Merge branch 'readme' of https://github.com/123malin/PaddleRec into readme

.travis.yml

@@ -16,15 +16,20 @@ before_install:
   # For pylint dockstring checker
   - sudo apt-get update 
   - sudo apt-get install -y python-pip libpython-dev
+  - sudo apt-get remove python-urllib3
+  - sudo apt-get purge python-urllib3
+  - sudo rm /usr/lib/python2.7/dist-packages/chardet-*
   - sudo pip install -U pip
+  - sudo pip install --upgrade setuptools
   - sudo pip install six --upgrade --ignore-installed six
-  - sudo pip install pillow
   - sudo pip install PyYAML
   - sudo pip install pylint pytest astroid isort pre-commit
   - sudo pip install kiwisolver
-  - sudo pip install paddlepaddle==1.7.2 --ignore-installed urllib3
-  - sudo pip uninstall -y rarfile
+  - sudo pip install scikit-build
+  - sudo pip install Pillow==5.3.0
+  - sudo pip install opencv-python==3.4.3.18
   - sudo pip install rarfile==3.0
+  - sudo pip install paddlepaddle==1.7.2
   - sudo python setup.py install
   - |
     function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }

README.md

@@ -124,7 +124,10 @@
 
 ```bash
 # 使用CPU进行单机训练
-python -m paddlerec.run -m paddlerec.models.rank.dnn  
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
+python -m paddlerec.run -m models/rank/dnn/config.yaml  
 ```
 
 
@@ -144,6 +147,7 @@ python -m paddlerec.run -m paddlerec.models.rank.dnn
 * [启动分布式训练](doc/distributed_train.md)
 * [启动预测](doc/predict.md)
 * [快速部署](doc/serving.md)
+* [预训练模型](doc/pre_train_model.md)
 
 
 ### 进阶教程

README_EN.md

@@ -119,7 +119,10 @@ We take the `dnn` algorithm as an example to get start of `PaddleRec`, and we ta
 
 ```bash
 # Training with cpu
-python -m paddlerec.run -m paddlerec.models.rank.dnn  
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
+python -m paddlerec.run -m models/rank/dnn/config.yaml  
 ```
 
 

core/engine/cluster/cloud/k8s_config.ini.template

@@ -19,6 +19,7 @@ afs_local_mount_point="/root/paddlejob/workspace/env_run/afs/"
 # 新k8s afs挂载帮助文档: http://wiki.baidu.com/pages/viewpage.action?pageId=906443193
 
 PADDLE_PADDLEREC_ROLE=WORKER
+PADDLEREC_CLUSTER_TYPE=K8S
 use_python3=<$ USE_PYTHON3 $>
 CPU_NUM=<$ CPU_NUM $>
 GLOG_v=0

core/engine/cluster/cloud/mpi_config.ini.template

@@ -17,6 +17,7 @@ output_path=<$ OUTPUT_PATH $>
 thirdparty_path=<$ THIRDPARTY_PATH $>
 
 PADDLE_PADDLEREC_ROLE=WORKER
+PADDLEREC_CLUSTER_TYPE=MPI
 use_python3=<$ USE_PYTHON3 $>
 CPU_NUM=<$ CPU_NUM $>
 GLOG_v=0

core/factory.py

@@ -22,6 +22,19 @@ trainers = {}
 
 
 def trainer_registry():
+    trainers["SingleTrainer"] = os.path.join(trainer_abs, "single_trainer.py")
+    trainers["ClusterTrainer"] = os.path.join(trainer_abs,
+                                              "cluster_trainer.py")
+    trainers["CtrCodingTrainer"] = os.path.join(trainer_abs,
+                                                "ctr_coding_trainer.py")
+    trainers["CtrModulTrainer"] = os.path.join(trainer_abs,
+                                               "ctr_modul_trainer.py")
+    trainers["TDMSingleTrainer"] = os.path.join(trainer_abs,
+                                                "tdm_single_trainer.py")
+    trainers["TDMClusterTrainer"] = os.path.join(trainer_abs,
+                                                 "tdm_cluster_trainer.py")
+    trainers["OnlineLearningTrainer"] = os.path.join(
+        trainer_abs, "online_learning_trainer.py")
     # Definition of procedure execution process
     trainers["CtrCodingTrainer"] = os.path.join(trainer_abs,
                                                 "ctr_coding_trainer.py")

core/trainer.py

@@ -107,6 +107,7 @@ class Trainer(object):
             self.device = Device.GPU
             gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
             self._place = fluid.CUDAPlace(gpu_id)
+            print("PaddleRec run on device GPU: {}".format(gpu_id))
             self._exe = fluid.Executor(self._place)
         elif device == "CPU":
             self.device = Device.CPU
@@ -146,6 +147,7 @@ class Trainer(object):
         elif engine.upper() == "CLUSTER":
             self.engine = EngineMode.CLUSTER
             self.is_fleet = True
+            self.which_cluster_type()
         else:
             raise ValueError("Not Support Engine {}".format(engine))
         self._context["is_fleet"] = self.is_fleet
@@ -165,6 +167,14 @@ class Trainer(object):
         self._context["is_pslib"] = (fleet_mode.upper() == "PSLIB")
         self._context["fleet_mode"] = fleet_mode
 
+    def which_cluster_type(self):
+        cluster_type = os.getenv("PADDLEREC_CLUSTER_TYPE", "MPI")
+        print("PADDLEREC_CLUSTER_TYPE: {}".format(cluster_type))
+        if cluster_type and cluster_type.upper() == "K8S":
+            self._context["cluster_type"] = "K8S"
+        else:
+            self._context["cluster_type"] = "MPI"
+
     def which_executor_mode(self):
         executor_mode = envs.get_runtime_environ("train.trainer.executor_mode")
         if executor_mode.upper() not in ["TRAIN", "INFER"]:

core/trainers/finetuning_trainer.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+"""
+General Trainer, applicable to many situations: Single/Cluster/Local_Cluster + PS/COLLECTIVE
+"""
+from __future__ import print_function
+
+import os
+
+from paddlerec.core.utils import envs
+from paddlerec.core.trainer import Trainer, EngineMode, FleetMode
+
+
+class FineTuningTrainer(Trainer):
+    """
+    Trainer for various situations
+    """
+
+    def __init__(self, config=None):
+        Trainer.__init__(self, config)
+        self.processor_register()
+        self.abs_dir = os.path.dirname(os.path.abspath(__file__))
+        self.runner_env_name = "runner." + self._context["runner_name"]
+
+    def processor_register(self):
+        print("processor_register begin")
+        self.regist_context_processor('uninit', self.instance)
+        self.regist_context_processor('network_pass', self.network)
+        self.regist_context_processor('startup_pass', self.startup)
+        self.regist_context_processor('train_pass', self.runner)
+        self.regist_context_processor('terminal_pass', self.terminal)
+
+    def instance(self, context):
+        instance_class_path = envs.get_global_env(
+            self.runner_env_name + ".instance_class_path", default_value=None)
+        if instance_class_path:
+            instance_class = envs.lazy_instance_by_fliename(
+                instance_class_path, "Instance")(context)
+        else:
+            if self.engine == EngineMode.SINGLE:
+                instance_class_name = "SingleInstance"
+            else:
+                raise ValueError(
+                    "FineTuningTrainer can only support SingleTraining.")
+
+            instance_path = os.path.join(self.abs_dir, "framework",
+                                         "instance.py")
+
+            instance_class = envs.lazy_instance_by_fliename(
+                instance_path, instance_class_name)(context)
+
+        instance_class.instance(context)
+
+    def network(self, context):
+        network_class_path = envs.get_global_env(
+            self.runner_env_name + ".network_class_path", default_value=None)
+        if network_class_path:
+            network_class = envs.lazy_instance_by_fliename(network_class_path,
+                                                           "Network")(context)
+        else:
+            if self.engine == EngineMode.SINGLE:
+                network_class_name = "FineTuningNetwork"
+            else:
+                raise ValueError(
+                    "FineTuningTrainer can only support SingleTraining.")
+
+            network_path = os.path.join(self.abs_dir, "framework",
+                                        "network.py")
+            network_class = envs.lazy_instance_by_fliename(
+                network_path, network_class_name)(context)
+
+        network_class.build_network(context)
+
+    def startup(self, context):
+        startup_class_path = envs.get_global_env(
+            self.runner_env_name + ".startup_class_path", default_value=None)
+        if startup_class_path:
+            startup_class = envs.lazy_instance_by_fliename(startup_class_path,
+                                                           "Startup")(context)
+        else:
+            if self.engine == EngineMode.SINGLE and not context["is_infer"]:
+                startup_class_name = "FineTuningStartup"
+            else:
+                raise ValueError(
+                    "FineTuningTrainer can only support SingleTraining.")
+
+            startup_path = os.path.join(self.abs_dir, "framework",
+                                        "startup.py")
+
+            startup_class = envs.lazy_instance_by_fliename(
+                startup_path, startup_class_name)(context)
+        startup_class.startup(context)
+
+    def runner(self, context):
+        runner_class_path = envs.get_global_env(
+            self.runner_env_name + ".runner_class_path", default_value=None)
+        if runner_class_path:
+            runner_class = envs.lazy_instance_by_fliename(runner_class_path,
+                                                          "Runner")(context)
+        else:
+            if self.engine == EngineMode.SINGLE and not context["is_infer"]:
+                runner_class_name = "SingleRunner"
+            else:
+                raise ValueError(
+                    "FineTuningTrainer can only support SingleTraining.")
+
+            runner_path = os.path.join(self.abs_dir, "framework", "runner.py")
+            runner_class = envs.lazy_instance_by_fliename(
+                runner_path, runner_class_name)(context)
+        runner_class.run(context)
+
+    def terminal(self, context):
+        terminal_class_path = envs.get_global_env(
+            self.runner_env_name + ".terminal_class_path", default_value=None)
+        if terminal_class_path:
+            terminal_class = envs.lazy_instance_by_fliename(
+                terminal_class_path, "Terminal")(context)
+            terminal_class.terminal(context)
+        else:
+            terminal_class_name = "TerminalBase"
+            if self.engine != EngineMode.SINGLE and self.fleet_mode != FleetMode.COLLECTIVE:
+                terminal_class_name = "PSTerminal"
+
+            terminal_path = os.path.join(self.abs_dir, "framework",
+                                         "terminal.py")
+            terminal_class = envs.lazy_instance_by_fliename(
+                terminal_path, terminal_class_name)(context)
+        terminal_class.terminal(context)
+        context['is_exit'] = True

core/trainers/framework/dataset.py

@@ -21,7 +21,7 @@ from paddlerec.core.utils import envs
 from paddlerec.core.utils import dataloader_instance
 from paddlerec.core.reader import SlotReader
 from paddlerec.core.trainer import EngineMode
-from paddlerec.core.utils.util import split_files
+from paddlerec.core.utils.util import split_files, check_filelist
 
 __all__ = ["DatasetBase", "DataLoader", "QueueDataset"]
 
@@ -121,14 +121,30 @@ class QueueDataset(DatasetBase):
         dataset.set_pipe_command(pipe_cmd)
         train_data_path = envs.get_global_env(name + "data_path")
 
-        file_list = [
-            os.path.join(train_data_path, x)
-            for x in os.listdir(train_data_path)
-        ]
+        hidden_file_list, file_list = check_filelist(
+            hidden_file_list=[],
+            data_file_list=[],
+            train_data_path=train_data_path)
+        if (hidden_file_list is not None):
+            print(
+                "Warning:please make sure there are no hidden files in the dataset folder and check these hidden files:{}".
+                format(hidden_file_list))
+
+        file_list.sort()
+        need_split_files = False
         if context["engine"] == EngineMode.LOCAL_CLUSTER:
+            # for local cluster: split files for multi process
+            need_split_files = True
+        elif context["engine"] == EngineMode.CLUSTER and context[
+                "cluster_type"] == "K8S":
+            # for k8s mount afs, split files for every node
+            need_split_files = True
+
+        if need_split_files:
             file_list = split_files(file_list, context["fleet"].worker_index(),
                                     context["fleet"].worker_num())
         print("File_list: {}".format(file_list))
+
         dataset.set_filelist(file_list)
         for model_dict in context["phases"]:
             if model_dict["dataset_name"] == dataset_name:

core/trainers/framework/network.py

@@ -23,7 +23,7 @@ from paddlerec.core.trainers.framework.dataset import DataLoader, QueueDataset
 
 __all__ = [
     "NetworkBase", "SingleNetwork", "PSNetwork", "PslibNetwork",
-    "CollectiveNetwork"
+    "CollectiveNetwork", "FineTuningNetwork"
 ]
 
 
@@ -109,6 +109,88 @@ class SingleNetwork(NetworkBase):
         context["status"] = "startup_pass"
 
 
+class FineTuningNetwork(NetworkBase):
+    """R
+    """
+
+    def __init__(self, context):
+        print("Running FineTuningNetwork.")
+
+    def build_network(self, context):
+        context["model"] = {}
+        for model_dict in context["phases"]:
+            context["model"][model_dict["name"]] = {}
+            train_program = fluid.Program()
+            startup_program = fluid.Program()
+            scope = fluid.Scope()
+            dataset_name = model_dict["dataset_name"]
+
+            with fluid.program_guard(train_program, startup_program):
+                with fluid.unique_name.guard():
+                    with fluid.scope_guard(scope):
+                        model_path = envs.os_path_adapter(
+                            envs.workspace_adapter(model_dict["model"]))
+                        model = envs.lazy_instance_by_fliename(
+                            model_path, "Model")(context["env"])
+
+                        model._data_var = model.input_data(
+                            dataset_name=model_dict["dataset_name"])
+
+                        if envs.get_global_env("dataset." + dataset_name +
+                                               ".type") == "DataLoader":
+                            model._init_dataloader(
+                                is_infer=context["is_infer"])
+                            data_loader = DataLoader(context)
+                            data_loader.get_dataloader(context, dataset_name,
+                                                       model._data_loader)
+
+                        model.net(model._data_var, context["is_infer"])
+
+                        finetuning_varnames = envs.get_global_env(
+                            "runner." + context["runner_name"] +
+                            ".finetuning_aspect_varnames",
+                            default_value=[])
+
+                        if len(finetuning_varnames) == 0:
+                            raise ValueError(
+                                "nothing need to be fine tuning, you may use other traning mode"
+                            )
+
+                        if len(finetuning_varnames) != 1:
+                            raise ValueError(
+                                "fine tuning mode can only accept one varname now"
+                            )
+
+                        varname = finetuning_varnames[0]
+                        finetuning_vars = train_program.global_block().vars[
+                            varname]
+                        finetuning_vars.stop_gradient = True
+                        optimizer = model.optimizer()
+                        optimizer.minimize(model._cost)
+
+            context["model"][model_dict["name"]][
+                "main_program"] = train_program
+            context["model"][model_dict["name"]][
+                "startup_program"] = startup_program
+            context["model"][model_dict["name"]]["scope"] = scope
+            context["model"][model_dict["name"]]["model"] = model
+            context["model"][model_dict["name"]][
+                "default_main_program"] = train_program.clone()
+            context["model"][model_dict["name"]]["compiled_program"] = None
+
+        context["dataset"] = {}
+        for dataset in context["env"]["dataset"]:
+            type = envs.get_global_env("dataset." + dataset["name"] + ".type")
+
+            if type == "QueueDataset":
+                dataset_class = QueueDataset(context)
+                context["dataset"][dataset[
+                    "name"]] = dataset_class.create_dataset(dataset["name"],
+                                                            context)
+
+        context["status"] = "startup_pass"
+
+
 class PSNetwork(NetworkBase):
     def __init__(self, context):
         print("Running PSNetwork.")

core/trainers/framework/runner.py

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import os
 import time
+import warnings
 import numpy as np
 import paddle.fluid as fluid
 
@@ -284,6 +285,7 @@ class RunnerBase(object):
             return (epoch_id + 1) % epoch_interval == 0
 
         def save_inference_model():
+            # get global env
             name = "runner." + context["runner_name"] + "."
             save_interval = int(
                 envs.get_global_env(name + "save_inference_interval", -1))
@@ -296,18 +298,44 @@ class RunnerBase(object):
             if feed_varnames is None or fetch_varnames is None or feed_varnames == "" or fetch_varnames == "" or \
                     len(feed_varnames) == 0 or len(fetch_varnames) == 0:
                 return
-            fetch_vars = [
-                fluid.default_main_program().global_block().vars[varname]
-                for varname in fetch_varnames
-            ]
+
+            # check feed var exist
+            for var_name in feed_varnames:
+                if var_name not in fluid.default_main_program().global_block(
+                ).vars:
+                    raise ValueError(
+                        "Feed variable: {} not in default_main_program, global block has follow vars: {}".
+                        format(var_name,
+                               fluid.default_main_program().global_block()
+                               .vars.keys()))
+
+            # check fetch var exist
+            fetch_vars = []
+            for var_name in fetch_varnames:
+                if var_name not in fluid.default_main_program().global_block(
+                ).vars:
+                    raise ValueError(
+                        "Fetch variable: {} not in default_main_program, global block has follow vars: {}".
+                        format(var_name,
+                               fluid.default_main_program().global_block()
+                               .vars.keys()))
+                else:
+                    fetch_vars.append(fluid.default_main_program()
+                                      .global_block().vars[var_name])
+
             dirname = envs.get_global_env(name + "save_inference_path", None)
 
             assert dirname is not None
             dirname = os.path.join(dirname, str(epoch_id))
 
             if is_fleet:
-                context["fleet"].save_inference_model(
-                    context["exe"], dirname, feed_varnames, fetch_vars)
+                warnings.warn(
+                    "Save inference model in cluster training is not recommended! Using save checkpoint instead.",
+                    category=UserWarning,
+                    stacklevel=2)
+                if context["fleet"].worker_index() == 0:
+                    context["fleet"].save_inference_model(
+                        context["exe"], dirname, feed_varnames, fetch_vars)
             else:
                 fluid.io.save_inference_model(dirname, feed_varnames,
                                               fetch_vars, context["exe"])
@@ -323,7 +351,8 @@ class RunnerBase(object):
                 return
             dirname = os.path.join(dirname, str(epoch_id))
             if is_fleet:
-                context["fleet"].save_persistables(context["exe"], dirname)
+                if context["fleet"].worker_index() == 0:
+                    context["fleet"].save_persistables(context["exe"], dirname)
             else:
                 fluid.io.save_persistables(context["exe"], dirname)
 

core/trainers/framework/startup.py

@@ -17,9 +17,13 @@ from __future__ import print_function
 import warnings
 
 import paddle.fluid as fluid
+import paddle.fluid.core as core
 from paddlerec.core.utils import envs
 
-__all__ = ["StartupBase", "SingleStartup", "PSStartup", "CollectiveStartup"]
+__all__ = [
+    "StartupBase", "SingleStartup", "PSStartup", "CollectiveStartup",
+    "FineTuningStartup"
+]
 
 
 class StartupBase(object):
@@ -65,6 +69,122 @@ class SingleStartup(StartupBase):
         context["status"] = "train_pass"
 
