, valid_auc: %f, valid_loss: %f" %
                         (global_step, valid_auc, np.mean(losses)))
             if global_step % args.save_steps == 0:
                 model_save_dir = os.path.join(args.output_dir,
@@ -269,8 +278,7 @@ def train(conf_dict, args):
                     ]
                     target_vars = [left_feat, pred]
                 fluid.io.save_inference_model(model_path, feed_var_names,
-                                            target_vars, exe,
-                                            test_prog)
+                                              target_vars, exe, test_prog)
                 logging.info("saving infer model in %s" % model_path)
 
         except fluid.core.EOFException:
@@ -282,8 +290,7 @@ def train(conf_dict, args):
     ce_info.append([np.mean(losses), end_time - start_time])
     #final save
     logging.info("the final step is %s" % global_step)
-    model_save_dir = os.path.join(args.output_dir,
-                                conf_dict["model_path"])
+    model_save_dir = os.path.join(args.output_dir, conf_dict["model_path"])
     model_path = os.path.join(model_save_dir, str(global_step))
     if not os.path.exists(model_save_dir):
         os.makedirs(model_save_dir)
@@ -296,8 +303,7 @@ def train(conf_dict, args):
             right.name,
         ]
         target_vars = [left_feat, pred]
-    fluid.io.save_inference_model(model_path, feed_var_names,
-                                target_vars, exe,
+    fluid.io.save_inference_model(model_path, feed_var_names, target_vars, exe,
                                   test_prog)
     logging.info("saving infer model in %s" % model_path)
     # used for continuous evaluation
@@ -322,7 +328,9 @@ def train(conf_dict, args):
         else:
             # Get Feeder and Reader
             get_test_examples = simnet_process.get_reader("test")
-        test_result = valid_and_test(test_prog,test_pyreader,get_test_examples,simnet_process,"test",exe,[pred.name])
+        test_result = valid_and_test(test_prog, test_pyreader,
+                                     get_test_examples, simnet_process, "test",
+                                     exe, [pred.name])
         if args.compute_accuracy:
             test_auc, test_acc = test_result
             logging.info("AUC of test is %f, Accuracy of test is %f" %
@@ -348,12 +356,13 @@ def test(conf_dict, args):
     startup_prog = fluid.Program()
 
     get_test_examples = simnet_process.get_reader("test")
-    batch_data = fluid.io.batch(get_test_examples, args.batch_size, drop_last=False)
+    batch_data = fluid.io.batch(
+        get_test_examples, args.batch_size, drop_last=False)
     test_prog = fluid.Program()
 
     conf_dict['dict_size'] = len(vocab)
 
-    net = utils.import_class("../models/matching",
+    net = utils.import_class("../shared_modules/models/matching",
                              conf_dict["net"]["module_name"],
                              conf_dict["net"]["class_name"])(conf_dict)
 
@@ -364,9 +373,7 @@ def test(conf_dict, args):
             with fluid.program_guard(test_prog, startup_prog):
                 with fluid.unique_name.guard():
                     test_pyreader, left, pos_right = create_model(
-                        args,
-                        pyreader_name = 'test_reader',
-                        is_inference=True)
+                        args, pyreader_name='test_reader', is_inference=True)
                     left_feat, pos_score = net.predict(left, pos_right)
                     pred = pos_score
             test_prog = test_prog.clone(for_test=True)
@@ -375,18 +382,13 @@ def test(conf_dict, args):
             with fluid.program_guard(test_prog, startup_prog):
                 with fluid.unique_name.guard():
                     test_pyreader, left, right = create_model(
-                        args,
-                        pyreader_name = 'test_reader',
-                        is_inference=True)
+                        args, pyreader_name='test_reader', is_inference=True)
                     left_feat, pred = net.predict(left, right)
             test_prog = test_prog.clone(for_test=True)
 
         exe.run(startup_prog)
 