 
+class FineTuningStartup(StartupBase):
+    """R
+    """
+
+    def __init__(self, context):
+        self.op_name_scope = "op_namescope"
+        self.clip_op_name_scope = "@CLIP"
+        self.self.op_role_var_attr_name = core.op_proto_and_checker_maker.kOpRoleVarAttrName(
+        )
+
+        print("Running SingleStartup.")
+
+    def _is_opt_role_op(self, op):
+        # NOTE: depend on oprole to find out whether this op is for
+        # optimize
+        op_maker = core.op_proto_and_checker_maker
+        optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
+        if op_maker.kOpRoleAttrName() in op.attr_names and \
+                int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role):
+            return True
+        return False
+
+    def _get_params_grads(self, program):
+        """
+        Get optimizer operators, parameters and gradients from origin_program
+        Returns:
+            opt_ops (list): optimize operators.
+            params_grads (dict): parameter->gradient.
+        """
+        block = program.global_block()
+        params_grads = []
+        # tmp set to dedup
+        optimize_params = set()
+        origin_var_dict = program.global_block().vars
+        for op in block.ops:
+            if self._is_opt_role_op(op):
+                # Todo(chengmo): Whether clip related op belongs to Optimize guard should be discussed
+                # delete clip op from opt_ops when run in Parameter Server mode
+                if self.op_name_scope in op.all_attrs(
+                ) and self.clip_op_name_scope in op.attr(self.op_name_scope):
+                    op._set_attr(
+                        "op_role",
+                        int(core.op_proto_and_checker_maker.OpRole.Backward))
+                    continue
+
+                if op.attr(self.op_role_var_attr_name):
+                    param_name = op.attr(self.op_role_var_attr_name)[0]
+                    grad_name = op.attr(self.op_role_var_attr_name)[1]
+                    if not param_name in optimize_params:
+                        optimize_params.add(param_name)
+                        params_grads.append([
+                            origin_var_dict[param_name],
+                            origin_var_dict[grad_name]
+                        ])
+        return params_grads
+
+    @staticmethod
+    def is_persistable(var):
+        """
+        Check whether the given variable is persistable.
+
+        Args:
+            var(Variable): The variable to be checked.
+
+        Returns:
+            bool: True if the given `var` is persistable
+            False if not.
+
+        Examples:
+            .. code-block:: python
+
+                import paddle.fluid as fluid
+                param = fluid.default_main_program().global_block().var('fc.b')
+                res = fluid.io.is_persistable(param)
+        """
+        if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \
+                var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \
+                var.desc.type() == core.VarDesc.VarType.READER:
+            return False
+        return var.persistable
+
+    def load(self, context, is_fleet=False, main_program=None):
+        dirname = envs.get_global_env(
+            "runner." + context["runner_name"] + ".init_model_path", None)
+        if dirname is None or dirname == "":
+            return
+        print("going to load ", dirname)
+
+        params_grads = self._get_params_grads(main_program)
+        update_params = [p for p, _ in params_grads]
+        need_load_vars = []
+        parameters = list(
+            filter(FineTuningStartup.is_persistable, main_program.list_vars()))
+
+        for param in parameters:
+            if param not in update_params:
+                need_load_vars.append(param)
+
+        fluid.io.load_vars(context["exe"], dirname, main_program,
+                           need_load_vars)
+        print("load from {} success".format(dirname))
+
+    def startup(self, context):
+        for model_dict in context["phases"]:
+            with fluid.scope_guard(context["model"][model_dict["name"]][
+                    "scope"]):
+                train_prog = context["model"][model_dict["name"]][
+                    "main_program"]
+                startup_prog = context["model"][model_dict["name"]][
+                    "startup_program"]
+                with fluid.program_guard(train_prog, startup_prog):
+                    context["exe"].run(startup_prog)
+                    self.load(context, main_program=train_prog)
+        context["status"] = "train_pass"
+
+
 class PSStartup(StartupBase):
     def __init__(self, context):
         print("Running PSStartup.")

core/utils/dataloader_instance.py

@@ -19,7 +19,7 @@ from paddlerec.core.utils.envs import get_global_env
 from paddlerec.core.utils.envs import get_runtime_environ
 from paddlerec.core.reader import SlotReader
 from paddlerec.core.trainer import EngineMode
-from paddlerec.core.utils.util import split_files
+from paddlerec.core.utils.util import split_files, check_filelist
 
 
 def dataloader_by_name(readerclass,
@@ -38,11 +38,27 @@ def dataloader_by_name(readerclass,
         assert package_base is not None
         data_path = os.path.join(package_base, data_path.split("::")[1])
 
-    files = [str(data_path) + "/%s" % x for x in os.listdir(data_path)]
+    hidden_file_list, files = check_filelist(
+        hidden_file_list=[], data_file_list=[], train_data_path=data_path)
+    if (hidden_file_list is not None):
+        print(
+            "Warning:please make sure there are no hidden files in the dataset folder and check these hidden files:{}".
+            format(hidden_file_list))
+
+    files.sort()
+
+    need_split_files = False
     if context["engine"] == EngineMode.LOCAL_CLUSTER:
+        # for local cluster: split files for multi process
+        need_split_files = True
+    elif context["engine"] == EngineMode.CLUSTER and context[
+            "cluster_type"] == "K8S":
+        # for k8s mount mode, split files for every node
+        need_split_files = True
+    print("need_split_files: {}".format(need_split_files))
+    if need_split_files:
         files = split_files(files, context["fleet"].worker_index(),
                             context["fleet"].worker_num())
-    print("file_list : {}".format(files))
 
     reader = reader_class(yaml_file)
     reader.init()
@@ -84,11 +100,27 @@ def slotdataloader_by_name(readerclass, dataset_name, yaml_file, context):
         assert package_base is not None
         data_path = os.path.join(package_base, data_path.split("::")[1])
 
-    files = [str(data_path) + "/%s" % x for x in os.listdir(data_path)]
+    hidden_file_list, files = check_filelist(
+        hidden_file_list=[], data_file_list=[], train_data_path=data_path)
+    if (hidden_file_list is not None):
+        print(
+            "Warning:please make sure there are no hidden files in the dataset folder and check these hidden files:{}".
+            format(hidden_file_list))
+
+    files.sort()
+
+    need_split_files = False
     if context["engine"] == EngineMode.LOCAL_CLUSTER:
+        # for local cluster: split files for multi process
+        need_split_files = True
+    elif context["engine"] == EngineMode.CLUSTER and context[
+            "cluster_type"] == "K8S":
+        # for k8s mount mode, split files for every node
+        need_split_files = True
+
+    if need_split_files:
         files = split_files(files, context["fleet"].worker_index(),
                             context["fleet"].worker_num())
-        print("file_list: {}".format(files))
 
     sparse = get_global_env(name + "sparse_slots", "#")
     if sparse == "":
@@ -138,11 +170,27 @@ def slotdataloader(readerclass, train, yaml_file, context):
         assert package_base is not None
         data_path = os.path.join(package_base, data_path.split("::")[1])
 
-    files = [str(data_path) + "/%s" % x for x in os.listdir(data_path)]
+    hidden_file_list, files = check_filelist(
+        hidden_file_list=[], data_file_list=[], train_data_path=data_path)
+    if (hidden_file_list is not None):
+        print(
+            "Warning:please make sure there are no hidden files in the dataset folder and check these hidden files:{}".
+            format(hidden_file_list))
+
+    files.sort()
+
+    need_split_files = False
     if context["engine"] == EngineMode.LOCAL_CLUSTER:
+        # for local cluster: split files for multi process
+        need_split_files = True
+    elif context["engine"] == EngineMode.CLUSTER and context[
+            "cluster_type"] == "K8S":
+        # for k8s mount mode, split files for every node
+        need_split_files = True
+
+    if need_split_files:
         files = split_files(files, context["fleet"].worker_index(),
                             context["fleet"].worker_num())
-        print("file_list: {}".format(files))
 
     sparse = get_global_env("sparse_slots", "#", namespace)
     if sparse == "":

core/utils/util.py

@@ -201,6 +201,28 @@ def split_files(files, trainer_id, trainers):
     return trainer_files[trainer_id]
 
 
+def check_filelist(hidden_file_list, data_file_list, train_data_path):
+    for root, dirs, files in os.walk(train_data_path):
+        if (files == None and dirs == None):
+            return None, None
+        else:
+            # use files and dirs
+            for file_name in files:
+                file_path = os.path.join(train_data_path, file_name)
+                if file_name[0] == '.':
+                    hidden_file_list.append(file_path)
+                else:
+                    data_file_list.append(file_path)
+            for dirs_name in dirs:
+                dirs_path = os.path.join(train_data_path, dirs_name)
+                if dirs_name[0] == '.':
+                    hidden_file_list.append(dirs_path)
+                else:
+                    #train_data_path = os.path.join(train_data_path, dirs_name)
+                    check_filelist(hidden_file_list, data_file_list, dirs_path)
+            return hidden_file_list, data_file_list
+
+
 class CostPrinter(object):
     """
     For count cost time && print cost log

doc/custom_reader.md

-# PaddleRec 自定义数据集及Reader
-
-用户自定义数据集及配置异步Reader，需要关注以下几个步骤：
-
-* [数据集整理](#数据集整理)
-* [在模型组网中加入输入占位符](#在模型组网中加入输入占位符)
-* [Reader实现](#Reader的实现)
-* [在yaml文件中配置Reader](#在yaml文件中配置reader)
-
-我们以CTR-DNN模型为例，给出了从数据整理，变量定义，Reader写法，调试的完整历程。
-
-* [数据及Reader示例-DNN](#数据及Reader示例-DNN)
-
-
-## 数据集整理
-
-PaddleRec支持模型自定义数据集。
-
-关于数据的tips：
-1. 数据量：
-
-    PaddleRec面向大规模数据设计，可以轻松支持亿级的数据读取，工业级的数据读写api：`dataset`在搜索、推荐、信息流等业务得到了充分打磨。
-2. 文件类型:
-
-    支持任意直接可读的文本数据，`dataset`同时支持`.gz`格式的文本压缩数据，无需额外代码，可直接读取。数据样本应以`\n`为标志，按行组织。
-
-3. 文件存放位置：
-
-    文件通常存放在训练节点本地，但同时，`dataset`支持使用`hadoop`远程读取数据，数据无需下载到本地，为dataset配置hadoop相关账户及地址即可。
-4. 数据类型
-
-    Reader处理的是以行为单位的`string`数据，喂入网络的数据需要转为`int`,`float`的数值数据，不支持`string`喂入网络，不建议明文保存及处理训练数据。
-5. Tips
-
-    Dataset模式下，训练线程与数据读取线程的关系强相关，为了多线程充分利用，`强烈建议将文件合理的拆为多个小文件`，尤其是在分布式训练场景下，可以均衡各个节点的数据量，同时加快数据的下载速度。
-
-## 在模型组网中加入输入占位符
-
-Reader读取文件后，产出的数据喂入网络，需要有占位符进行接收。占位符在Paddle中使用`fluid.data`或`fluid.layers.data`进行定义。`data`的定义可以参考[fluid.data](https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/fluid_cn/data_cn.html#data)以及[fluid.layers.data](https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/layers_cn/data_cn.html#data)。
-
-假如您希望输入三个数据，分别是维度32的数据A，维度变长的稀疏数据B，以及一个一维的标签数据C，并希望梯度可以经过该变量向前传递，则示例如下：
-
-数据A的定义：
-```python
-var_a = fluid.data(name='A', shape= [-1, 32], dtype='float32')
-```
-
-数据B的定义，变长数据的使用可以参考[LoDTensor](https://www.paddlepaddle.org.cn/documentation/docs/zh/beginners_guide/basic_concept/lod_tensor.html#cn-user-guide-lod-tensor)：
-```python
-var_b = fluid.data(name='B', shape=[-1, 1], lod_level=1, dtype='int64')
-```
-
-数据C的定义：
-```python
-var_c = fluid.data(name='C', shape=[-1, 1], dtype='int32')
-var_c.stop_gradient = False
-```
-
-当我们完成以上三个数据的定义后，在PaddleRec的模型定义中，还需将其加入model基类成员变量`self._data_var`
-
-```python
-self._data_var.append(var_a)
-self._data_var.append(var_b)
-self._data_var.append(var_c)
-```
-至此，我们完成了在组网中定义输入数据的工作。
-
-## Reader的实现
-
-### Reader的实现范式
-
-Reader的逻辑需要一个单独的python文件进行描述。我们试写一个`test_reader.py`，实现的具体流程如下：
-1. 首先我们需要引入Reader基类
-
-    ```python
-    from paddlerec.core.reader import ReaderBase
-    ```
-2. 创建一个子类，继承Reader的基类，训练所需Reader命名为`TrainerReader`
-    ```python
-    class TrainerReader(ReaderBase):
-        def init(self):
-            pass
-
-        def generator_sample(self, line):
-            pass
-    ```
-
-3. 在`init(self)`函数中声明一些在数据读取中会用到的变量，必要时可以在`config.yaml`文件中配置变量，利用`env.get_global_env()`拿到。
-   
-    比如，我们希望从yaml文件中读取一个数据预处理变量`avg=10`，目的是将数据A的数据缩小10倍，可以这样实现：
-
-    首先更改yaml文件，在某个space下加入该变量
-
-    ```yaml
-    ...
-    train:
-        reader:
-            avg: 10
-    ...
-    ```
-
-
-    再更改Reader的init函数
-
-    ```python
-    from paddlerec.core.utils import envs
-    class TrainerReader(Reader):
-        def init(self):
-            self.avg = envs.get_global_env("avg", None, "train.reader")
-
-        def generator_sample(self, line):
-            pass
-    ```
-
-4. 继承并实现基类中的`generate_sample(self, line)`函数，逐行读取数据。
-   - 该函数应返回一个可以迭代的reader方法(带有yield的函数不再是一个普通的函数，而是一个生成器generator，成为了可以迭代的对象，等价于一个数组、链表、文件、字符串etc.)
-   - 在这个可以迭代的函数中，如示例代码中的`def reader()`，我们定义数据读取的逻辑。以行为单位的数据进行截取，转换及预处理。
-   - 最后，我们需要将数据整理为特定的格式，才能够被PaddleRec的Reader正确读取，并灌入的训练的网络中。简单来说，数据的输出顺序与我们在网络中创建的`inputs`必须是严格一一对应的，并转换为类似字典的形式。
-    
-    示例： 假设数据ABC在文本数据中，每行以这样的形式存储：
-    ```shell
-    0.1,0.2,0.3...3.0,3.1,3.2 \t 99999,99998,99997 \t 1 \n
-    ```
-
-    则示例代码如下：
-    ```python
-    from paddlerec.core.utils import envs
-    class TrainerReader(Reader):
-        def init(self):
-            self.avg = envs.get_global_env("avg", None, "train.reader")
-
-        def generator_sample(self, line):
-            
-            def reader(self, line):
-                # 先分割 '\n'， 再以 '\t'为标志分割为list
-                variables = (line.strip('\n')).split('\t')
-
-                # A是第一个元素，并且每个数据之间使用','分割
-                var_a = variables[0].split(',') # list
-                var_a = [float(i) / self.avg for i in var_a] # 将str数据转换为float
-                
-
-                # B是第二个元素，同样以 ',' 分割
-                var_b = variables[1].split(',') # list
-                var_b = [int(i) for i in var_b] # 将str数据转换为int
-
-                # C是第三个元素, 只有一个元素，没有分割符
-                var_c = variables[2]
-                var_c = int(var_c) # 将str数据转换为int
-                var_c = [var_c] # 将单独的数据元素置入list中
-
-                # 将数据与数据名结合，组织为dict的形式
-                # 如下，output形式为{ A: var_a, B: var_b, C: var_c}
-                variable_name = ['A', 'B', 'C']
-                output = zip(variable_name, [var_a] + [var_b] + [var_c])
-
-                # 将数据输出，使用yield方法，将该函数变为了一个可迭代的对象
-                yield output
-
-    ```
-    
-    至此，我们完成了Reader的实现。
-
-
-### 在yaml文件中配置Reader
-
-在模型的yaml配置文件中，主要的修改是三个，如下
-
-```yaml
-reader:
-    batch_size: 2
-    class: "{workspace}/reader.py"
-    train_data_path: "{workspace}/data/train_data"
-    reader_debug_mode: False
-```
-
-batch_size: 顾名思义，是小批量训练时的样本大小
-class: 运行改模型所需reader的路径
-train_data_path: 训练数据所在文件夹
-reader_debug_mode: 测试reader语法，及输出是否符合预期的debug模式的开关
-
-
-## 数据及Reader示例-DNN
-
-Reader代码来源于[criteo_reader.py](../models/rank/criteo_reader.py), 组网代码来源于[model.py](../models/rank/dnn/model.py)
-
-### Criteo数据集格式
-
-CTR-DNN训练及测试数据集选用[Display Advertising Challenge](https://www.kaggle.com/c/criteo-display-ad-challenge/)所用的Criteo数据集。该数据集包括两部分：训练集和测试集。训练集包含一段时间内Criteo的部分流量，测试集则对应训练数据后一天的广告点击流量。
-每一行数据格式如下所示：
-```bash
-<label> <integer feature 1> ... <integer feature 13> <categorical feature 1> ... <categorical feature 26>
-```
-其中```<label>```表示广告是否被点击，点击用1表示，未点击用0表示。```<integer feature>```代表数值特征（连续特征），共有13个连续特征。```<categorical feature>```代表分类特征（离散特征），共有26个离散特征。相邻两个特征用```\t```分隔，缺失特征用空格表示。测试集中```<label>```特征已被移除。
-
-### Criteo数据集的预处理
-
-数据预处理共包括两步：
-- 将原始训练集按9:1划分为训练集和验证集
-- 数值特征（连续特征）需进行归一化处理，但需要注意的是，对每一个特征```<integer feature i>```，归一化时用到的最大值并不是用全局最大值，而是取排序后95%位置处的特征值作为最大值，同时保留极值。
-
-### CTR网络输入的定义
-
-正如前所述，Criteo数据集中，分为连续数据与离散（稀疏）数据，所以整体而言，CTR-DNN模型的数据输入层包括三个，分别是：`dense_input`用于输入连续数据，维度由超参数`dense_feature_dim`指定，数据类型是归一化后的浮点型数据。`sparse_input_ids`用于记录离散数据，在Criteo数据集中，共有26个slot，所以我们创建了名为`C1~C26`的26个稀疏参数输入，并设置`lod_level=1`，代表其为变长数据，数据类型为整数；最后是每条样本的`label`，代表了是否被点击，数据类型是整数，0代表负样例，1代表正样例。
-
-在Paddle中数据输入的声明使用`paddle.fluid.layers.data()`，会创建指定类型的占位符，数据IO会依据此定义进行数据的输入。
-
-稀疏参数输入的定义:
-```python
-def sparse_inputs():
-    ids = envs.get_global_env("hyper_parameters.sparse_inputs_slots", None)
-
-    sparse_input_ids = [
-        fluid.layers.data(name="S" + str(i),
-                            shape=[1],
-                            lod_level=1,
-                            dtype="int64") for i in range(1, ids)
-    ]
-    return sparse_input_ids
-```
-
-稠密参数输入的定义：
-```python
-def dense_input():
-    dim = envs.get_global_env("hyper_parameters.dense_input_dim", None)
-
-    dense_input_var = fluid.layers.data(name="D",
-                                        shape=[dim],
-                                        dtype="float32")
-    return dense_input_var
-```
-
-标签的定义：
-```python
-def label_input():
-    label = fluid.layers.data(name="click", shape=[1], dtype="int64")
-    return label
-```
-
-组合起来，正确的声明他们：
-```python
-self.sparse_inputs = sparse_inputs()
-self.dense_input = dense_input()
-self.label_input = label_input()
-
-self._data_var.append(self.dense_input)
-
-for input in self.sparse_inputs:
-    self._data_var.append(input)
-
-self._data_var.append(self.label_input)
-
-```
-
-
-### Criteo Reader写法
-
-```python
-# 引入PaddleRec的Reader基类
-from paddlerec.core.reader import ReaderBase
-# 引入PaddleRec的读取yaml配置文件的方法
-from paddlerec.core.utils import envs
-
-# 定义TrainReader，需要继承 paddlerec.core.reader.Reader
-class Reader(ReaderBase)::
-
-    # 数据预处理逻辑，继承自基类
-    # 如果无需处理， 使用pass跳过该函数的执行
-    def init(self):
-        self.cont_min_ = [0, -3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
-        self.cont_max_ = [20, 600, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50]
-        self.cont_diff_ = [20, 603, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50]
-        self.hash_dim_ = envs.get_global_env("hyper_parameters.sparse_feature_number", None, "train.model")
-        self.continuous_range_ = range(1, 14)
-        self.categorical_range_ = range(14, 40)
-
-    # 读取数据方法，继承自基类
-    # 实现可以迭代的reader函数，逐行处理数据
-    def generate_sample(self, line):
-        """
-        Read the data line by line and process it as a dictionary
-        """
-
-        def reader():
-            """
-            This function needs to be implemented by the user, based on data format
-            """
-            features = line.rstrip('\n').split('\t')
-
-            dense_feature = []
-            sparse_feature = []
-            for idx in self.continuous_range_:
-                if features[idx] == "":
-                    dense_feature.append(0.0)
-                else:
-                    dense_feature.append(
-                        (float(features[idx]) - self.cont_min_[idx - 1]) /
-                        self.cont_diff_[idx - 1])
-
-            for idx in self.categorical_range_:
-                sparse_feature.append(
-                    [hash(str(idx) + features[idx]) % self.hash_dim_])
-            label = [int(features[0])]
-            feature_name = ["D"]
-            for idx in self.categorical_range_:
-                feature_name.append("S" + str(idx - 13))
-            feature_name.append("label")
-            yield zip(feature_name, [dense_feature] + sparse_feature + [label])
-
-        return reader
-```
-
-
-### 调试Reader
-
-在Linux下运行时，默认启动`Dataset`模式，在Win/Mac下运行时，默认启动`Dataloader`模式。
-
-通过在`config.yaml`中添加或修改`reader_debug_mode=True`打开debug模式，只会结合组网运行reader的部分，读取10条样本，并print，方便您观察格式是否符合预期或隐藏bug。
-```yaml
-reader:
-    batch_size: 2
-    class: "{workspace}/../criteo_reader.py"
-    train_data_path: "{workspace}/data/train"
-    reader_debug_mode: True
-```
-
-修改后，使用paddlerec.run执行该修改后的yaml文件，可以观察输出。
-```bash
-python -m paddlerec.run -m ./models/rank/dnn/config.yaml -e single
-```
-
-### Dataset调试
-
-dataset输出的数据格式如下：
-` dense_input:size ; dense_input:value ; sparse_input:size ; sparse_input:value ; ... ; sparse_input:size ; sparse_input:value ; label:size ; label:value `
-
-基本规律是对于每个变量，会先输出其维度大小，再输出其具体值。
-
-直接debug `criteo_reader`理想的输出为(截取了一个片段)：
-```bash
-...
-13 0.0 0.00497512437811 0.05 0.08 0.207421875 0.028 0.35 0.08 0.082 0.0 0.4 0.0 0.08 1 737395 1 210498 1 903564 1 286224 1 286835 1 906818 1 90
-6116 1 67180 1 27346 1 51086 1 142177 1 95024 1 157883 1 873363 1 600281 1 812592 1 228085 1 35900 1 880474 1 984402 1 100885 1 26235 1 410878 1 798162 1 499868 1 306163 1 0
-...
-```
-可以看到首先输出的是13维的dense参数，随后是分立的sparse参数，最后一个是1维的label，数值为0，输出符合预期。
-
->使用Dataset的一些注意事项
-> - Dataset的基本原理：将数据print到缓存，再由C++端的代码实现读取，因此，我们不能在dataset的读取代码中，加入与数据读取无关的print信息，会导致C++端拿到错误的数据信息。
-> - dataset目前只支持在`unbuntu`及`CentOS`等标准Linux环境下使用，在`Windows`及`Mac`下使用时，会产生预料之外的错误，请知悉。
-
-### DataLoader调试
-
-dataloader的输出格式为`list: [ list[var_1], list[var_2], ... , list[var_3]]`，每条样本的数据会被放在一个 **list[list]** 中，list[0]为第一个variable。
-
-直接debug `criteo_reader`理想的输出为(截取了一个片段)：
-```bash
-...
-[[0.0, 0.004975124378109453, 0.05, 0.08, 0.207421875, 0.028, 0.35, 0.08, 0.082, 0.0, 0.4, 0.0, 0.08], [560746], [902436], [262029], [182633], [368411], [735166], [321120], [39572], [185732], [140298], [926671], [81559], [461249], [728372], [915018], [907965], [818961], [850958], [311492], [980340], [254960], [175041], [524857], [764893], [526288], [220126], [0]]
-...
-```
-可以看到首先输出的是13维的dense参数的list，随后是分立的sparse参数，各自在一个list中，最后一个是1维的label的list，数值为0，输出符合预期。

doc/distributed_train.md

@@ -48,7 +48,7 @@
 
 ```yaml
 # workspace
-workspace: "paddlerec.models.rank.dnn"
+workspace: "models/rank/dnn"
 
 mode: [single_cpu_train]
 runner:

doc/model_develop.md

@@ -92,7 +92,7 @@ def input_data(self, is_infer=False, **kwargs):
         return train_inputs
 ```
 
-更多数据读取教程，请参考[自定义数据集及Reader](custom_dataset_reader.md)
+更多数据读取教程，请参考[自定义数据集及Reader](custom_reader.md)
 
 
 ### 组网的定义

doc/pre_train_model.md

+# PaddleRec 预训练模型
+
+PaddleRec基于业务实践，使用真实数据，产出了推荐领域算法的若干预训练模型，方便开发者进行算法调研。
+
+## 文本分类预训练模型
+
+### 获取地址
+
+```bash
+wget xxx.tar.gz
+```
+
+### 使用方法
+
+解压后，得到的是一个paddle的模型文件夹，使用`PaddleRec/models/contentunderstanding/classification_finetue`模型进行加载

doc/train.md

@@ -20,7 +20,7 @@ python -m paddlerec.run -m paddlerec.models.xxx.yyy
 例如启动`recall`下的`word2vec`模型的默认配置;
 
 ```shell
-python -m paddlerec.run -m paddlerec.models.recall.word2vec
+python -m paddlerec.run -m models/recall/word2vec
 ```
 
 ### 2. 启动内置模型的个性化配置训练

models/contentunderstanding/classification/config.yaml

@@ -12,13 +12,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.contentunderstanding.classification"
+workspace: "models/contentunderstanding/classification"
 
 dataset:
 - name: data1
-  batch_size: 5
+  batch_size: 10
   type: DataLoader
-  data_path: "{workspace}/data/train_data"
+  data_path: "{workspace}/data/train"
+  data_converter: "{workspace}/reader.py"
+- name: dataset_infer
+  batch_size: 2
+  type: DataLoader
+  data_path: "{workspace}/data/test"
   data_converter: "{workspace}/reader.py"
 
 hyper_parameters:
@@ -26,23 +31,47 @@ hyper_parameters:
     class: Adagrad
     learning_rate: 0.001
   is_sparse: False
+  dict_dim: 33257
+  max_len: 100
+  cnn_dim: 128
+  cnn_filter_size1: 1
+  cnn_filter_size2: 2
+  cnn_filter_size3: 3
+  emb_dim: 128
+  hid_dim: 96
+  class_dim: 2
 
-mode: runner1
+mode: [train_runner,infer_runner]
 
 runner:
-- name: runner1
+- name: train_runner
   class: train
-  epochs: 10
+  epochs: 16
   device: cpu
-  save_checkpoint_interval: 2
-  save_inference_interval: 4
+  save_checkpoint_interval: 1
+  save_inference_interval: 1
   save_checkpoint_path: "increment"
   save_inference_path: "inference"
   save_inference_feed_varnames: []
   save_inference_fetch_varnames: []
+  init_model_path: "" 
+  print_interval: 10
+  phases: phase_train
+- name: infer_runner
+  class: infer
+  # device to run training or infer
+  device: cpu
+  print_interval: 1
+  init_model_path: "increment/14" # load model path
+  phases: phase_infer
 
+  
 phase:
-- name: phase1
+- name: phase_train
   model: "{workspace}/model.py"
   dataset_name: data1
   thread_num: 1
+- name: phase_infer
+  model: "{workspace}/model.py" # user-defined model
+  dataset_name: dataset_infer # select dataset by name
+  thread_num: 1

models/contentunderstanding/classification/data/preprocess.py

+# encoding=utf-8
+import os
+import sys
+
+
+def build_word_dict():
+    word_file = "word_dict.txt"
+    f = open(word_file, "r")
+    lines = f.readlines()
+    word_list_ids = range(1, len(lines) + 1)
+    word_dict = dict(zip([word.strip() for word in lines], word_list_ids))
+    f.close()
+    return word_dict
+
+
+def build_token_data(word_dict, txt_file, token_file):
+    max_text_size = 100
+
+    f = open(txt_file, "r")
+    fout = open(token_file, "w")
+    lines = f.readlines()
+    i = 0
+
+    for line in lines:
+        line = line.strip("\n").split("\t")
+        text = line[0].strip("\n").split(" ")
+        tokens = []
+        label = line[1]
+        for word in text:
+            if word in word_dict:
+                tokens.append(str(word_dict[word]))
+            else:
+                tokens.append("0")
+
+        seg_len = len(tokens)
+        if seg_len < 5:
+            continue
+        if seg_len >= max_text_size:
+            tokens = tokens[:max_text_size]
+            seg_len = max_text_size
+        else:
+            tokens = tokens + ["0"] * (max_text_size - seg_len)
+        text_tokens = " ".join(tokens)
+        fout.write(text_tokens + " " + str(seg_len) + " " + label + "\n")
+        if (i + 1) % 100 == 0:
+            print(str(i + 1) + " lines OK")
+        i += 1
+
+    fout.close()
+    f.close()
+
+
+word_dict = build_word_dict()
+
+txt_file = "test.tsv"
+token_file = "test.txt"
+build_token_data(word_dict, txt_file, token_file)
+
+txt_file = "dev.tsv"
+token_file = "dev.txt"
+build_token_data(word_dict, txt_file, token_file)
+
+txt_file = "train.tsv"
+token_file = "train.txt"
+build_token_data(word_dict, txt_file, token_file)

models/contentunderstanding/classification/data/test/test.txt

+5681 17044 4352 7574 16576 3574 32952 12211 18835 28961 15320 2019 21675 30604 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 14 1
+9054 31881 4449 12211 12488 5975 3574 28592 2547 2547 14132 3574 24908 5975 24285 10010 3574 31872 20925 9886 12211 26530 3567 30818 19640 22506 28312 19887 12211 28212 8576 3574 28592 12306 14132 539 33049 9039 14160 113 3567 19675 5511 2111 623 12068 12211 3574 18416 12068 19680 12211 30781 21946 1525 9886 3574 28109 31201 3567 25710 30503 30781 12068 19887 12211 22052 3574 2050 5402 10217 31201 1525 9698 14160 19887 3574 26209 24908 539 33049 9039 32949 8890 29693 3566 3566 11053 30781 26853 3567 3567 0 0 0 0 0 0 0 0 92 0
+19640 32771 31526 16576 13354 3574 5087 30781 7902 19037 12211 0 3574 4756 15048 11063 0 15019 16576 2019 29812 2276 22804 13275 2019 24599 12211 30294 6983 26606 1467 3574 18448 8052 16576 23091 32440 11034 16576 3574 1470 6983 1346 31382 13354 3574 11711 10074 28587 5030 19058 16576 2019 16497 6890 12223 30035 6983 1112 18448 30837 11280 24599 2019 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 64 0
+7513 19838 3562 32737 15474 3562 1887 15474 0 0 18835 19813 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 12 1
+30325 3574 30788 12211 25843 11533 30150 8937 11309 8690 12211 14166 2200 3574 15802 0 20424 14166 25336 113 16576 11533 24294 12211 26301 16576 3574 28592 16191 12211 8690 13743 0 517 12211 0 0 23958 3574 31019 19680 13841 15337 12211 23958 30781 28630 3574 8690 12700 11280 12211 23958 24908 20409 7481 8052 6094 4002 30245 3574 1526 9904 27032 31347 24006 12211 14166 0 9910 24908 12211 0 2019 25469 17293 27438 29774 13757 24908 22301 28505 25450 12211 14039 3574 28801 4621 4879 3574 623 9904 23958 14166 18417 4895 113 11114 2018 113 100 1
+113 16576 17947 28955 12211 24253 3574 22068 30167 12211 14039 30818 28640 7801 2019 7985 30167 5402 6805 0 12211 27645 33067 30151 3574 11110 12211 10710 4549 22708 4308 24908 25975 12211 26957 0 2019 17942 25575 227 19641 1525 13129 113 15492 23224 3574 21163 15565 23273 29004 12452 13233 27573 12211 12046 2019 302 19367 16576 27914 0 0 113 12211 28035 0 13743 13330 24390 12466 1525 12537 3574 18131 2019 9315 25720 27416 2276 15038 18162 10024 28955 3574 10097 18162 26594 12211 21949 3574 30788 12133 26362 1779 27386 21017 14295 1525 454 100 1
+33022 4169 19038 25096 3574 19185 113 25010 0 0 10511 17460 28972 6574 3574 1409 0 10010 3574 33022 129 16186 10511 17460 15182 3574 20235 10511 17460 11226 27150 13166 3562 18835 19038 5391 3574 22195 8052 28892 31948 10960 3574 13367 29338 15048 11030 22185 18621 28776 5205 2019 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 52 0
+23439 330 0 0 29655 12211 3574 4211 3574 19650 19640 13757 3562 0 0 8990 330 0 0 18920 12211 31924 6688 31857 15364 3574 19641 30781 18416 28952 9209 12211 118 10710 16912 3562 0 0 27771 330 0 0 10126 30325 3574 15374 4348 0 6356 28420 24193 29526 12211 10523 21872 3571 24383 1580 3574 17536 1525 14745 21674 10710 4952 14871 3574 14590 20306 7695 0 32718 3562 0 0 13260 330 0 0 5847 30325 3574 25951 26995 21163 22787 15535 20889 3574 27914 5391 130 2276 15243 6356 0 16576 3562 0 0 100 1
+24908 32568 24044 28952 16576 27914 28955 3574 14160 13543 16582 5536 2019 11711 3527 19675 12211 15474 3574 0 14160 31857 30927 2019 18416 9231 12486 12211 20374 3574 1111 30173 19058 3574 31857 31825 3574 30170 15501 21070 2019 31383 19640 5004 3574 31858 12211 6408 2733 8034 24870 12730 12211 16401 2019 18416 19640 9072 18416 12211 2313 12211 20374 3574 18416 2313 25575 19315 31383 20374 20161 24160 3574 11711 3527 3574 31383 20374 31857 28378 2019 1296 5402 23273 16576 2019 16497 28952 2019 9512 15038 5536 3574 11711 10486 15168 19641 21994 0 2019 100 1
+0 7902 5402 29107 16576 15535 15535 15535 0 19634 21017 12211 26505 14160 15129 0 15535 15535 15535 26211 4002 9749 23360 16576 15535 15535 15535 26040 15535 15535 15535 15535 11698 32986 19641 0 22421 15535 15535 15535 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 40 0
+28955 17755 3574 1735 18232 19262 12992 12230 3574 18416 30781 7388 19680 19643 16576 12211 3574 28952 9209 3574 16572 22360 2019 19680 19643 6414 12211 2011 27666 2012 3574 13757 32205 3574 14754 11280 12211 22186 7628 1827 17413 3574 19641 30781 31383 12211 4853 2019 33140 113 6047 6414 3310 31383 3574 4654 22360 6580 26147 12211 18696 2019 12306 6414 20539 3574 12680 22360 18624 8051 29384 1146 2019 18046 33188 16582 29384 12211 17311 13222 3574 18416 7453 28961 8014 3574 11711 18416 28961 17658 3574 29384 30781 19893 19643 15073 12211 32171 12211 2019 100 0
+28955 12211 30964 14590 28961 4412 29183 29493 6393 17111 29183 11670 12211 19636 23233 28961 4412 29183 25469 1112 16603 14590 16720 28961 9749 32365 23958 12211 33245 1525 11271 29183 29607 4694 8052 12068 32247 26813 29183 12229 6856 3674 330 30326 972 32948 29183 18416 28961 20161 1120 19641 30054 28955 330 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 55 0
+28587 26594 16393 14439 20100 8452 12211 11738 3574 20288 2276 2770 9051 29266 3574 27097 12211 0 14648 7902 5827 4308 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 22 1
+19083 3561 20034 30173 8356 3574 18416 18016 6154 13757 30827 23410 4879 5213 3566 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 15 1
+28587 14745 2018 1580 3574 19636 9052 14160 19683 16576 0 0 6007 5361 26370 5391 785 3574 0 17010 28587 27857 19048 20558 9051 3574 6007 0 0 22897 18323 1447 2019 0 0 32391 17536 24961 19048 9749 18448 3574 24283 6356 7648 26789 2019 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 47 0
+24908 18920 1400 665 16167 12211 17293 3574 13518 28952 8393 23504 3574 31266 12211 30781 4477 2019 4654 18896 4289 13841 4822 3574 24908 27376 15243 18416 8052 20077 17493 17317 3574 14842 16949 3574 12081 28961 2276 0 14399 20158 14398 16335 12211 3699 7697 6318 69 2019 11924 8053 27376 12211 14039 3574 21210 23273 3574 1732 30818 17942 22561 3083 2019 17268 12700 28892 9108 16576 26203 19037 23872 3574 14988 31773 3574 33140 1725 24908 0 8053 8052 13841 3574 25944 0 2019 4032 5025 13841 19185 12211 14039 3574 665 0 12211 4822 6988 100 1
+29728 31619 6149 5402 113 7317 11738 3574 31482 11924 16576 17657 6541 9761 3574 31224 5402 21141 3574 6356 16191 19640 14451 26154 7192 16076 3567 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 27 0
+29302 11364 19059 13652 12211 3574 7898 30781 6356 7961 14954 21752 7340 2019 29302 11401 8328 3574 20384 20034 1460 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 21 0
+4592 12211 31382 11030 3574 7961 6356 136 11714 31881 31478 3574 7957 11533 17413 3574 18835 14451 14550 11533 389 3574 14444 20444 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 24 1
+18416 24908 0 5233 22185 12211 29183 18956 30781 9668 8904 15168 18416 16108 29183 18416 29123 4351 28845 11709 11731 30486 21200 3574 4351 32986 8052 13757 11711 16497 25138 18448 3006 30326 20837 6356 16060 11231 13757 18448 11731 29173 3576 18835 27924 11711 11533 11225 3574 17386 15934 7288 0 26216 12211 1542 3574 24908 12511 18416 16060 11231 32842 18448 11731 29173 3574 18956 9668 31387 755 32986 18416 28972 18855 30781 18448 3006 30326 20837 30781 8052 13757 15048 18448 11731 29173 12211 3574 19640 18584 18416 32986 25710 18416 2276 29173 12211 22052 24908 100 0