-        utils.init_checkpoint(
-            exe,
-            args.init_checkpoint,
-            main_program=test_prog)
+        utils.init_checkpoint(exe, args.init_checkpoint, main_program=test_prog)
 
         test_exe = exe
         test_pyreader.decorate_paddle_reader(batch_data)
@@ -398,15 +400,18 @@ def test(conf_dict, args):
         output = []
         while True:
             try:
-                output = test_exe.run(program=test_prog,fetch_list=fetch_list)
+                output = test_exe.run(program=test_prog, fetch_list=fetch_list)
                 if args.task_mode == "pairwise":
-                    pred_list += list(map(lambda item: float(item[0]), output[0]))
+                    pred_list += list(
+                        map(lambda item: float(item[0]), output[0]))
                     predictions_file.write(u"\n".join(
-                        map(lambda item: str((item[0] + 1) / 2), output[0])) + "\n")
+                        map(lambda item: str((item[0] + 1) / 2), output[0])) +
+                                           "\n")
                 else:
                     pred_list += map(lambda item: item, output[0])
                     predictions_file.write(u"\n".join(
-                        map(lambda item: str(np.argmax(item)), output[0])) + "\n")
+                        map(lambda item: str(np.argmax(item)), output[0])) +
+                                           "\n")
             except fluid.core.EOFException:
                 test_pyreader.reset()
                 break
@@ -450,36 +455,36 @@ def infer(conf_dict, args):
     startup_prog = fluid.Program()
 
     get_infer_examples = simnet_process.get_infer_reader
-    batch_data = fluid.io.batch(get_infer_examples, args.batch_size, drop_last=False)
+    batch_data = fluid.io.batch(
+        get_infer_examples, args.batch_size, drop_last=False)
 
     test_prog = fluid.Program()
 
     conf_dict['dict_size'] = len(vocab)
 
-    net = utils.import_class("../models/matching",
+    net = utils.import_class("../shared_modules/models/matching",
                              conf_dict["net"]["module_name"],
                              conf_dict["net"]["class_name"])(conf_dict)
 
     if args.task_mode == "pairwise":
         with fluid.program_guard(test_prog, startup_prog):
             with fluid.unique_name.guard():
-                infer_pyreader, left, pos_right = create_model(args, pyreader_name = 'infer_reader', is_inference = True)
+                infer_pyreader, left, pos_right = create_model(
+                    args, pyreader_name='infer_reader', is_inference=True)
                 left_feat, pos_score = net.predict(left, pos_right)
                 pred = pos_score
         test_prog = test_prog.clone(for_test=True)
     else:
         with fluid.program_guard(test_prog, startup_prog):
             with fluid.unique_name.guard():
-                infer_pyreader, left, right = create_model(args, pyreader_name = 'infer_reader', is_inference = True)
+                infer_pyreader, left, right = create_model(
+                    args, pyreader_name='infer_reader', is_inference=True)
                 left_feat, pred = net.predict(left, right)
         test_prog = test_prog.clone(for_test=True)
 
     exe.run(startup_prog)
 
-    utils.init_checkpoint(
-        exe,
-        args.init_checkpoint,
-        main_program=test_prog)
+    utils.init_checkpoint(exe, args.init_checkpoint, main_program=test_prog)
 
     test_exe = exe
     infer_pyreader.decorate_sample_list_generator(batch_data)
@@ -491,7 +496,7 @@ def infer(conf_dict, args):
     infer_pyreader.start()
     while True:
         try:
-                output = test_exe.run(program=test_prog,fetch_list=fetch_list)
+            output = test_exe.run(program=test_prog, fetch_list=fetch_list)
             if args.task_mode == "pairwise":
                 preds_list += list(
                     map(lambda item: str((item[0] + 1) / 2), output[0]))
@@ -514,6 +519,7 @@ def get_cards():
         num = len(cards.split(","))
     return num
 
+
 if __name__ == "__main__":
 
     args = ArgConfig()

README.md

@@ -149,7 +149,7 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
 [**PaddleNLP**](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP) 是基于 PaddlePaddle 深度学习框架开发的自然语言处理 (NLP) 工具，算法，模型和数据的开源项目。百度在 NLP 领域十几年的深厚积淀为 PaddleNLP 提供了强大的核心动力。使用 PaddleNLP，您可以得到：
 