models/contentunderstanding/classification/data/train_data/part-0.txt

-12 27 13  0 25 52 89 20 39  4 9 1
-78 10 61 58 29 79 85 16 46 41 9 1
-81 77 44  4  5 57 43 97 42 89 6 0
- 7 77 86  3 98 89 56 24  7 59 9 1
-65 89 99 27 65 98 16 89 42  0 3 0
-66 14 48 38 66  5 56 89 98 19 4 1
-78  7 10 20 77 16 37 43 59 23 6 1
-84 95 28 35  0 82 55 19 13 81 7 0
-34 32 98 37 43 51  6 38 20 40 9 0
-75 36 13 51 70 24 62 90 32 91 7 1
-13  5 49 21 57 21 67 85 74 14 1 0
-68 13 86 16 52 50 23 11 65 99 1 1
-15 20 75 55 15 90 54 54 15 91 9 0
-44 56 15 88 57  3 62 53 89 57 8 1
-23  8 40 25 60 33  8 69 44 88 7 1
-63 94  5 43 23 70 31 67 21 55 6 0
-44 11 64 92 10 37 30 84 19 71 5 1
-89 18 71 13 16 58 47 60 77 87 7 1
-13 48 56 39 98 53 32 93 13 91 7 0
-56 78 67 68 27 11 77 48 45 10 1 1
-52 12 14  5  2  8  3 36 33 59 6 0
-86 42 91 81  2  9 21  0 44  7 9 1
-96 27 82 55 81 30 91 41 91 58 2 1
-97 69 76 47 80 62 23 30 87 22 7 1
-42 56 25 47 42 18 80 53 15 57 7 0
-34 73 75 88 61 79 40 74 87 87 6 1
- 7 91  9 24 42 60 76 31 10 13 4 0
-21  1 46 59 61 54 99 54 89 55 5 1
-67 21  1 29 88  5  3 85 39 22 5 1
-90 99  7  8 17 77 73  3 32 10 5 0
-30 44 26 32 37 74 90 71 42 29 9 1
-79 68  3 24 21 37 35  3 76 23 6 1
- 3 66  7  4  2 88 94 64 47 81 6 1
-10 48 16 49 96 93 61 97 84 39 3 1
-73 28 67 59 89 92 17 24 52 71 3 1
-98  4 35 62 91  2 78 51 72 93 1 1
-37 42 96 10 48 49 84 45 59 47 5 1
-13 24  7 49 63 78 29 75 45 92 7 1
- 1  6 95 23 38 34 85 94 33 47 6 1
-99 63 65 39 72 73 91 20 16 45 9 0
-35  8 81 24 62  0 95  0 52 46 4 1
-58 66 88 42 86 94 91  8 18 92 7 0
-12 62 56 43 99 31 63 80 11  7 4 1
-22 36  1 39 69 20 56 75 17 15 7 0
-25 97 62 50 99 98 32  2 98 75 7 1
- 7 59 98 68 62 19 28 28 60 27 7 0
-39 63 43 45 43 11 40 81  4 25 6 0
-81 95 27 84 71 45 87 65 40 50 1 0
-82 21 69 55 71 92 52 65 90 16 3 0
-24  6  5 22 36 34 66 71  3 52 2 0
- 5 14 66 71 49 10 52 81 32 14 1 0
- 8 94 52 23 60 27 43 19 89 91 9 0
-26 14 36 37 28 94 46 96 11 80 8 1
-89 19 77 66 48 75 62 58 90 81 8 1
-25 43 95 21 25 81 39 79  9 74 9 0
-25  2 64 27 67 36 59 68 99 66 5 1
-13 46 41 55 89 93 79 83 32 52 6 0
-49 77 57  9 91 49 86 50 32  5 2 0
-94  7 53 54 70 69  5 51 59 91 5 1
-24 72 94 13 17 12  2 67  0 89 6 1
-70 38 19 27 38 87 72 41 98 84 6 1
-89 76 82  4 69 64 97 77 88 58 9 0
-67 41 99  1 80 38 96 24 67 59 3 1
-42 83 50 19 97 99 99 50 46 76 8 1
-43 99 63 40 93 15  3 57 11  0 1 0
-16 65 31 43 89 37 98 63 29 69 8 1
-39  5 65 45 12 82 46 87 82 93 8 0
-34 69 82 13  4 20 92 58 46 83 2 1
-46 79 87 57 87 23 72 95 37 88 8 0
-41 72 81 71 60 15 32  1  9 97 3 0
-84 98 15 78 39 82 89 74 46 32 9 0
-16 18 92 80 50 44 98 45 15 41 3 1
-74 78 81 40 17 65 38 21 27  9 1 0
-14 69 68 50 57 11 62  2 89 54 6 0
-70 29 79 29 44 56 33 27 25  4 3 1
-44 20 87 67 65 41 93 37 99 78 1 1
-93 57 87 11 33 40 21  3 47 87 9 1
- 8  3 24 49 99 48 40 22 99 41 2 0
-19 90  9 83 93 22 36 96 44 73 7 1
- 4 73  2 88 79 90 32 48 45 12 5 0
-24 58 34 67 85 62 84 48 14 79 5 1
-54 69 19 18 59 78 84 48 61 46 4 0
-72 69 95 26 30 74 49 30 95 61 8 0
-73 29 46 39 48 30 97 63 89 34 9 1
-51 32 44 22 70 69 91 81 74 52 3 0
-99 66 89 71 31 42  5 40 21 12 6 0
-58 26 59 56 91 49 79 57 57 74 6 1
-30 36 59 74  6 30 17  1 99 38 4 0
-43 48 77 86 67 25 38 36  3 91 4 1
-67 24 51 34 37  8 98 76 84 13 1 1
-73 47 88 15 32 99 67 26 28 89 3 1
-91 66 11 86  5 12 15 43 79 89 1 1
-15 60 43 58 61  0 62 32 98 29 9 0
-80 36 78 42 70 52  2 10 42 41 6 1
-36 16 46 34 96 39  8 21 86 54 5 1
-80 72 13  1 28 49 73 90 81 34 1 0
-73 64 86  9 94 49 44 38 47 64 2 0
-69 90 69 36 60 45 39  7 41 72 8 0
-31 86 54 82 81 77 93 99 68 63 1 1
-95 76 97 36 40 12  4 95 59 64 4 1
-88 20 64 40 27 11 96 40 41 73 6 0
-28 72 70 43 34 54 98 43 29 63 5 0
-78 72  4 47 47 38 73  8 65 40 3 1
-91 64 51 93  8 78 53 15 42 32 4 0
-34 36 45  9 16  0 51 40 90 29 2 1
-80 93 65 80 11 19 26 61 29  8 4 0
-94 11 60 36 58 98 43 90 64  1 1 0
-42 54 89 86 80 72 81 48 19 67 5 0
-81 25 30 60 59 20 75 38 75 29 6 0
-84 16 48 28 23 20 53 13 32 90 1 0
-58 31 77 68 27 88 51 97 70 93 8 1
-63 67 85  6 35 22 28 65  8  7 3 0
-54 75 93 58 98  9 15 37 61 38 6 1
-56 24 50 62 63 47  9  4 58 30 8 1
-64 91 32 68 50 90 51 86 52  6 1 1
-55 50 46 41 28  1 11 39 75  9 1 0
-23 27 98 73 25  7 89 48  7 44 4 1
-86 98 68  1 74 46 15 92 59 25 9 1
-95 86 72 13 33 60 62 83 96 84 1 0
- 9 58 37 50 57 16 78  0 21 80 2 0
-82 94 74 42  3 60 61 93 34 22 3 1
-16 97 97 14 47 50 90 35  9 58 5 0
-70 94 82 42 85 88 59 58  6 68 9 0
-14 58 24 44  8 29 12 18 26 80 7 0
-22 23  7 82 39 28 96 92 23 40 5 1
-40 31 72 94 20 81 89  4 42  1 5 0
-57 63 71 41 28  2 39 67 90 54 6 0
- 9 74  4 41 11 31 15 21 44 32 6 1
-31 28 66 66 61 78 72 80 82 88 3 1
-79 18  1 59 35 62  0 72 78 97 7 0
-14 19 30 63 38 37 12 15 54 15 6 1
-54 91 37 79 60 35 55 62 94 84 7 1
-10 55 78 96 45 55 35 56 54 70 6 1
-23 46 15 93 66 11 32 45 74 25 4 0
-51 55  9  9 88 59 21 66 87 12 1 1
-90 22 38 66 12  9 30 48 55 85 1 1
-39 23 82 29 57 76 79 56  3 19 2 0
- 7 72 76 15 90 23 40 40 33 39 4 1
-60 64 34 11 18 18 38 39 53 37 1 1
-85 72 51 47 83 90 32 96 78 23 9 1
-85 51 96 31 83 70 57 65 15  0 6 0
-41 11 56 94 40  6 62 86 68 83 7 0
-34 82 44 30  2  2 94 62 41 27 6 1
-54 86 50 83 76 65  0 87 80 70 7 0
-97 50 65 78  2 90 28  5 12 56 5 1
-34 19 68 93 11  9 14 87 22 70 9 0
-63 77 27 20 20 37 65 51 29 29 9 1
-22 79 98 57 56 97 43 49  4 80 4 1
- 6  4 35 54  4 36  1 79 85 35 6 0
-12 55 68 61 91 43 49  5 93 27 8 0
-64 22 69 16 63 20 28 60 13 35 7 1
- 9 19 60 89 62 29 47 33  6 13 4 0
-14 15 39 86 47 75  7 70 57 60 6 1
-90 63 12 43 28 46 39 97 83 42 6 0
-49  3  3 64 59 46 30 13 61 10 2 0
-79 47 29 47 54 38 50 66 18 63 5 1
-98 67  1 22 66 32 91 77 63 33 3 0
-72 22 10 27 28 44 29 66 71  1 7 0
-20 52 19 23  9 38  1 93 83 73 5 0
-88 57 22 64 93 66 20 90 78  2 7 1
-90 86 41 28 14 25 86 73  7 21 4 0
-63 91  0 29  2 78 86 76  9 20 4 1
- 3 57 91 37 21 85 80 99 18 79 1 1
-69 95 36  6 85 47 83 83 61 52 4 0
-72  4 34 16 59 78 56 70 27 44 9 1
-58 42  6 53 21  7 83 38 86 66 5 0
-22 86 22 21 86 22 83 38 62 19 4 0
-14 63 20 53 98 76 10 22 35 76 9 1
-16 88 13 66 37 33 11 40 61 97 2 1
-60  9 98 35 51 11 98 73 67 26 6 1
-25 48 87 93 58 58 15  9 23 13 7 1
-61 47 47 36 97 22 63 35  9 38 5 1
-94 49 41 38  0 81 59 39 13 65 3 0
-88 82 71 96 76 16 57 24 72 36 5 1
-28 46  8 95 94 86 63  1 42 63 6 0
-12 95 29 66 64 77 19 26 73 53 4 0
-19  5 52 34 13 62  6  4 25 58 5 0
-18 39 39 56 73 29  5 15 13 82 1 1
-50 66 99 67 76 25 43 12 24 67 9 0
-74 56 61 97 23 63 22 63  6 83 2 1
-10 96 13 49 43 20 58 19 99 58 7 1
- 2 95 31  4 99 91 27 90 85 32 3 0
-41 23 20 71 41 75 75 35 16 12 3 1
-21 33 87 57 19 27 94 36 80 10 6 0
- 8  0 25 74 14 61 86  8 42 82 9 0
-23 33 91 19 84 99 95 92 29 31 8 0
-94 94  5  6 98 23 37 65 14 25 6 1
-42 16 39 32  2 20 86 81 90 91 8 0
-72 39 20 63 88 52 65 81 77 96 4 0
-48 73 65 75 89 36 75 36 11 35 8 0
-79 74  3 29 63 20 76 46  8 82 5 0
- 7 46 38 77 79 92 71 98 30 35 6 0
-44 69 93 31 22 68 91 70 32 86 5 0
-45 38 77 87 64 44 69 19 28 82 9 0
-93 63 92 84 22 44 51 94  4 99 9 0
-77 10 49 29 59 55 44  7 95 39 2 0
-10 85 99  9 91 29 64 14 50 24 6 1
-74  4 21 12 77 36 71 51 50 31 9 1
-66 76 28 18 23 49 33 31  6 44 1 1
-92 50 90 64 95 58 93  4 78 88 6 1
-69 79 76 47 46 26 30 40 33 58 8 1
-97 12 87 82  6 18 57 49 49 58 1 1
-70 79 55 86 29 88 55 39 17 74 5 1
-65 51 45 62 54 17 59 12 29 79 5 0
- 5 63 82 51 54 97 54 36 57 46 3 0
-74 77 52 10 12  9 34 95  2  0 5 0
-50 20 22 89 50 70 55 98 80 50 1 0
-61 80  7  3 78 36 44 37 90 18 9 0
-81 13 55 57 88 81 66 55 18 34 2 1
-52 30 54 70 28 56 48 82 67 20 8 1
- 0 41 15 63 27 90 12 16 56 79 3 0
-69 89 54  1 93 10 15  2 25 59 8 0
-74 99 17 93 96 82 38 77 98 85 4 0
- 8 59 17 92 60 21 59 76 55 73 2 1
-53 56 79 19 29 94 86 96 62 39 3 1
-23 44 25 63 41 94 65 10  8 40 9 1
- 7 18 80 43 20 70 14 59 72 17 9 0
-84 97 79 14 37 64 23 68  8 24 2 0
-63 94 98 77  8 62 10 77 63 56 4 0
- 8 63 74 34 49 22 52 54 44 93 3 0
-94 48 92 58 82 48 53 34 96 25 2 0
-33 15  3 95 48 93  9 69 44 77 7 1
-69 72 80 77 64 24 52 21 36 49 2 0
-59 34 54 66 60 19 76 79 16 70 5 1
- 8 83  9 91 67 79 31 20 31 88 2 0
-64 95 46 95 78 63  4 60 66 63 7 1
-10 39 78 45 36  4 89 94 68 75 7 0
-81 52 70 11 48 15 40 63 29 14 8 1
-94 49 30 14 53 12 53 42 77 82 8 1
-40 88 46 20 54 84 76 15  2 73 2 1
-71 50 79 54 17 58 30 16 17 99 1 1
-74 79 74 61 61 36 28 39 89 36 6 0
-53 45 45 23 51 32 93 26 10  8 3 0
- 1 97  6 67 88 20 41 63 49  6 8 0
- 3 64 41 19 41 80 75 71 69 90 8 0
-31 90 38 93 52  0 38 86 41 68 9 1
-50 94 53  9 73 59 94  7 24 57 3 0
-87 11  4 62 96  7  0 59 46 11 6 1
-77 67 56 88 45 62 10 51 86 27 6 1
-62 62 59 99 83 84 79 97 56 37 5 0
-19 55  0 37 44 44  2  7 54 50 5 1
-23 60 11 83  6 48 20 77 54 31 6 0
-27 53 52 30  3 70 57 38 47 96 5 0
-75 14  5 83 72 46 47 64 14 12 7 0
-29 95 36 63 59 49 38 44 13 15 2 1
-38  3 70 89  2 94 89 74 33  6 8 1
-28 56 49 43 83 34  7 63 36 13 7 0
-25 90 23 85 50 65 36 10 64 38 5 0
-35 94 48 38 99 71 42 39 61 75 8 1
-28 73 34 22 51  8 52 98 74 19 8 1
-12 40 65 12  7 96 73 65 12 90 5 0
-42 42 48 16 80 14 48 29 29 45 5 0
-58 20  4  0 69 99 15  4 16  4 1 1
-93 30 90  5 23 63 25 30 99 32 7 1
-91 23 20 26 84 78 58 76 58 90 5 1
-33  2 36 59 55  9 79 34 92 57 9 0
-80 63 84 73 22 40 70 94 59 34 5 0
-49 95 50 32 90 22 18 66 46 32 2 0
-47 72  3 94 33 78 87 43 11 67 5 0
-76 44 86 81 95 48 79 46 11 65 8 1
-59 51 97 75 17  5 40 59 32 62 6 0
-41 13 58  7 54 84  8 84 27 55 1 0
-24 80 44 26 86 99 68 80 81 22 9 0
-12 45 16 44 66 76 33 53  3 20 9 0
-22  3 79  6 32 38 75 66 15 25 9 1
-51 48 26 53 33 26 18 74  9 39 5 1
-35 67 89 91 29 81 23 52 19 11 6 0
-64 50 43  1 43 49 19 20 84 19 8 0
-34  4  9 77 24 61 55 82 42 76 9 0
-37 84 94 33 67 60  3 95 78  8 9 0
-82 10 54 12 47 23 78 97  6 51 5 0
-70 40 38 47  5 38 83 70 37 90 2 0
-42 21 62 27 43 47 82 80 88 49 4 0
-68 68 67 12 38 13 32 30 93 27 3 1
- 5 44 98 28  5 81 20 56 10 34 9 1
-40 46 11 33 73 62 68 70 66 85 4 0
- 9 46 11 84  6 31 18 89 66 32 1 1
- 6 78 44 98 77 29 69 39 62 78 1 0
-47 90 18  0  3  8 12 20 51 75 4 1
-21 29 74 19 12 29 41 22 63 47 8 1
-22 59 64 62 18 89 19 92 87  8 8 0
- 6 21 24 58 14 53 18 93 62 15 8 0
-20 33 88 25 37 52  1 72 74 11 2 0
-90 49 28 53 28 80 22 81  0 46 9 0
-87 31 51 27 15 31 68 93  5  4 7 1
-21 72 60  2 24 79 22 24 77 61 9 0
-20  4  6 40 28 14 16 78 58 99 7 1
-80 35 98 20 91 35 47 29  3 19 2 1
-57 21 24 61 60 39 83 34 53  2 2 0
-74 86 78 78 18 44 20 94 85 71 4 1
-27 48 44 92 10 18 74 54 25 85 2 0
-74 77 28 75 74 91 69 36 95 68 7 0
-32 84 17 18 55 79 59 57 21 69 2 1
-69 77 40 98 83 40  4 66 39 83 1 1
-63 24 32 39 75 92 81 49  2 51 5 1
-35 40 84 71  3 16 82 91 44 52 8 0
-21 78 66  4 57 27 21 89  4 34 7 1
-94 18 57 49 88 26 29 76 56 67 6 0
-14 91 71 30  5 36 28 74 16 73 3 1
-93 36 43 46 77 44 59 19 56 84 3 0
-11 16  2 67 11 96 20 91 20 59 2 1
-72 79 26 99 90 71 56 46 35 99 3 0
-29 87 20 40 13 14 14 40 61 27 6 0
-41 64 28 51 56 52 87 67 37 91 6 1
-33 14  5 30 99 54 27 80 54 55 4 1
-60 44 73 91 71 53 54 95 59 81 6 0
-69 33 11 83  4 53 34 39 43 84 1 0
-73 31 19  4 50 20 66 73 94 88 4 0
-30 49 41 76  5 21 88 69 76  3 2 0
-18 50 27 76 67 38 87 16 52 87 5 1
-33 36 80  8 43 82 89 76 37  3 5 0
-98 21 61 24 58 13  9 85 56 74 1 1
-84 27 50 96  9 56 30 31 85 65 1 1
-65 74 40  2  8 40 18 57 30 38 1 1
-76 44 64  6 10 32 84 70 74 24 1 1
-14 29 59 34 27  8  0 37 27 68 3 0
- 6 47  5 77 15 41 93 49 59 83 4 1
-39 88 43 89 32 98 82  0  5 12 9 0
-78 79 30 26 58  6  9 58 37 65 8 1
-25 28 66 41 70 87 76 62 29 39 7 1

models/contentunderstanding/classification/model.py

@@ -20,28 +20,32 @@ from paddlerec.core.model import ModelBase
 class Model(ModelBase):
     def __init__(self, config):
         ModelBase.__init__(self, config)
-        self.dict_dim = 100
-        self.max_len = 10
-        self.cnn_dim = 32
-        self.cnn_filter_size = 128
-        self.emb_dim = 8
-        self.hid_dim = 128
-        self.class_dim = 2
-        self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse",
-                                             False)
+        self.dict_dim = envs.get_global_env("hyper_parameters.dict_dim")
+        self.max_len = envs.get_global_env("hyper_parameters.max_len")
+        self.cnn_dim = envs.get_global_env("hyper_parameters.cnn_dim")
+        self.cnn_filter_size1 = envs.get_global_env(
+            "hyper_parameters.cnn_filter_size1")
+        self.cnn_filter_size2 = envs.get_global_env(
+            "hyper_parameters.cnn_filter_size2")
+        self.cnn_filter_size3 = envs.get_global_env(
+            "hyper_parameters.cnn_filter_size3")
+        self.emb_dim = envs.get_global_env("hyper_parameters.emb_dim")
+        self.hid_dim = envs.get_global_env("hyper_parameters.hid_dim")
+        self.class_dim = envs.get_global_env("hyper_parameters.class_dim")
+        self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse")
 
     def input_data(self, is_infer=False, **kwargs):
         data = fluid.data(
             name="input", shape=[None, self.max_len], dtype='int64')
-        label = fluid.data(name="label", shape=[None, 1], dtype='int64')
         seq_len = fluid.data(name="seq_len", shape=[None], dtype='int64')
-        return [data, label, seq_len]
+        label = fluid.data(name="label", shape=[None, 1], dtype='int64')
+        return [data, seq_len, label]
 
     def net(self, input, is_infer=False):
         """ network definition """
         data = input[0]
-        label = input[1]
-        seq_len = input[2]
+        seq_len = input[1]
+        label = input[2]
 
         # embedding layer
         emb = fluid.embedding(
@@ -50,15 +54,31 @@ class Model(ModelBase):
             is_sparse=self.is_sparse)
         emb = fluid.layers.sequence_unpad(emb, length=seq_len)
         # convolution layer
-        conv = fluid.nets.sequence_conv_pool(
+        conv1 = fluid.nets.sequence_conv_pool(
+            input=emb,
+            num_filters=self.cnn_dim,
+            filter_size=self.cnn_filter_size1,
+            act="tanh",
+            pool_type="max")
+
+        conv2 = fluid.nets.sequence_conv_pool(
             input=emb,
             num_filters=self.cnn_dim,
-            filter_size=self.cnn_filter_size,
+            filter_size=self.cnn_filter_size2,
             act="tanh",
             pool_type="max")
 
+        conv3 = fluid.nets.sequence_conv_pool(
+            input=emb,
+            num_filters=self.cnn_dim,
+            filter_size=self.cnn_filter_size3,
+            act="tanh",
+            pool_type="max")
+
+        convs_out = fluid.layers.concat(input=[conv1, conv2, conv3], axis=1)
+
         # full connect layer
-        fc_1 = fluid.layers.fc(input=[conv], size=self.hid_dim)
+        fc_1 = fluid.layers.fc(input=convs_out, size=self.hid_dim, act="tanh")
         # softmax layer
         prediction = fluid.layers.fc(input=[fc_1],
                                      size=self.class_dim,
@@ -70,5 +90,7 @@ class Model(ModelBase):
         self._cost = avg_cost
         if is_infer:
             self._infer_results["acc"] = acc
+            self._infer_results["loss"] = avg_cost
         else:
             self._metrics["acc"] = acc
+            self._metrics["loss"] = avg_cost

models/contentunderstanding/classification/reader.py

@@ -23,9 +23,10 @@ class Reader(ReaderBase):
 
     def _process_line(self, l):
         l = l.strip().split()
-        data = l[0:10]
-        seq_len = l[10:11]
-        label = l[11:]
+        data = l[0:100]
+        seq_len = l[100:101]
+        label = l[101:]
+
         return data, label, seq_len
 
     def generate_sample(self, line):
@@ -37,6 +38,6 @@ class Reader(ReaderBase):
             data = [int(i) for i in data]
             label = [int(i) for i in label]
             seq_len = [int(i) for i in seq_len]
-            yield [('data', data), ('label', label), ('seq_len', seq_len)]
+            yield [('data', data), ('seq_len', seq_len), ('label', label)]
 
         return data_iter

models/contentunderstanding/classification/readme.md

+# classification文本分类模型
+
+以下是本例的简要目录结构及说明： 
+
+```
+├── data #样例数据
+	├── train
+		├── train.txt #训练数据样例
+	├── test
+    	├── test.txt #测试数据样例
+	├── preprocess.py #数据处理程序
+├── __init__.py
+├── README.md #文档
+├── model.py #模型文件
+├── config.yaml #配置文件
+├── reader.py #读取程序
+```
+
+注：在阅读该示例前，建议您先了解以下内容：
+[paddlerec入门教程](https://github.com/PaddlePaddle/PaddleRec/blob/master/README.md)  
+
+
+## 内容
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [运行环境](#运行环境)
+- [快速开始](#快速开始)
+- [效果复现](#效果复现)
+- [进阶使用](#进阶使用)
+- [FAQ](#FAQ)
+
+
+## 模型简介
+TextCNN网络是2014年提出的用来做文本分类的卷积神经网络，由于其结构简单、效果好，在文本分类、推荐等NLP领域应用广泛。对于文本分类问题，常见的方法无非就是抽取文本的特征。然后再基于抽取的特征训练一个分类器。 然而研究证明，TextCnn在文本分类问题上有着更加卓越的表现。从直观上理解，TextCNN通过一维卷积来获取句子中n-gram的特征表示。TextCNN对文本浅层特征的抽取能力很强，在短文本领域专注于意图分类时效果很好，应用广泛，且速度较快。  
+Yoon Kim在论文[EMNLP 2014][Convolutional neural networks for sentence classication](https://www.aclweb.org/anthology/D14-1181.pdf)提出了TextCNN并给出基本的结构。将卷积神经网络CNN应用到文本分类任务，利用多个不同size的kernel来提取句子中的关键信息（类似于多窗口大小的ngram），从而能够更好地捕捉局部相关性。模型的主体结构如图所示：  
+<p align="center">
+<img align="center" src="../../../doc/imgs/cnn-ckim2014.png">
+<p>
+
+## 数据准备
+情感倾向分析（Sentiment Classification，简称Senta）针对带有主观描述的中文文本，可自动判断该文本的情感极性类别并给出相应的置信度。情感类型分为积极、消极。情感倾向分析能够帮助企业理解用户消费习惯、分析热点话题和危机舆情监控，为企业提供有利的决策支持。  
+情感是人类的一种高级智能行为，为了识别文本的情感倾向，需要深入的语义建模。另外，不同领域（如餐饮、体育）在情感的表达各不相同，因而需要有大规模覆盖各个领域的数据进行模型训练。为此，我们通过基于深度学习的语义模型和大规模数据挖掘解决上述两个问题。效果上，我们基于开源情感倾向分类数据集ChnSentiCorp进行评测，模型在测试集上的准确率如表所示：  
+
+| 模型 | dev | test | 
+| :------| :------ | :------
+| TextCNN | 90.75% | 92.19% |
+
+
+您可以直接执行以下命令下载我们分词完毕后的数据集,文件解压之后，senta_data目录下会存在训练数据（train.tsv）、开发集数据（dev.tsv）、测试集数据（test.tsv）以及对应的词典（word_dict.txt）：  
+
+``` 
+wget https://baidu-nlp.bj.bcebos.com/sentiment_classification-dataset-1.0.0.tar.gz
+tar -zxvf sentiment_classification-dataset-1.0.0.tar.gz
+```
+数据格式为一句中文的评价语句，和一个代表情感信息的标签。两者之间用/t分隔，中文的评价语句已经分词，词之间用空格分隔。  
+```
+15.4寸 笔记本 的 键盘 确实 爽 ， 基本 跟 台式机 差不多 了 ， 蛮 喜欢 数字 小 键盘 ， 输 数字 特 方便 ， 样子 也 很 美观 ， 做工 也 相当 不错    1
+跟 心灵 鸡汤 没 什么 本质 区别 嘛 ， 至少 我 不 喜欢 这样 读 经典 ， 把 经典 都 解读 成 这样 有点 去 中国 化 的 味道 了 0
+```
+
+## 运行环境
+PaddlePaddle>=1.7.2
+
+python 2.7/3.5/3.6/3.7
+
+PaddleRec >=0.1
+
+os : windows/linux/macos
+
+
+## 快速开始
+本文提供了样例数据可以供您快速体验，在paddlerec目录下直接执行下面的命令即可启动训练： 
+
+```
+python -m paddlerec.run -m models/contentunderstanding/classification/config.yaml
+```   
+
+
+## 效果复现
+为了方便使用者能够快速的跑通每一个模型，我们在每个模型下都提供了样例数据。如果需要复现readme中的效果,请按如下步骤依次操作即可。  
+1. 确认您当前所在目录为PaddleRec/models/contentunderstanding/classification  
+2. 下载并解压数据集，命令如下：  
+``` 
+wget https://baidu-nlp.bj.bcebos.com/sentiment_classification-dataset-1.0.0.tar.gz
+tar -zxvf sentiment_classification-dataset-1.0.0.tar.gz
+```
+3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，将preprocess.py复制到senta_data文件中并执行，即可将数据集中提供的dev.tsv，test.tsv，train.tsv转化为可直接训练的dev.txt，test.txt，train.txt.
+```
+cp ./data/preprocess.py ./senta_data/
+cd senta_data/
+python preprocess.py
+```
+4. 创建存放训练集和测试集的目录，将数据放入目录中。
+```
+mkdir train
+mv train.txt train
+mkdir test
+mv dev.txt  test
+cd ..
+```  
+5. 打开文件config.yaml,更改其中的参数  
+
+将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
+将data1下的batch_size值从10改为128    
+将data1下的data_path改为：{workspace}/senta_data/train  
+将dataset_infer下的batch_size值从2改为256  
+将dataset_infer下的data_path改为：{workspace}/senta_data/test  
+
+6.  执行命令，开始训练：
+```
+python -m paddlerec.run -m ./config.yaml
+```
+
+7. 运行结果：
+```
+PaddleRec: Runner infer_runner Begin
+Executor Mode: infer
+processor_register begin
+Running SingleInstance.
+Running SingleNetwork.
+Running SingleInferStartup.
+Running SingleInferRunner.
+load persistables from increment/14
+batch: 1, acc: [0.91796875], loss: [0.2287855]
+batch: 2, acc: [0.91796875], loss: [0.22827303]
+batch: 3, acc: [0.90234375], loss: [0.27907994]
+```
+## 进阶使用
+  
+## FAQ

models/contentunderstanding/readme.md

@@ -22,12 +22,12 @@
 
 |       模型        |       简介        |       论文        |
 | :------------------: | :--------------------: | :---------: |
-| TagSpace | 标签推荐 | [EMNLP 2014][TagSpace: Semantic Embeddings from Hashtags](https://research.fb.com/publications/tagspace-semantic-embeddings-from-hashtags/) |
+| TagSpace | 标签推荐 | [EMNLP 2014][TagSpace: Semantic Embeddings from Hashtags](https://www.aclweb.org/anthology/D14-1194.pdf) |
 | Classification | 文本分类 | [EMNLP 2014][Convolutional neural networks for sentence classication](https://www.aclweb.org/anthology/D14-1181.pdf) |
 
 下面是每个模型的简介（注：图片引用自链接中的论文）
 
-[TagSpace模型](https://research.fb.com/publications/tagspace-semantic-embeddings-from-hashtags)
+[TagSpace模型](https://www.aclweb.org/anthology/D14-1194.pdf)
 <p align="center">
 <img align="center" src="../../doc/imgs/tagspace.png">
 <p>
@@ -37,86 +37,173 @@
 <img align="center" src="../../doc/imgs/cnn-ckim2014.png">
 <p>
 
-##使用教程(快速开始)
+## 使用教程(快速开始)
 ```
-python -m paddlerec.run -m paddlerec.models.contentunderstanding.tagspace
-python -m paddlerec.run -m paddlerec.models.contentunderstanding.classification
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd PaddleRec
+python -m paddlerec.run -m models/contentunderstanding/tagspace/config.yaml
+python -m paddlerec.run -m models/contentunderstanding/classification/config.yaml
 ```
 
 ## 使用教程（复现论文）
 
-###注意
+### 注意
 
 为了方便使用者能够快速的跑通每一个模型，我们在每个模型下都提供了样例数据。如果需要复现readme中的效果请使用以下提供的脚本下载对应数据集以及数据预处理。
 
-### 数据处理
 
 **（1）TagSpace**
 
+### 数据处理
 [数据地址](https://github.com/mhjabreel/CharCNN/tree/master/data/) , [备份数据地址](https://paddle-tagspace.bj.bcebos.com/data.tar)
- 
+
 数据格式如下
 ```
 "3","Wall St. Bears Claw Back Into the Black (Reuters)","Reuters - Short-sellers, Wall Street's dwindling\band of ultra-cynics, are seeing green again."
 ```
 
-数据解压后，将文本数据转为paddle数据，先将数据放到训练数据目录和测试数据目录
-
+本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，将原始数据存放在raw_big_train_data和raw_big_test_data两个目录下，并在python3环境下运行我们提供的text2paddle.py文件。即可生成可以直接用于训练的数据目录test_big_data和train_big_data。命令如下：
 ```
 mkdir raw_big_train_data
 mkdir raw_big_test_data
 mv train.csv raw_big_train_data
 mv test.csv raw_big_test_data
+python3 text2paddle.py raw_big_train_data/ raw_big_test_data/ train_big_data test_big_data big_vocab_text.txt big_vocab_tag.txt
 ```
 
-运行脚本text2paddle.py 生成paddle输入格式
+运行后的data目录：  
+
+```
+big_vocab_tag.txt  #标签词汇数
+big_vocab_text.txt #文本词汇数
+data.tar  #数据集
+raw_big_train_data  #数据集中原始的训练集
+raw_big_test_data  #数据集中原始的测试集
+train_data  #样例训练集
+test_data  #样例测试集
+train_big_data  #数据集经处理后的训练集
+test_big_data  #数据集经处理后的测试集
+text2paddle.py  #预处理文件
+```
 
+处理完成的数据格式如下：
 ```
-python text2paddle.py raw_big_train_data/ raw_big_test_data/ train_big_data test_big_data big_vocab_text.txt big_vocab_tag.txt
+2,27 7062 8390 456 407 8 11589 3166 4 7278 31046 33 3898 2897 426 1
+2,27 9493 836 355 20871 300 81 19 3 4125 9 449 462 13832 6 16570 1380 2874 5 0 797 236 19 3688 2106 14 8615 7 209 304 4 0 123 1
+2,27 12754 637 106 3839 1532 66 0 379 6 0 1246 9 307 33 161 2 8100 36 0 350 123 101 74 181 0 6657 4 0 1222 17195 1
 ```
 
+
 ### 训练
+退回tagspace目录中，打开文件config.yaml,更改其中的参数  
+将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
+将dataset下sample_1的batch_size值从10改为128   
+将dataset下sample_1的data_path改为：{workspace}/data/train_big_data  
+将dataset下inferdata的batch_size值从10改为500 
+将dataset下inferdata的data_path改为：{workspace}/data/test_big_data 
+执行命令，开始训练：
 ```
-cd modles/contentunderstanding/tagspace
-python -m paddlerec.run -m ./config.yaml # 自定义修改超参后，指定配置文件，使用自定义配置
+python -m paddlerec.run -m ./config.yaml
 ```
 
 ### 预测
+在跑完训练后，模型会开始在验证集上预测。
+运行结果：
+```
+PaddleRec: Runner infer_runner Begin
+Executor Mode: infer
+processor_register begin
+Running SingleInstance.
+Running SingleNetwork.
+Running SingleInferStartup.
+Running SingleInferRunner.
+load persistables from increment/9
+batch: 1, acc: [0.91], loss: [0.02495437]
+batch: 2, acc: [0.936], loss: [0.01941476]
+batch: 3, acc: [0.918], loss: [0.02116447]
+batch: 4, acc: [0.916], loss: [0.0219945]
+batch: 5, acc: [0.902], loss: [0.02242816]
+batch: 6, acc: [0.9], loss: [0.02421589]
+batch: 7, acc: [0.9], loss: [0.026441]
+batch: 8, acc: [0.934], loss: [0.01797657]
+batch: 9, acc: [0.932], loss: [0.01687362]
+batch: 10, acc: [0.926], loss: [0.02047823]
+batch: 11, acc: [0.918], loss: [0.01998716]
+batch: 12, acc: [0.898], loss: [0.0229556]
+batch: 13, acc: [0.928], loss: [0.01736144]
+batch: 14, acc: [0.93], loss: [0.01911209]
 ```
-# 修改对应模型的config.yaml, workspace配置为当前目录的绝对路径
-# 修改对应模型的config.yaml，mode配置infer_runner
-# 示例: mode: train_runner -> mode: infer_runner
-# infer_runner中 class配置为 class: infer
-# 修改phase阶段为infer的配置，参照config注释
 
-# 修改完config.yaml后 执行:
-python -m paddlerec.run -m ./config.yaml
+**（2）Classification**
+
+### 数据处理
+情感倾向分析（Sentiment Classification，简称Senta）针对带有主观描述的中文文本，可自动判断该文本的情感极性类别并给出相应的置信度。情感类型分为积极、消极。情感倾向分析能够帮助企业理解用户消费习惯、分析热点话题和危机舆情监控，为企业提供有利的决策支持。  
+情感是人类的一种高级智能行为，为了识别文本的情感倾向，需要深入的语义建模。另外，不同领域（如餐饮、体育）在情感的表达各不相同，因而需要有大规模覆盖各个领域的数据进行模型训练。为此，我们通过基于深度学习的语义模型和大规模数据挖掘解决上述两个问题。效果上，我们基于开源情感倾向分类数据集ChnSentiCorp进行评测。  
+您可以直接执行以下命令下载我们分词完毕后的数据集,文件解压之后，senta_data目录下会存在训练数据（train.tsv）、开发集数据（dev.tsv）、测试集数据（test.tsv）以及对应的词典（word_dict.txt）：
+
+``` 
+wget https://baidu-nlp.bj.bcebos.com/sentiment_classification-dataset-1.0.0.tar.gz
+tar -zxvf sentiment_classification-dataset-1.0.0.tar.gz
 ```
 
-**（2）Classification**
+数据格式为一句中文的评价语句，和一个代表情感信息的标签。两者之间用/t分隔，中文的评价语句已经分词，词之间用空格分隔。  
 
-### 训练
 ```
-cd modles/contentunderstanding/classification
-python -m paddlerec.run -m ./config.yaml # 自定义修改超参后，指定配置文件，使用自定义配置
+15.4寸 笔记本 的 键盘 确实 爽 ， 基本 跟 台式机 差不多 了 ， 蛮 喜欢 数字 小 键盘 ， 输 数字 特 方便 ， 样子 也 很 美观 ， 做工 也 相当 不错    1
+跟 心灵 鸡汤 没 什么 本质 区别 嘛 ， 至少 我 不 喜欢 这样 读 经典 ， 把 经典 都 解读 成 这样 有点 去 中国 化 的 味道 了 0
+```
+本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，将preprocess.py复制到senta_data文件中并执行，即可将数据集中提供的dev.tsv，test.tsv，train.tsv转化为可直接训练的dev.txt，test.txt，train.txt.
+```
+cp ./data/preprocess.py ./senta_data/
+cd senta_data/
+python preprocess.py
 ```
 
-### 预测
+### 训练
+创建存放训练集和测试集的目录，将数据放入目录中。
 ```
-# 修改对应模型的config.yaml, workspace配置为当前目录的绝对路径
-# 修改对应模型的config.yaml，mode配置infer_runner
-# 示例: mode: train_runner -> mode: infer_runner
-# infer_runner中 class配置为 class: infer
-# 修改phase阶段为infer的配置，参照config注释
+mkdir train
+mv train.txt train
+mkdir test
+mv dev.txt  test
+cd ..
+```  
+
+打开文件config.yaml,更改其中的参数  
+将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
+将data1下的batch_size值从10改为128    
+将data1下的data_path改为：{workspace}/senta_data/train  
+将dataset_infer下的batch_size值从2改为256  
+将dataset_infer下的data_path改为：{workspace}/senta_data/test  
 
-# 修改完config.yaml后 执行:
+执行命令，开始训练：
+```
 python -m paddlerec.run -m ./config.yaml
 ```
 