 - **丰富而全面的 NLP 任务支持：**
-  - PaddleNLP 为您提供了多粒度，多场景的应用支持。涵盖了从[分词](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[词性标注](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[命名实体识别](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)等 NLP 基础技术，到[文本分类](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/sentiment_classification)，[文本相似度计算](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/similarity_net)，[语义表示](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK)，[文本生成](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleTextGEN)等 NLP 核心技术。同时，PaddleNLP 还提供了针对常见 NLP 大型应用系统（如[阅读理解](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMRC)，[对话系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleDialogue)，[机器翻译系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMT)等）的特定核心技术和工具组件，模型和预训练参数等，让您在 NLP 领域畅通无阻。
+  - PaddleNLP 为您提供了多粒度，多场景的应用支持。涵盖了从[分词](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[词性标注](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)，[命名实体识别](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/lexical_analysis)等 NLP 基础技术，到[文本分类](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/sentiment_classification)，[文本相似度计算](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/similarity_net)，[语义表示](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models)，[文本生成](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/seq2seq)等 NLP 核心技术。同时，PaddleNLP 还提供了针对常见 NLP 大型应用系统（如[阅读理解](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_reading_comprehension)，[对话系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/dialogue_system)，[机器翻译系统](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_translation)等）的特定核心技术和工具组件，模型和预训练参数等，让您在 NLP 领域畅通无阻。
 - **稳定可靠的 NLP 模型和强大的预训练参数：**
   - PaddleNLP集成了百度内部广泛使用的 NLP 工具模型，为您提供了稳定可靠的 NLP 算法解决方案。基于百亿级数据的预训练参数和丰富的预训练模型，助您轻松提高模型效果，为您的 NLP 业务注入强大动力。
 - **持续改进和技术支持，零基础搭建 NLP 应用：**
@@ -167,14 +167,14 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
 
 #### 语义表示
 
-[PaddleLARK](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK) (Paddle LAngauge Representation ToolKit) 是传统语言模型的进一步发展，通过在大规模语料上训练得到的通用的语义表示模型，可以助益其他自然语言处理任务，是通用预训练 + 特定任务精调范式的体现。PaddleLARK 集成了 ELMO，BERT，ERNIE 1.0，ERNIE 2.0，XLNet 等热门中英文预训练模型。
+[pretrain_langauge_models](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models) (Paddle LAngauge Representation ToolKit) 是传统语言模型的进一步发展，通过在大规模语料上训练得到的通用的语义表示模型，可以助益其他自然语言处理任务，是通用预训练 + 特定任务精调范式的体现。pretrain_langauge_models 集成了 ELMO，BERT，ERNIE 1.0，ERNIE 2.0，XLNet 等热门中英文预训练模型。
 
 | 模型                                                         | 简介                                                         |
 | ------------------------------------------------------------ | ------------------------------------------------------------ |
 | [ERNIE](https://github.com/PaddlePaddle/ERNIE)(Enhanced Representation from kNowledge IntEgration) | 百度自研的语义表示模型，通过建模海量数据中的词、实体及实体关系，学习真实世界的语义知识。相较于 BERT 学习原始语言信号，ERNIE 直接对先验语义知识单元进行建模，增强了模型语义表示能力。 |
-| [BERT](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK/BERT)(Bidirectional Encoder Representation from Transformers) | 一个迁移能力很强的通用语义表示模型， 以 Transformer 为网络基本组件，以双向 Masked Language Model和 Next Sentence Prediction 为训练目标，通过预训练得到通用语义表示，再结合简单的输出层，应用到下游的 NLP 任务，在多个任务上取得了 SOTA 的结果。 |
-| [XLNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK/XLNet)(XLNet: Generalized Autoregressive Pretraining for Language Understanding) | 重要的语义表示模型之一，引入 Transformer-XL 为骨架，以 Permutation Language Modeling 为优化目标，在若干下游任务上优于 BERT 的性能。 |
-| [ELMo](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleLARK/ELMo)(Embeddings from Language Models) | 重要的通用语义表示模型之一，以双向 LSTM 为网路基本组件，以 Language Model 为训练目标，通过预训练得到通用的语义表示，将通用的语义表示作为 Feature 迁移到下游 NLP 任务中，会显著提升下游任务的模型性能。 |
+| [BERT](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models/BERT)(Bidirectional Encoder Representation from Transformers) | 一个迁移能力很强的通用语义表示模型， 以 Transformer 为网络基本组件，以双向 Masked Language Model和 Next Sentence Prediction 为训练目标，通过预训练得到通用语义表示，再结合简单的输出层，应用到下游的 NLP 任务，在多个任务上取得了 SOTA 的结果。 |
+| [XLNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models/XLNet)(XLNet: Generalized Autoregressive Pretraining for Language Understanding) | 重要的语义表示模型之一，引入 Transformer-XL 为骨架，以 Permutation Language Modeling 为优化目标，在若干下游任务上优于 BERT 的性能。 |
+| [ELMo](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/pretrain_langauge_models/ELMo)(Embeddings from Language Models) | 重要的通用语义表示模型之一，以双向 LSTM 为网路基本组件，以 Language Model 为训练目标，通过预训练得到通用的语义表示，将通用的语义表示作为 Feature 迁移到下游 NLP 任务中，会显著提升下游任务的模型性能。 |
 