+### 预测
+在跑完训练后，模型会开始在验证集上预测。
+运行结果：  
+
+```
+PaddleRec: Runner infer_runner Begin
+Executor Mode: infer
+processor_register begin
+Running SingleInstance.
+Running SingleNetwork.
+Running SingleInferStartup.
+Running SingleInferRunner.
+load persistables from increment/14
+batch: 1, acc: [0.91796875], loss: [0.2287855]
+batch: 2, acc: [0.91796875], loss: [0.22827303]
+batch: 3, acc: [0.90234375], loss: [0.27907994]
+```
+
+
 ## 效果对比
 ### 模型效果 (测试)
 
-|       数据集        |       模型       |       loss        |       auc          |       acc         |       mae          |
-| :------------------: | :--------------------: | :---------: |:---------: | :---------: |:---------: |
-|       ag news dataset        |       TagSpace       |       --        |       --          |       --          |       --          |
-|       --        |       Classification       |       --        |       --          |       --          |       --          |
+|       数据集        |       模型       |       loss         |       acc         |
+| :------------------: | :--------------------: | :---------: |:---------: | 
+|       ag news dataset        |       TagSpace       |       0.2282        |       0.9179          | 
+|       ChnSentiCorp        |       Classification       |       0.9177        |       0.0199          | 

models/contentunderstanding/tagspace/config.yaml

@@ -12,20 +12,25 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.contentunderstanding.tagspace"
+workspace: "models/contentunderstanding/tagspace"
 
 dataset:
 - name: sample_1
-  type: QueueDataset
-  batch_size: 5
+  type: DataLoader
+  batch_size: 10
   data_path: "{workspace}/data/train_data"
   data_converter: "{workspace}/reader.py"
+- name: inferdata
+  type: DataLoader
+  batch_size: 10
+  data_path: "{workspace}/data/test_data"
+  data_converter: "{workspace}/reader.py" 
 
 hyper_parameters:
   optimizer:
     class: Adagrad
     learning_rate: 0.001
-  vocab_text_size: 11447
+  vocab_text_size: 75378
   vocab_tag_size: 4
   emb_dim: 10
   hid_dim: 1000
@@ -34,22 +39,34 @@ hyper_parameters:
   neg_size: 3
   num_devices: 1
 
-mode: runner1
+mode: [runner1,infer_runner]
 
 runner:
 - name: runner1
   class: train
   epochs: 10
   device: cpu
-  save_checkpoint_interval: 2
-  save_inference_interval: 4
+  save_checkpoint_interval: 1
+  save_inference_interval: 1
   save_checkpoint_path: "increment"
   save_inference_path: "inference"
   save_inference_feed_varnames: []
   save_inference_fetch_varnames: []
+  phases: phase1
+- name: infer_runner
+  class: infer
+  # device to run training or infer
+  device: cpu
+  print_interval: 1
+  init_model_path: "increment/9" # load model path
+  phases: phase_infer
 
 phase:
 - name: phase1
   model: "{workspace}/model.py"
   dataset_name: sample_1
   thread_num: 1
+- name: phase_infer
+  model: "{workspace}/model.py"
+  dataset_name: inferdata
+  thread_num: 1

models/contentunderstanding/tagspace/data/text2paddle.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+import sys
+import six
+import collections
+import os
+import csv
+import re
+import sys
+if six.PY2:
+    reload(sys)
+    sys.setdefaultencoding('utf-8')
+
+
+def word_count(column_num, input_file, word_freq=None):
+    """
+    compute word count from corpus
+    """
+    if word_freq is None:
+        word_freq = collections.defaultdict(int)
+    data_file = csv.reader(input_file)
+    for row in data_file:
+        for w in re.split(r'\W+', row[column_num].strip()):
+            word_freq[w] += 1
+    return word_freq
+
+
+def build_dict(column_num=2, min_word_freq=0, train_dir="", test_dir=""):
+    """
+    Build a word dictionary from the corpus,  Keys of the dictionary are words,
+    and values are zero-based IDs of these words.
+    """
+    word_freq = collections.defaultdict(int)
+    files = os.listdir(train_dir)
+    for fi in files:
+        with open(os.path.join(train_dir, fi), "r", encoding='utf-8') as f:
+            word_freq = word_count(column_num, f, word_freq)
+    files = os.listdir(test_dir)
+    for fi in files:
+        with open(os.path.join(test_dir, fi), "r", encoding='utf-8') as f:
+            word_freq = word_count(column_num, f, word_freq)
+
+    word_freq = [x for x in six.iteritems(word_freq) if x[1] > min_word_freq]
+    word_freq_sorted = sorted(word_freq, key=lambda x: (-x[1], x[0]))
+    words, _ = list(zip(*word_freq_sorted))
+    word_idx = dict(list(zip(words, six.moves.range(len(words)))))
+    return word_idx
+
+
+def write_paddle(text_idx, tag_idx, train_dir, test_dir, output_train_dir,
+                 output_test_dir):
+    files = os.listdir(train_dir)
+    if not os.path.exists(output_train_dir):
+        os.mkdir(output_train_dir)
+    for fi in files:
+        with open(os.path.join(train_dir, fi), "r", encoding='utf-8') as f:
+            with open(
+                    os.path.join(output_train_dir, fi), "w",
+                    encoding='utf-8') as wf:
+                data_file = csv.reader(f)
+                for row in data_file:
+                    tag_raw = re.split(r'\W+', row[0].strip())
+                    pos_index = tag_idx.get(tag_raw[0])
+                    wf.write(str(pos_index) + ",")
+                    text_raw = re.split(r'\W+', row[2].strip())
+                    l = [text_idx.get(w) for w in text_raw]
+                    for w in l:
+                        wf.write(str(w) + " ")
+                    wf.write("\n")
+
+    files = os.listdir(test_dir)
+    if not os.path.exists(output_test_dir):
+        os.mkdir(output_test_dir)
+    for fi in files:
+        with open(os.path.join(test_dir, fi), "r", encoding='utf-8') as f:
+            with open(
+                    os.path.join(output_test_dir, fi), "w",
+                    encoding='utf-8') as wf:
+                data_file = csv.reader(f)
+                for row in data_file:
+                    tag_raw = re.split(r'\W+', row[0].strip())
+                    pos_index = tag_idx.get(tag_raw[0])
+                    wf.write(str(pos_index) + ",")
+                    text_raw = re.split(r'\W+', row[2].strip())
+                    l = [text_idx.get(w) for w in text_raw]
+                    for w in l:
+                        wf.write(str(w) + " ")
+                    wf.write("\n")
+
+
+def text2paddle(train_dir, test_dir, output_train_dir, output_test_dir,
+                output_vocab_text, output_vocab_tag):
+    print("start constuct word dict")
+    vocab_text = build_dict(2, 0, train_dir, test_dir)
+    with open(output_vocab_text, "w", encoding='utf-8') as wf:
+        wf.write(str(len(vocab_text)) + "\n")
+
+    vocab_tag = build_dict(0, 0, train_dir, test_dir)
+    with open(output_vocab_tag, "w", encoding='utf-8') as wf:
+        wf.write(str(len(vocab_tag)) + "\n")
+
+    print("construct word dict done\n")
+    write_paddle(vocab_text, vocab_tag, train_dir, test_dir, output_train_dir,
+                 output_test_dir)
+
+
+train_dir = sys.argv[1]
+test_dir = sys.argv[2]
+output_train_dir = sys.argv[3]
+output_test_dir = sys.argv[4]
+output_vocab_text = sys.argv[5]
+output_vocab_tag = sys.argv[6]
+text2paddle(train_dir, test_dir, output_train_dir, output_test_dir,
+            output_vocab_text, output_vocab_tag)

models/contentunderstanding/tagspace/model.py

@@ -16,7 +16,6 @@ import paddle.fluid as fluid
 import paddle.fluid.layers.nn as nn
 import paddle.fluid.layers.tensor as tensor
 import paddle.fluid.layers.control_flow as cf
-
 from paddlerec.core.model import ModelBase
 from paddlerec.core.utils import envs
 
@@ -98,14 +97,19 @@ class Model(ModelBase):
             tensor.fill_constant_batch_size_like(
                 input=loss_part2, shape=[-1, 1], value=0.0, dtype='float32'),
             loss_part2)
-        avg_cost = nn.mean(loss_part3)
+        avg_cost = fluid.layers.mean(loss_part3)
+
         less = tensor.cast(cf.less_than(cos_neg, cos_pos), dtype='float32')
+        label_ones = fluid.layers.fill_constant_batch_size_like(
+            input=cos_neg, dtype='float32', shape=[-1, 1], value=1.0)
         correct = nn.reduce_sum(less)
+        total = fluid.layers.reduce_sum(label_ones)
+        acc = fluid.layers.elementwise_div(correct, total)
         self._cost = avg_cost
 
         if is_infer:
-            self._infer_results["correct"] = correct
-            self._infer_results["cos_pos"] = cos_pos
+            self._infer_results["acc"] = acc
+            self._infer_results["loss"] = self._cost
         else:
-            self._metrics["correct"] = correct
-            self._metrics["cos_pos"] = cos_pos
+            self._metrics["acc"] = acc
+            self._metrics["loss"] = self._cost

models/contentunderstanding/tagspace/readme.md

+# tagspace文本分类模型
+
+以下是本例的简要目录结构及说明： 
+
+```
+├── data #样例数据
+	├── train_data
+		├── small_train.csv #训练数据样例
+	├── test_data
+    	├── small_test.csv #测试数据样例
+	├── text2paddle.py #数据处理程序
+├── __init__.py
+├── README.md #文档
+├── model.py #模型文件
+├── config.yaml #配置文件
+├── reader.py #读取程序
+```
+
+注：在阅读该示例前，建议您先了解以下内容：
+
+[paddlerec入门教程](https://github.com/PaddlePaddle/PaddleRec/blob/master/README.md)
+
+## 内容
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [运行环境](#运行环境)
+- [快速开始](#快速开始)
+- [效果复现](#效果复现)
+- [进阶使用](#进阶使用)
+- [FAQ](#FAQ)
+
+
+## 模型简介
+tagspace模型是一种对文本打标签的方法，它主要学习从短文到相关主题标签的映射。论文中主要利用CNN做doc向量， 然后优化 f(w,t+),f(w,t-)的距离作为目标函数，得到了 t（标签）和doc在一个特征空间的向量表达，这样就可以找 doc的hashtags了。  
+
+论文[TAGSPACE: Semantic Embeddings from Hashtags](https://www.aclweb.org/anthology/D14-1194.pdf)中的网络结构如图所示，一层输入层，一个卷积层，一个pooling层以及最后一个全连接层进行降维。
+<p align="center">
+<img align="center" src="../../../doc/imgs/tagspace.png">
+<p>
+
+## 数据准备
+[数据地址](https://github.com/mhjabreel/CharCNN/tree/master/data/) , [备份数据地址](https://paddle-tagspace.bj.bcebos.com/data.tar)
+
+数据格式如下：  
+```
+"3","Wall St. Bears Claw Back Into the Black (Reuters)","Reuters - Short-sellers, Wall Street's dwindling\band of ultra-cynics, are seeing green again."
+```
+
+## 运行环境
+PaddlePaddle>=1.7.2  
+
+python 2.7/3.5/3.6/3.7  
+
+PaddleRec >=0.1  
+
+os : windows/linux/macos    
+
+
+## 快速开始
+本文提供了样例数据可以供您快速体验，在paddlerec目录下直接执行下面的命令即可启动训练： 
+
+```
+python -m paddlerec.run -m models/contentunderstanding/tagspace/config.yaml
+```   
+
+
+## 效果复现
+为了方便使用者能够快速的跑通每一个模型，我们在每个模型下都提供了样例数据。如果需要复现readme中的效果,请按如下步骤依次操作即可。  
+1. 确认您当前所在目录为PaddleRec/models/contentunderstanding/tagspace  
+2. 在data目录下载并解压数据集，命令如下：  
+``` 
+cd data
+wget https://paddle-tagspace.bj.bcebos.com/data.tar
+tar -xvf data.tar
+```
+3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，将原始数据存放在raw_big_train_data和raw_big_test_data两个目录下，并在python3环境下运行我们提供的text2paddle.py文件。即可生成可以直接用于训练的数据目录test_big_data和train_big_data。命令如下：
+```
+mkdir raw_big_train_data
+mkdir raw_big_test_data
+mv train.csv raw_big_train_data
+mv test.csv raw_big_test_data
+python3 text2paddle.py raw_big_train_data/ raw_big_test_data/ train_big_data test_big_data big_vocab_text.txt big_vocab_tag.txt
+```
+
+运行后的data目录：  
+
+```
+big_vocab_tag.txt  #标签词汇数
+big_vocab_text.txt #文本词汇数
+data.tar  #数据集
+raw_big_train_data  #数据集中原始的训练集
+raw_big_test_data  #数据集中原始的测试集
+train_data  #样例训练集
+test_data  #样例测试集
+train_big_data  #数据集经处理后的训练集
+test_big_data  #数据集经处理后的测试集
+text2paddle.py  #预处理文件
+```
+
+处理完成的数据格式如下：
+```
+2,27 7062 8390 456 407 8 11589 3166 4 7278 31046 33 3898 2897 426 1
+2,27 9493 836 355 20871 300 81 19 3 4125 9 449 462 13832 6 16570 1380 2874 5 0 797 236 19 3688 2106 14 8615 7 209 304 4 0 123 1
+2,27 12754 637 106 3839 1532 66 0 379 6 0 1246 9 307 33 161 2 8100 36 0 350 123 101 74 181 0 6657 4 0 1222 17195 1
+```
+
+4. 退回tagspace目录中，打开文件config.yaml,更改其中的参数  
+将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
+将dataset下sample_1的batch_size值从10改为128   
+将dataset下sample_1的data_path改为：{workspace}/data/train_big_data  
+将dataset下inferdata的batch_size值从10改为500 
+将dataset下inferdata的data_path改为：{workspace}/data/test_big_data 
+
+5.  执行命令，开始训练：
+```
+python -m paddlerec.run -m ./config.yaml
+```
+6. 运行结果：
+```
+PaddleRec: Runner infer_runner Begin
+Executor Mode: infer
+processor_register begin
+Running SingleInstance.
+Running SingleNetwork.
+Running SingleInferStartup.
+Running SingleInferRunner.
+load persistables from increment/9
+batch: 1, acc: [0.91], loss: [0.02495437]
+batch: 2, acc: [0.936], loss: [0.01941476]
+batch: 3, acc: [0.918], loss: [0.02116447]
+batch: 4, acc: [0.916], loss: [0.0219945]
+batch: 5, acc: [0.902], loss: [0.02242816]
+batch: 6, acc: [0.9], loss: [0.02421589]
+batch: 7, acc: [0.9], loss: [0.026441]
+batch: 8, acc: [0.934], loss: [0.01797657]
+batch: 9, acc: [0.932], loss: [0.01687362]
+batch: 10, acc: [0.926], loss: [0.02047823]
+batch: 11, acc: [0.918], loss: [0.01998716]
+batch: 12, acc: [0.898], loss: [0.0229556]
+batch: 13, acc: [0.928], loss: [0.01736144]
+batch: 14, acc: [0.93], loss: [0.01911209]
+```
+
+## 进阶使用
+  
+## FAQ

models/demo/movie_recommand/rank/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.demo.movie_recommand"
+workspace: "models/demo/movie_recommand"
 
 # list of dataset
 dataset:

models/demo/movie_recommand/recall/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.demo.movie_recommand"
+workspace: "models/demo/movie_recommand"
 
 # list of dataset
 dataset:

models/match/dssm/config.yaml

@@ -13,54 +13,56 @@
 # limitations under the License.
 
 
-workspace: "paddlerec.models.match.dssm"
+workspace: "models/match/dssm"
 
 dataset:
 - name: dataset_train
-  batch_size: 4
-  type: QueueDataset
+  batch_size: 8
+  type: DataLoader # or QueueDataset
   data_path: "{workspace}/data/train" 
   data_converter: "{workspace}/synthetic_reader.py"
 - name: dataset_infer
   batch_size: 1
-  type: QueueDataset
-  data_path: "{workspace}/data/train"
+  type: DataLoader # or QueueDataset
+  data_path: "{workspace}/data/test"
   data_converter: "{workspace}/synthetic_evaluate_reader.py"
 
 hyper_parameters:
   optimizer:
     class: sgd
-    learning_rate: 0.01
+    learning_rate: 0.001
     strategy: async
-  trigram_d: 1000
-  neg_num: 4
+  trigram_d: 1439
+  neg_num: 1
   fc_sizes: [300, 300, 128]
   fc_acts: ['tanh', 'tanh', 'tanh']
 
-mode: train_runner
+mode: [train_runner,infer_runner]
 # config of each runner.
 # runner is a kind of paddle training class, which wraps the train/infer process.
 runner:
 - name: train_runner
   class: train
   # num of epochs
-  epochs: 4
+  epochs: 3
   # device to run training or infer
   device: cpu
-  save_checkpoint_interval: 2 # save model interval of epochs
-  save_inference_interval: 4 # save inference
+  save_checkpoint_interval: 1 # save model interval of epochs
+  save_inference_interval: 1 # save inference
   save_checkpoint_path: "increment" # save checkpoint path
   save_inference_path: "inference" # save inference path
   save_inference_feed_varnames: ["query", "doc_pos"] # feed vars of save inference
   save_inference_fetch_varnames: ["cos_sim_0.tmp_0"] # fetch vars of save inference
   init_model_path: "" # load model path
   print_interval: 2
+  phases: phase1
 - name: infer_runner
   class: infer
   # device to run training or infer
   device: cpu
   print_interval: 1
   init_model_path: "increment/2" # load model path
+  phases: phase2
 
 # runner will run all the phase in each epoch
 phase:
@@ -68,7 +70,7 @@ phase:
   model: "{workspace}/model.py" # user-defined model
   dataset_name: dataset_train # select dataset by name
   thread_num: 1
-#- name: phase2
-#  model: "{workspace}/model.py" # user-defined model
-#  dataset_name: dataset_infer # select dataset by name
-#  thread_num: 1
+- name: phase2
+  model: "{workspace}/model.py" # user-defined model
+  dataset_name: dataset_infer # select dataset by name
+  thread_num: 1

models/match/dssm/data/preprocess.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+#encoding=utf-8
+
+import os
+import sys
+import numpy as np
+import random
+
+f = open("./zhidao", "r")
+lines = f.readlines()
+f.close()
+
+#建立字典
+word_dict = {}
+for line in lines:
+    line = line.strip().split("\t")
+    text = line[0].split(" ") + line[1].split(" ")
+    for word in text:
+        if word in word_dict:
+            continue
+        else:
+            word_dict[word] = len(word_dict) + 1
+
+f = open("./zhidao", "r")
+lines = f.readlines()
+f.close()
+
+lines = [line.strip().split("\t") for line in lines]
+
+#建立以query为key，以负例为value的字典
+neg_dict = {}
+for line in lines:
+    if line[2] == "0":
+        if line[0] in neg_dict:
+            neg_dict[line[0]].append(line[1])
+        else:
+            neg_dict[line[0]] = [line[1]]
+
+#建立以query为key，以正例为value的字典
+pos_dict = {}
+for line in lines:
+    if line[2] == "1":
+        if line[0] in pos_dict:
+            pos_dict[line[0]].append(line[1])
+        else:
+            pos_dict[line[0]] = [line[1]]
+
+#划分训练集和测试集
+query_list = list(pos_dict.keys())
+#print(len(query))
+random.shuffle(query_list)
+train_query = query_list[:90]
+test_query = query_list[90:]
+
+#获得训练集
+train_set = []
+for query in train_query:
+    for pos in pos_dict[query]:
+        if query not in neg_dict:
+            continue
+        for neg in neg_dict[query]:
+            train_set.append([query, pos, neg])
+random.shuffle(train_set)
+
+#获得测试集
+test_set = []
+for query in test_query:
+    for pos in pos_dict[query]:
+        test_set.append([query, pos, 1])
+    if query not in neg_dict:
+        continue
+    for neg in neg_dict[query]:
+        test_set.append([query, neg, 0])
+random.shuffle(test_set)
+
+#训练集中的query,pos,neg转化为词袋
+f = open("train.txt", "w")
+for line in train_set:
+    query = line[0].strip().split(" ")
+    pos = line[1].strip().split(" ")
+    neg = line[2].strip().split(" ")
+    query_token = [0] * (len(word_dict) + 1)
+    for word in query:
+        query_token[word_dict[word]] = 1
+    pos_token = [0] * (len(word_dict) + 1)
+    for word in pos:
+        pos_token[word_dict[word]] = 1
+    neg_token = [0] * (len(word_dict) + 1)
+    for word in neg:
+        neg_token[word_dict[word]] = 1
+    f.write(','.join([str(x) for x in query_token]) + "\t" + ','.join([
+        str(x) for x in pos_token
+    ]) + "\t" + ','.join([str(x) for x in neg_token]) + "\n")
+f.close()
+
+#测试集中的query和pos转化为词袋
+f = open("test.txt", "w")
+fa = open("label.txt", "w")
+for line in test_set:
+    query = line[0].strip().split(" ")
+    pos = line[1].strip().split(" ")
+    label = line[2]
+    query_token = [0] * (len(word_dict) + 1)
+    for word in query:
+        query_token[word_dict[word]] = 1
+    pos_token = [0] * (len(word_dict) + 1)
+    for word in pos:
+        pos_token[word_dict[word]] = 1
+    f.write(','.join([str(x) for x in query_token]) + "\t" + ','.join(
+        [str(x) for x in pos_token]) + "\n")
+    fa.write(str(label) + "\n")
+f.close()
+fa.close()

models/match/dssm/model.py

@@ -73,6 +73,7 @@ class Model(ModelBase):
 
         query_fc = fc(inputs[0], self.hidden_layers, self.hidden_acts,
                       ['query_l1', 'query_l2', 'query_l3'])
+
         doc_pos_fc = fc(inputs[1], self.hidden_layers, self.hidden_acts,
                         ['doc_pos_l1', 'doc_pos_l2', 'doc_pos_l3'])
         R_Q_D_p = fluid.layers.cos_sim(query_fc, doc_pos_fc)
@@ -93,7 +94,7 @@ class Model(ModelBase):
         prob = fluid.layers.softmax(concat_Rs, axis=1)
 
         hit_prob = fluid.layers.slice(
-            prob, axes=[0, 1], starts=[0, 0], ends=[4, 1])
+            prob, axes=[0, 1], starts=[0, 0], ends=[8, 1])
         loss = -fluid.layers.reduce_sum(fluid.layers.log(hit_prob))
         avg_cost = fluid.layers.mean(x=loss)
         self._cost = avg_cost

models/match/dssm/readme.md

+# DSSM文本匹配模型
+
+以下是本例的简要目录结构及说明： 
+
+```
+├── data #样例数据
+	├── train
+		├── train.txt #训练数据样例
+	├── test
+    	├── test.txt #测试数据样例
+	├── preprocess.py #数据处理程序
+├── __init__.py
+├── README.md #文档
+├── model.py #模型文件
+├── config.yaml #配置文件
+├── synthetic_reader.py #读取训练集的程序
+├── synthetic_evaluate_reader.py #读取测试集的程序
+├── transform.py #将数据整理成合适的格式方便计算指标
+├── run.sh #全量数据集中的训练脚本，从训练到预测并计算指标
+
+```
+
+注：在阅读该示例前，建议您先了解以下内容：
+
+[paddlerec入门教程](https://github.com/PaddlePaddle/PaddleRec/blob/master/README.md)
+
+## 内容
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [运行环境](#运行环境)
+- [快速开始](#快速开始)
+- [效果复现](#效果复现)
+- [进阶使用](#进阶使用)
+- [FAQ](#FAQ)
+
+
+## 模型简介
+DSSM是Deep Structured Semantic Model的缩写，即我们通常说的基于深度网络的语义模型，其核心思想是将query和doc映射到到共同维度的语义空间中，通过最大化query和doc语义向量之间的余弦相似度，从而训练得到隐含语义模型，达到检索的目的。DSSM有很广泛的应用，比如：搜索引擎检索，广告相关性，问答系统，机器翻译等。    
+DSSM 的输入采用 BOW（Bag of words）的方式，相当于把字向量的位置信息抛弃了，整个句子里的词都放在一个袋子里了。将一个句子用这种方式转化为一个向量输入DNN中。  
+Query 和 Doc 的语义相似性可以用这两个向量的 cosine 距离表示，然后通过softmax 函数选出与Query语义最相似的样本 Doc 。  
+
+模型的具体细节可以阅读论文[DSSM](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/cikm2013_DSSM_fullversion.pdf):
+<p align="center">
+<img align="center" src="../../../doc/imgs/dssm.png">
+<p>
+
+## 数据准备
+我们公开了自建的测试集，包括百度知道、ECOM、QQSIM、UNICOM 四个数据集。这里我们选取百度知道数据集来进行训练。执行以下命令可以获取上述数据集。
+```
+wget --no-check-certificate https://baidu-nlp.bj.bcebos.com/simnet_dataset-1.0.0.tar.gz
+tar xzf simnet_dataset-1.0.0.tar.gz
+rm simnet_dataset-1.0.0.tar.gz
+```
+
+## 运行环境
+PaddlePaddle>=1.7.2
+
+python 2.7/3.5/3.6/3.7
+
+PaddleRec >=0.1
+
+os : windows/linux/macos
+
+## 快速开始
+本文提供了样例数据可以供您快速体验，在paddlerec目录下执行下面的命令即可快速启动训练： 
+
+```
+python -m paddlerec.run -m models/match/dssm/config.yaml
+```   
+
+输出结果示例：
+```
+PaddleRec: Runner train_runner Begin
+Executor Mode: train
+processor_register begin
+Running SingleInstance.
+Running SingleNetwork.
+file_list : ['models/match/dssm/data/train/train.txt']
+Running SingleStartup.
+Running SingleRunner.
+!!! The CPU_NUM is not specified, you should set CPU_NUM in the environment variable list.
+CPU_NUM indicates that how many CPUPlace are used in the current task.
+And if this parameter are set as N (equal to the number of physical CPU core) the program may be faster.
+export CPU_NUM=32 # for example, set CPU_NUM as number of physical CPU core which is 32.
+!!! The default number of CPU_NUM=1.
+I0821 06:56:26.224299 31061 parallel_executor.cc:440] The Program will be executed on CPU using ParallelExecutor, 1 cards are used, so 1 programs are executed in parallel.
+I0821 06:56:26.231163 31061 build_strategy.cc:365] SeqOnlyAllReduceOps:0, num_trainers:1
+I0821 06:56:26.237023 31061 parallel_executor.cc:307] Inplace strategy is enabled, when build_strategy.enable_inplace = True
+I0821 06:56:26.240788 31061 parallel_executor.cc:375] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0
+batch: 2, LOSS: [4.538238]
+batch: 4, LOSS: [4.16424]
+batch: 6, LOSS: [3.8121371]
+batch: 8, LOSS: [3.4250507]
+batch: 10, LOSS: [3.2285979]
+batch: 12, LOSS: [3.2116117]
+batch: 14, LOSS: [3.1406002]
+epoch 0 done, use time: 0.357971906662, global metrics: LOSS=[3.0968776]
+batch: 2, LOSS: [2.6843479]
+batch: 4, LOSS: [2.546976]
+batch: 6, LOSS: [2.4103594]
+batch: 8, LOSS: [2.301374]
+batch: 10, LOSS: [2.264183]
+batch: 12, LOSS: [2.315862]
+batch: 14, LOSS: [2.3409634]
+epoch 1 done, use time: 0.22123003006, global metrics: LOSS=[2.344321]
+batch: 2, LOSS: [2.0882485]
+batch: 4, LOSS: [2.006743]
+batch: 6, LOSS: [1.9231766]
+batch: 8, LOSS: [1.8850241]
+batch: 10, LOSS: [1.8829436]
+batch: 12, LOSS: [1.9336565]
+batch: 14, LOSS: [1.9784685]
+epoch 2 done, use time: 0.212922096252, global metrics: LOSS=[1.9934461]
+PaddleRec Finish
+```
+## 效果复现
+为了方便使用者能够快速的跑通每一个模型，我们在每个模型下都提供了样例数据。如果需要复现readme中的效果,请按如下步骤依次操作即可。  
+1. 确认您当前所在目录为PaddleRec/models/match/dssm
+2. 在data目录下载并解压数据集，命令如下：  
+``` 
+cd data
+wget --no-check-certificate https://baidu-nlp.bj.bcebos.com/simnet_dataset-1.0.0.tar.gz
+tar xzf simnet_dataset-1.0.0.tar.gz
+rm simnet_dataset-1.0.0.tar.gz
+```
+3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，可以看见目录中存在一个名为zhidao的文件。然后能可以在python3环境下运行我们提供的preprocess.py文件。即可生成可以直接用于训练的数据目录test.txt,train.txt和label.txt。将其放入train和test目录下以备训练时调用。命令如下：
+```
+mv data/zhidao ./
+rm -rf data
+python3 preprocess.py
+rm -f ./train/train.txt
+mv train.txt ./train
+rm -f ./test/test.txt
+mv test.txt test
+cd ..
+```
+经过预处理的格式：  
+训练集为三个稀疏的BOW方式的向量：query,pos,neg  
+测试集为两个稀疏的BOW方式的向量：query,pos  
+label.txt中对应的测试集中的标签
+
+4. 退回dssm目录中，打开文件config.yaml,更改其中的参数  
+
+将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
+将dataset_train中的batch_size从8改为128
+将文件model.py中的 hit_prob = fluid.layers.slice(prob, axes=[0, 1], starts=[0, 0], ends=[8, 1])  
+    改为hit_prob = fluid.layers.slice(prob, axes=[0, 1], starts=[0, 0], ends=[128, 1]).当您需要改变batchsize的时候，end中第一个参数也需要随之变化
+
+5.  执行脚本，开始训练.脚本会运行python -m paddlerec.run -m ./config.yaml启动训练，并将结果输出到result文件中。然后启动transform.py整合数据，最后计算出正逆序指标：
+```
+sh run.sh
+```
+
+输出结果示例：
+```
+................run.................
+!!! The CPU_NUM is not specified, you should set CPU_NUM in the environment variable list.
+CPU_NUM indicates that how many CPUPlace are used in the current task.
+And if this parameter are set as N (equal to the number of physical CPU core) the program may be faster.
+
+export CPU_NUM=32 # for example, set CPU_NUM as number of physical CPU core which is 32.
+
+!!! The default number of CPU_NUM=1.
+I0821 07:16:04.512531 32200 parallel_executor.cc:440] The Program will be executed on CPU using ParallelExecutor, 1 cards are used, so 1 programs are executed in parallel.
+I0821 07:16:04.515708 32200 build_strategy.cc:365] SeqOnlyAllReduceOps:0, num_trainers:1
+I0821 07:16:04.518872 32200 parallel_executor.cc:307] Inplace strategy is enabled, when build_strategy.enable_inplace = True
+I0821 07:16:04.520995 32200 parallel_executor.cc:375] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0
+75
+pnr: 2.25581395349
+query_num: 11
+pair_num: 184 184
+equal_num: 44
+正序率： 0.692857142857
+97 43
+```
+6. 提醒：因为采取较小的数据集进行训练和测试，得到指标的浮动程度会比较大。如果得到的指标不合预期，可以多次执行步骤5，即可获得合理的指标。
+
+## 进阶使用
+  
+## FAQ

models/match/dssm/run.sh

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+#!/bin/bash
+echo "................run................."
+python -m paddlerec.run -m ./config.yaml >result1.txt
+grep -i "query_doc_sim" ./result1.txt >./result2.txt
+sed '$d' result2.txt >result.txt
+rm -f result1.txt
+rm -f result2.txt
+python transform.py
+sort -t $'\t' -k1,1 -k 2nr,2 pair.txt >result.txt
+rm -f pair.txt
+python ../../../tools/cal_pos_neg.py result.txt

models/match/dssm/transform.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+import random
+import numpy as np
+import sklearn.metrics
+
+label = []
+filename = './data/label.txt'
+f = open(filename, "r")
+f.readline()
+num = 0
+for line in f.readlines():
+    num = num + 1
+    line = line.strip()
+    label.append(line)
+f.close()
+print(num)
+
+filename = './result.txt'
+sim = []
+for line in open(filename):
+    line = line.strip().split(",")
+    line[1] = line[1].split(":")
+    line = line[1][1].strip(" ")
+    line = line.strip("[")
+    line = line.strip("]")
+    sim.append(float(line))
+
+filename = './data/test/test.txt'
+f = open(filename, "r")
+f.readline()
+query = []
+for line in f.readlines():
+    line = line.strip().split("\t")
+    query.append(line[0])
+f.close()
+
+filename = 'pair.txt'
+f = open(filename, "w")
+for i in range(len(sim)):
+    f.write(str(query[i]) + "\t" + str(sim[i]) + "\t" + str(label[i]) + "\n")
+f.close()

models/match/match-pyramid/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-workspace: "paddlerec.models.match.match-pyramid"
+workspace: "models/match/match-pyramid"
 
 dataset:
 - name: dataset_train

models/match/match-pyramid/readme.md

 # match-pyramid文本匹配模型
 
-## 介绍
+以下是本例的简要目录结构及说明： 
+
+```
+├── data #样例数据
+    ├── process.py #数据处理脚本
+    ├── relation.test.fold1.txt #评估计算指标时用到的关系文件
+    ├── train
+    	├── train.txt #训练数据样例
+    ├── test
+    	├── test.txt #测试数据样例
+├── __init__.py
+├── README.md #文档
+├── model.py #模型文件
+├── config.yaml #配置文件
+├── data_process.sh #数据下载和处理脚本
+├── eval.py #计算指标的评估程序
+├── run.sh #一键运行程序
+├── test_reader.py #测试集读取程序
+├── train_reader.py #训练集读取程序
+```
+
+注：在阅读该示例前，建议您先了解以下内容：
+
+[paddlerec入门教程](https://github.com/PaddlePaddle/PaddleRec/blob/master/README.md)
+
+## 内容
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [运行环境](#运行环境)
+- [快速开始](#快速开始)
+- [论文复现](#论文复现)
+- [进阶使用](#进阶使用)
+- [FAQ](#FAQ)
+
+
+## 模型简介
 在许多自然语言处理任务中，匹配两个文本是一个基本问题。一种有效的方法是从单词，短语和句子中提取有意义的匹配模式以产生匹配分数。受卷积神经网络在图像识别中的成功启发，神经元可以根据提取的基本视觉模式（例如定向的边角和边角）捕获许多复杂的模式，所以我们尝试将文本匹配建模为图像识别问题。本模型对齐原作者庞亮开源的tensorflow代码：https://github.com/pl8787/MatchPyramid-TensorFlow/blob/master/model/model_mp.py， 实现了下述论文中提出的Match-Pyramid模型：
 
 ```text
@@ -19,8 +55,23 @@
 3.关系文件：关系文件被用来存储两个句子之间的关系，如query 和document之间的关系。例如：relation.train.fold1.txt, relation.test.fold1.txt  
 4.嵌入层文件：我们将预训练的词向量存储在嵌入文件中。例如：embed_wiki-pdc_d50_norm  
 
-## 数据下载和预处理
-本文提供了数据集的下载以及一键生成训练和测试数据的预处理脚本，您可以直接一键运行:bash data_process.sh  
+## 运行环境
+PaddlePaddle>=1.7.2
+python 2.7/3.5/3.6/3.7
+PaddleRec >=0.1
+os : windows/linux/macos
+
+## 快速开始
+
+本文提供了样例数据可以供您快速体验，在paddlerec目录下直接执行下面的命令即可启动训练： 
+
+```
+python -m paddlerec.run -m models/match/match-pyramid/config.yaml
+```   
+
+## 论文复现
+1. 确认您当前所在目录为PaddleRec/models/match/match-pyramid
+2. 本文提供了原数据集的下载以及一键生成训练和测试数据的预处理脚本，您可以直接一键运行:bash data_process.sh  
 执行该脚本，会从国内源的服务器上下载Letor07数据集，删除掉data文件夹中原有的relation.test.fold1.txt和relation.train.fold1.txt，并将完整的数据集解压到data文件夹。随后运行 process.py 将全量训练数据放置于`./data/train`，全量测试数据放置于`./data/test`。并生成用于初始化embedding层的embedding.npy文件  
 执行该脚本的理想输出为：  
 ```
@@ -69,9 +120,11 @@ data/embed_wiki-pdc_d50_norm
 [./data/relation.test.fold1.txt]
         Instance size: 13652
 ```
+3. 打开文件config.yaml,更改其中的参数  
+
+将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）
 
-## 一键训练并测试评估
-本文提供了一键执行训练，测试和评估的脚本，您可以直接一键运行：bash run.sh  
+4. 随后，您直接一键运行：bash run.sh  即可得到复现的论文效果
 执行该脚本后，会执行python -m paddlerec.run -m ./config.yaml 命令开始训练并测试模型，将测试的结果保存到result.txt文件，最后通过执行eval.py进行评估得到数据的map指标  
 执行该脚本的理想输出为：  
 ```
@@ -79,16 +132,7 @@ data/embed_wiki-pdc_d50_norm
 13651
 336
 ('map=', 0.420878322843591)
-```
-
-## 每个文件的作用
-paddlerec可以：  
-通过config.yaml规定模型的参数  
-通过model.py规定模型的组网  
-使用train_reader.py读取训练集中的数据  
-使用test_reader.py读取测试集中的数据。  
-本文额外提供：  
-data_process.sh用来一键处理数据  
-run.sh用来一键启动训练，直接得出测试结果  
-eval.py通过保存的测试结果，计算map指标  
-如需详细了解paddlerec的使用方法请参考https://github.com/PaddlePaddle/PaddleRec/blob/master/README_CN.md 页面下方的教程。    
+```  
+## 进阶使用
+  
+## FAQ

models/match/match-pyramid/run.sh

 #!/bin/bash
 echo "................run................."
 python -m paddlerec.run -m ./config.yaml >result1.txt
-grep -A1 "prediction" ./result1.txt >./result.txt
+grep -i "prediction" ./result1.txt >./result.txt
 rm -f result1.txt
 python eval.py

models/match/multiview-simnet/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 # workspace
-workspace: "paddlerec.models.match.multiview-simnet"
+workspace: "models/match/multiview-simnet"
 
 # list of dataset
 dataset:

models/match/readme.md

 # 匹配模型库
 
 ## 简介
-我们提供了常见的匹配任务中使用的模型算法的PaddleRec实现, 单机训练&预测效果指标以及分布式训练&预测性能指标等。实现的模型包括 [DSSM](http://gitlab.baidu.com/tangwei12/paddlerec/tree/develop/models/match/dssm)、[MultiView-Simnet](http://gitlab.baidu.com/tangwei12/paddlerec/tree/develop/models/match/multiview-simnet)。
+我们提供了常见的匹配任务中使用的模型算法的PaddleRec实现, 单机训练&预测效果指标以及分布式训练&预测性能指标等。实现的模型包括 [DSSM](http://gitlab.baidu.com/tangwei12/paddlerec/tree/develop/models/match/dssm)、[MultiView-Simnet](http://gitlab.baidu.com/tangwei12/paddlerec/tree/develop/models/match/multiview-simnet)、[match-pyramid](https://github.com/PaddlePaddle/PaddleRec/tree/master/models/match/match-pyramid)。
 