 #### 文本相似度计算
 
@@ -182,7 +182,7 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
 
 #### 文本生成
 
-[PaddleTextGEN](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleTextGEN) (Paddle Text Generation) ,一个基于 PaddlePaddle 的文本生成框架，提供了一些列经典文本生成模型案例，如 vanilla seq2seq，seq2seq with attention，variational seq2seq 模型等。
+[seq2seq](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/seq2seq) (Paddle Text Generation) ,一个基于 PaddlePaddle 的文本生成框架，提供了一些列经典文本生成模型案例，如 vanilla seq2seq，seq2seq with attention，variational seq2seq 模型等。
 
 ### NLP 系统应用
 
@@ -195,7 +195,7 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
 
 #### 阅读理解
 
-[PaddleMRC](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMRC) (Paddle Machine Reading Comprehension)，集合了百度在阅读理解领域相关的模型，工具，开源数据集等一系列工作。
+[machine_reading_comprehension](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_reading_comprehension) (Paddle Machine Reading Comprehension)，集合了百度在阅读理解领域相关的模型，工具，开源数据集等一系列工作。
 
 | 模型                                                         | 简介                                                         |
 | ------------------------------------------------------------ | ------------------------------------------------------------ |
@@ -205,16 +205,16 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
 
 #### 机器翻译
 
-[PaddleMT](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleMT) ，全称为Paddle Machine Translation，基于Transformer的经典机器翻译模型，基于论文 [Attention Is All You Need](https://arxiv.org/abs/1706.03762)。
+[machine_translation](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_translation) ，全称为Paddle Machine Translation，基于Transformer的经典机器翻译模型，基于论文 [Attention Is All You Need](https://arxiv.org/abs/1706.03762)。
 
 #### 对话系统
 
-[PaddleDialogue](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleDialogue) 包含对话系统方向的模型、数据集和工具。
+[dialogue_system](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/dialogue_system) 包含对话系统方向的模型、数据集和工具。
 
 | 模型                                                         | 简介                                                         |
 | ------------------------------------------------------------ | ------------------------------------------------------------ |
-| [DGU](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleDialogue/dialogue_general_understanding) (Dialogue General Understanding，通用对话理解模型) | 覆盖了包括**检索式聊天系统**中 context-response matching 任务和**任务完成型对话系统**中**意图识别**，**槽位解析**，**状态追踪**等常见对话系统任务，在 6 项国际公开数据集中都获得了最佳效果。 |
-| [ADEM](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/PaddleDialogue/auto_dialogue_evaluation) (Auto Dialogue Evaluation Model) | 评估开放领域对话系统的回复质量，能够帮助企业或个人快速评估对话系统的回复质量，减少人工评估成本。 |
+| [DGU](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/dialogue_system/dialogue_general_understanding) (Dialogue General Understanding，通用对话理解模型) | 覆盖了包括**检索式聊天系统**中 context-response matching 任务和**任务完成型对话系统**中**意图识别**，**槽位解析**，**状态追踪**等常见对话系统任务，在 6 项国际公开数据集中都获得了最佳效果。 |
+| [ADEM](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/dialogue_system/auto_dialogue_evaluation) (Auto Dialogue Evaluation Model) | 评估开放领域对话系统的回复质量，能够帮助企业或个人快速评估对话系统的回复质量，减少人工评估成本。 |
 | [Proactive Conversation](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/Research/ACL2019-DuConv) | 包含百度开源的知识驱动的开放领域对话数据集 [DuConv](https://ai.baidu.com/broad/subordinate?dataset=duconv)，以及基线模型。对应论文 [Proactive Human-Machine Conversation with Explicit Conversation Goals](https://arxiv.org/abs/1906.05572) 发表于 ACL2019。 |
 | [DAM](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/Research/ACL2018-DAM)（Deep Attention Matching Network，深度注意力机制模型） | 开放领域多轮对话匹配模型，对应论文 [Multi-Turn Response Selection for Chatbots with Deep Attention Matching Network](https://aclweb.org/anthology/P18-1103/) 发表于 ACL2018。 |