 模型算法库在持续添加中，欢迎关注。
 
@@ -18,6 +18,8 @@
 | :------------------: | :--------------------: | :---------: |
 | DSSM | Deep Structured Semantic Models | [CIKM 2013][Learning Deep Structured Semantic Models for Web Search using Clickthrough Data](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/cikm2013_DSSM_fullversion.pdf) |
 | MultiView-Simnet | Multi-view Simnet for Personalized recommendation | [WWW 2015][A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/frp1159-songA.pdf) |
+| match-pyramid | Text Matching as Image Recognition | [arXiv W2016][Text Matching as Image Recognition](https://arxiv.org/pdf/1602.06359.pdf) |
+
 
 下面是每个模型的简介（注：图片引用自链接中的论文）
 
@@ -31,21 +33,26 @@
 <img align="center" src="../../doc/imgs/multiview-simnet.png">
 <p>
 
+[match-pyramid](https://arxiv.org/pdf/1602.06359.pdf):
+<p align="center">
+<img align="center" src="../../doc/imgs/match-pyramid.png">
+<p>
+
 ## 使用教程(快速开始)
-### 训练
-```shell
-python -m paddlerec.run -m paddlerec.models.match.dssm # dssm
-python -m paddlerec.run -m paddlerec.models.match.multiview-simnet # multiview-simnet
+### 训练&预测
+每个模型都提供了样例数据可以供您快速体验，在paddlerec目录下直接执行下面的命令即可启动训练：
+```
+python -m paddlerec.run -m models/match/dssm/config.yaml # dssm
+python -m paddlerec.run -m models/match/multiview-simnet/config.yaml # multiview-simnet
+python -m paddlerec.run -m models/contentunderstanding/match-pyramid/config.yaml #match-pyramid
 ```
+### 效果复现
+每个模型下的readme中都有详细的效果复现的教程，您可以进入模型的目录中详细查看  
 
-### 预测
-```shell
-# 修改对应模型的config.yaml, workspace配置为当前目录的绝对路径
-# 修改对应模型的config.yaml，mode配置infer_runner
-# 示例: mode: train_runner -> mode: infer_runner
-# infer_runner中 class配置为 class: infer
-# 修改phase阶段为infer的配置，参照config注释
+### 模型效果 (测试)
 
-# 修改完config.yaml后 执行:
-python -m paddlerec.run -m ./config.yaml # 以dssm为例
-```
+|       数据集        |       模型       |      正逆序比          |       map       |  
+| :------------------: | :--------------------: | :---------: |:---------: |
+|       zhidao       |       DSSM       |       2.25        |       --          | 
+|       Letor07        |       match-pyramid       |       --        |      0.42          | 
+|       zhidao        |       multiview-simnet       |       1.72        |       --          |

models/multitask/esmm/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-workspace: "paddlerec.models.multitask.esmm"
+workspace: "models/multitask/esmm"
 
 dataset:
   - name: dataset_train

models/multitask/mmoe/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.multitask.mmoe"
+workspace: "models/multitask/mmoe"
 
 dataset:
 - name: dataset_train

models/multitask/readme.md

@@ -44,9 +44,12 @@
 
 ## 使用教程(快速开始)
 ```shell
-python -m paddlerec.run -m paddlerec.models.multitask.mmoe # mmoe
-python -m paddlerec.run -m paddlerec.models.multitask.share-bottom # share-bottom
-python -m paddlerec.run -m paddlerec.models.multitask.esmm # esmm
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
+python -m paddlerec.run -m models/multitask/mmoe/config.yaml # mmoe
+python -m paddlerec.run -m models/multitask/share-bottom/config.yaml # share-bottom
+python -m paddlerec.run -m models/multitask/esmm/config.yaml # esmm
 ```
 
 ## 使用教程（复现论文）

models/multitask/share-bottom/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.multitask.share-bottom"
+workspace: "models/multitask/share-bottom"
 
 dataset:
 - name: dataset_train

models/rank/AutoInt/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.AutoInt"
+workspace: "models/rank/AutoInt"
 
 
 dataset:

models/rank/BST/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.BST"
+workspace: "models/rank/BST"
 
 dataset:
 - name: sample_1

models/rank/afm/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.afm"
+workspace: "models/rank/afm"
 
 dataset:
   - name: train_sample

models/rank/dcn/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.dcn"
+workspace: "models/rank/dcn"
 
 dataset:
   - name: train_sample

models/rank/deep_crossing/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.deep_crossing"
+workspace: "models/rank/deep_crossing"
 
 dataset:
   - name: train_sample

models/rank/deepfm/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.deepfm"
+workspace: "models/rank/deepfm"
 
 
 dataset:

models/rank/dien/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.dien"
+workspace: "models/rank/dien"
 
 dataset:
   - name: sample_1

models/rank/din/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.din"
+workspace: "models/rank/din"
 
 dataset:
   - name: sample_1

models/rank/dnn/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 # workspace
-workspace: "paddlerec.models.rank.dnn"
+workspace: "models/rank/dnn"
 
 # list of dataset
 dataset:
@@ -67,7 +67,6 @@ runner:
   save_inference_path: "inference" # save inference path
   save_inference_feed_varnames: [] # feed vars of save inference
   save_inference_fetch_varnames: [] # fetch vars of save inference
-  init_model_path: "" # load model path
   print_interval: 10
   phases: [phase1]
 

models/rank/ffm/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.ffm"
+workspace: "models/rank/ffm"
 
 dataset:
   - name: train_sample

models/rank/fgcnn/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.fgcnn"
+workspace: "models/rank/fgcnn"
 
 dataset:
   - name: train_sample

models/rank/fibinet/README.md

@@ -132,7 +132,7 @@ CPU环境
 
 ### 运行
 ```
-python -m paddlerec.run -m paddlerec.models.rank.fibinet
+python -m paddlerec.run -m models/rank/fibinet
 ```
 
 

models/rank/fibinet/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 # workspace
-workspace: "paddlerec.models.rank.fibinet"
+workspace: "models/rank/fibinet"
 
 # list of dataset
 dataset:

models/rank/flen/README.md

@@ -110,7 +110,7 @@ CPU环境
 ### 运行
 
 ```
-python -m paddlerec.run -m paddlerec.models.rank.flen
+python -m paddlerec.run -m models/rank/flen
 ```
 
 ## 论文复现

models/rank/flen/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 # workspace
-workspace: "paddlerec.models.rank.flen"
+workspace: "models/rank/flen"
 
 # list of dataset
 dataset:

models/rank/fm/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.fm"
+workspace: "models/rank/fm"
 
 dataset:
   - name: train_sample

models/rank/fnn/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.fnn"
+workspace: "models/rank/fnn"
 
 dataset:
   - name: train_sample

models/rank/logistic_regression/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.logistic_regression"
+workspace: "models/rank/logistic_regression"
 
 
 dataset:

models/rank/nfm/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.nfm"
+workspace: "models/rank/nfm"
 
 dataset:
   - name: train_sample

models/rank/pnn/config.yaml

@@ -15,7 +15,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.pnn"
+workspace: "models/rank/pnn"
 
 dataset:
   - name: train_sample

models/rank/readme.md

@@ -107,7 +107,7 @@ sh run.sh
 ### 训练
 ```
 cd modles/rank/dnn # 进入选定好的排序模型的目录 以DNN为例
-python -m paddlerec.run -m paddlerec.models.rank.dnn # 使用内置配置
+python -m paddlerec.run -m models/rank/dnn/config.yaml # 使用内置配置
 # 如果需要使用自定义配置，config.yaml中workspace需要使用改模型目录的绝对路径
 # 自定义修改超参后，指定配置文件，使用自定义配置
 python -m paddlerec.run -m ./config.yaml

models/rank/wide_deep/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.rank.wide_deep"
+workspace: "models/rank/wide_deep"
 
 
 dataset:

models/rank/xdeepfm/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 debug: false
-workspace: "paddlerec.models.rank.xdeepfm"
+workspace: "models/rank/xdeepfm"
 
 dataset:
   - name: sample_1

models/recall/fasttext/config.yaml

@@ -11,7 +11,7 @@
 # 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.
-workspace: "paddlerec.models.recall.fasttext"
+workspace: "models/recall/fasttext"
 
 # list of dataset
 dataset:

models/recall/gnn/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 # workspace
-workspace: "paddlerec.models.recall.gnn"
+workspace: "models/recall/gnn"
 
 # list of dataset
 dataset:

models/recall/gnn/readme.md

@@ -165,7 +165,7 @@ CPU环境
 
 ### 运行
 ```
-python -m paddlerec.run -m paddlerec.models.recall.gnn
+python -m paddlerec.run -m models/recall/gnn/config.yaml
 ```
 
 ### 结果展示

models/recall/gru4rec/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.recall.gru4rec"
+workspace: "models/recall/gru4rec"
 
 dataset:
 - name: dataset_train

models/recall/look-alike_recall/README.md

@@ -129,7 +129,7 @@ CPU环境
 
 ### 运行
 ```
-python -m paddlerec.run -m paddlerec.models.recall.look-alike_recall
+python -m paddlerec.run -m models/recall/look-alike_recall/config.yaml
 ```
 
 

models/recall/look-alike_recall/config.yaml

@@ -14,7 +14,7 @@
 
 # global settings 
 debug: false
-workspace: "paddlerec.models.recall.look-alike_recall"
+workspace: "models/recall/look-alike_recall"
 
 dataset:
 - name: sample_1

models/recall/ncf/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.recall.ncf"
+workspace: "models/recall/ncf"
 
 dataset:
 - name: dataset_train

models/recall/readme.md

@@ -62,12 +62,15 @@
 ## 使用教程(快速开始)
 ### 
 ```shell
-python -m paddlerec.run -m paddlerec.models.recall.word2vec # word2vec
-python -m paddlerec.run -m paddlerec.models.recall.ssr # ssr
-python -m paddlerec.run -m paddlerec.models.recall.gru4rec # gru4rec
-python -m paddlerec.run -m paddlerec.models.recall.gnn # gnn
-python -m paddlerec.run -m paddlerec.models.recall.ncf # ncf
-python -m paddlerec.run -m paddlerec.models.recall.youtube_dnn # youtube_dnn
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
+python -m paddlerec.run -m models/recall/word2vec/config.yaml # word2vec
+python -m paddlerec.run -m models/recall/ssr/config.yaml # ssr
+python -m paddlerec.run -m models/recall/gru4rec/config.yaml # gru4rec
+python -m paddlerec.run -m models/recall/gnn/config.yaml # gnn
+python -m paddlerec.run -m models/recall/ncf/config.yaml # ncf
+python -m paddlerec.run -m models/recall/youtube_dnn/config.yaml # youtube_dnn
 ```
 
 ## 使用教程（复现论文）
@@ -87,6 +90,9 @@ sh data_prepare.sh
 
 ### 训练
 ```bash
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
 cd modles/recall/gnn # 进入选定好的召回模型的目录 以gnn为例
 python -m paddlerec.run -m ./config.yaml # 自定义修改超参后，指定配置文件，使用自定义配置
 ```

models/recall/ssr/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.recall.ssr"
+workspace: "models/recall/ssr"
 
 dataset:
 - name: dataset_train

models/recall/word2vec/config.yaml

@@ -11,7 +11,7 @@
 # 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.
-workspace: "paddlerec.models.recall.word2vec" 
+workspace: "models/recall/word2vec"
 
 # list of dataset
 dataset:

models/recall/youtube_dnn/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-workspace: "paddlerec.models.recall.youtube_dnn"
+workspace: "models/recall/youtube_dnn"
 
 dataset:
 - name: dataset_train

models/rerank/listwise/config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-workspace: "paddlerec.models.rerank.listwise"
+workspace: "models/rerank/listwise"
 
 dataset:
 - name: dataset_train

models/rerank/readme.md

@@ -28,7 +28,10 @@
 
 ## 使用教程(快速开始)
 ```shell
-python -m paddlerec.run -m paddlerec.models.rerank.listwise # listwise
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
+python -m paddlerec.run -m models/rerank/listwise/config.yaml # listwise
 ```
 
 ## 使用教程（复现论文）

models/treebased/tdm/README.md

@@ -8,7 +8,10 @@
 2. 基于单机模型，可以进行分布式的参数服务器训练
 
 ```shell
-python -m paddlerec.run -m paddlerec.models.treebased.tdm
+git clone https://github.com/PaddlePaddle/PaddleRec.git paddle-rec
+cd paddle-rec
+
+python -m paddlerec.run -m models/treebased/tdm/config.yaml
 ```
 
 ## 树结构的准备

models/treebased/tdm/config.yaml

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-workspace: "paddlerec.models.treebased.tdm"
+workspace: "models/treebased/tdm"
 
 # list of dataset
 dataset:

run.py

@@ -16,7 +16,6 @@ import os
 import subprocess
 import sys
 import argparse
-import tempfile
 import warnings
 
 import copy
@@ -39,6 +38,7 @@ def engine_registry():
     engines["TRANSPILER"]["INFER"] = single_infer_engine
     engines["TRANSPILER"]["LOCAL_CLUSTER_TRAIN"] = local_cluster_engine
     engines["TRANSPILER"]["CLUSTER_TRAIN"] = cluster_engine
+    engines["TRANSPILER"]["ONLINE_LEARNING"] = online_learning
     engines["PSLIB"]["TRAIN"] = local_mpi_engine
     engines["PSLIB"]["LOCAL_CLUSTER_TRAIN"] = local_mpi_engine
     engines["PSLIB"]["CLUSTER_TRAIN"] = cluster_mpi_engine
@@ -259,6 +259,20 @@ def single_infer_engine(args):
     return trainer
 
 
+def online_learning(args):
+    trainer = "OnlineLearningTrainer"
+    single_envs = {}
+    single_envs["train.trainer.trainer"] = trainer
+    single_envs["train.trainer.threads"] = "2"
+    single_envs["train.trainer.engine"] = "online_learning"
+    single_envs["train.trainer.platform"] = envs.get_platform()
+    print("use {} engine to run model: {}".format(trainer, args.model))
+
+    set_runtime_envs(single_envs, args.model)
+    trainer = TrainerFactory.create(args.model)
+    return trainer
+
+
 def cluster_engine(args):
     def master():
         from paddlerec.core.engine.cluster.cluster import ClusterEngine

setup.cfg

-[easy_install]
-index_url=http://pip.baidu.com/pypi/simple
\ No newline at end of file

setup.py

@@ -38,15 +38,18 @@ readme = ""
 
 def build(dirname):
     package_dir = os.path.dirname(os.path.abspath(__file__))
+
     shutil.copytree(
-        package_dir, dirname, ignore=shutil.ignore_patterns(".git"))
+        package_dir,
+        dirname,
+        ignore=shutil.ignore_patterns(".git", "models", "build", "dist",
+                                      "*.md"))
+
     os.mkdir(os.path.join(dirname, "paddlerec"))
     shutil.move(
         os.path.join(dirname, "core"), os.path.join(dirname, "paddlerec"))
     shutil.move(
         os.path.join(dirname, "doc"), os.path.join(dirname, "paddlerec"))
-    shutil.move(
-        os.path.join(dirname, "models"), os.path.join(dirname, "paddlerec"))
     shutil.move(
         os.path.join(dirname, "tests"), os.path.join(dirname, "paddlerec"))
     shutil.move(
@@ -63,17 +66,8 @@ def build(dirname):
     package_dir = {'': dirname}
     package_data = {}
 
-    models_copy = [
-        'data/*.txt', 'data/*/*.txt', '*.yaml', '*.sh', 'tree/*.npy',
-        'tree/*.txt', 'data/sample_data/*', 'data/sample_data/train/*',
-        'data/sample_data/infer/*', 'data/*/*.csv', 'Criteo_data/*',
-        'Criteo_data/sample_data/train/*'
-    ]
-
     engine_copy = ['*/*.sh', '*/*.template']
     for package in packages:
-        if package.startswith("paddlerec.models."):
-            package_data[package] = models_copy
         if package.startswith("paddlerec.core.engine"):
             package_data[package] = engine_copy
 
@@ -98,16 +92,6 @@ build(dirname)
 shutil.rmtree(dirname)
 
 print(u'''
-\033[32m
-██████╗  █████╗ ██████╗ ██████╗ ██╗     ███████╗██████╗ ███████╗ ██████╗
-██╔══██╗██╔══██╗██╔══██╗██╔══██╗██║     ██╔════╝██╔══██╗██╔════╝██╔════╝
-██████╔╝███████║██║  ██║██║  ██║██║     █████╗  ██████╔╝█████╗  ██║
-██╔═══╝ ██╔══██║██║  ██║██║  ██║██║     ██╔══╝  ██╔══██╗██╔══╝  ██║
-██║     ██║  ██║██████╔╝██████╔╝███████╗███████╗██║  ██║███████╗╚██████╗
-╚═╝     ╚═╝  ╚═╝╚═════╝ ╚═════╝ ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝ ╚═════╝
-\033[0m
-\033[34m
 Installation Complete. Congratulations!
 How to use it ? Please visit our webside: https://github.com/PaddlePaddle/PaddleRec
-\033[0m
 ''')

tools/cal_pos_neg.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+#!/usr/bin/python  
+#-*- coding:utf-8 -*-  
+############################  
+#File Name: cal_pos_neg.py
+#Author: youqiheng 
+#Mail: youqiheng@baidu.com
+#Created Time: 2018-04-15 21:59:45
+############################ 
+"""
+docstring
+"""
+
+import os
+import sys
+
+if len(sys.argv) < 2:
+    print "usage:python %s input" % (sys.argv[0])
+    sys.exit(-1)
+
+fin = file(sys.argv[1])
+pos_num = 0
+neg_num = 0
+
+score_list = []
+label_list = []
+last_query = "-1"
+
+#0       12.786960       1
+#0       -1.480890       0
+cnt = 0
+query_num = 0
+pair_num = 0
+equal_num = 0
+for line in fin:
+    cols = line.strip().split("\t")
+    cnt += 1
+    if cnt % 500000 == 0:
+        print "cnt:", cnt, 1.0 * pos_num / neg_num
+    if len(cols) != 3:
+        continue
+
+    cur_query = cols[0]
+    if cur_query != last_query:
+        query_num += 1
+        for i in xrange(0, len(score_list)):
+            for j in xrange(i + 1, len(score_list)):
+                if label_list[i] == label_list[j]:
+                    continue
+                pair_num += 1
+                if (score_list[i] - score_list[j]) * (
+                        label_list[i] - label_list[j]) < 0:
+                    neg_num += 1
+                elif (score_list[i] - score_list[j]) * (
+                        label_list[i] - label_list[j]) > 0:
+                    pos_num += 1
+                else:
+                    equal_num += 1
+        score_list = []
+        label_list = []
+
+    last_query = cur_query
+
+    label = int(cols[2])
+
+    score_list.append(round(float(cols[1]), 6))
+    label_list.append(int(cols[2]))
+
+fin.close()
+
+for i in xrange(0, len(score_list)):
+    for j in xrange(i + 1, len(score_list)):
+        if label_list[i] == label_list[j]:
+            continue
+        pair_num += 1
+        if (score_list[i] - score_list[j]) * (label_list[i] - label_list[j]
+                                              ) < 0:
+            neg_num += 1
+        elif (score_list[i] - score_list[j]) * (label_list[i] - label_list[j]
+                                                ) > 0:
+            pos_num += 1
+        else:
+            equal_num += 1
+
+if neg_num > 0:
+    print "pnr:", 1.0 * pos_num / neg_num
+    print "query_num:", query_num
+    print "pair_num:", pos_num + neg_num + equal_num, pair_num
+    print "equal_num:", equal_num
+    print "正序率：", 1.0 * pos_num / (pos_num + neg_num)
+print pos_num, neg_num