#52: 完成launcher 和 config

data/mrc_ner/config/config.json

@@ -0,0 +1,86 @@
+{
+  "training_dataset": {
+    "__type__": "MSRAFlatNerDataset",
+    "__name_space__": "mrc_ner",
+    "dataset_file_path": "/Users/panxu/MyProjects/github/easytext/data/dataset/mrc_msra_ner/sample.json"
+  },
+
+  "validation_dataset": {
+    "__type__": "MSRAFlatNerDataset",
+    "__name_space__": "mrc_ner",
+    "dataset_file_path": "/Users/panxu/MyProjects/github/easytext/data/dataset/mrc_msra_ner/sample.json"
+  },
+
+  "model_collate": {
+    "__type__": "BertModelCollate",
+    "__name_space__": "mrc_ner",
+    "tokenizer": {
+      "__type__": "BertTokenizer",
+      "bert_dir": "/Users/panxu/MyProjects/github/easytext/data/pretrained/bert/chinese_roberta_wwm_large_ext_pytorch"
+    },
+
+    "max_length": 128
+  },
+
+
+
+  "model": {
+    "__type__": "BertRnnWithCrf",
+    "__name_space__": "mrc_ner",
+    "bert_dir": "/Users/panxu/MyProjects/github/easytext/data/pretrained/bert/chinese_roberta_wwm_large_ext_pytorch",
+    "dropout": 0.1
+  },
+
+  "loss": {
+    "__type__": "MRCBCELoss",
+    "__name_space__": "mrc_ner",
+    "start_weight": 1.0,
+    "end_weight": 1.0,
+    "match_weight": 1.0
+  },
+
+  "metric": {
+    "__type__": "MrcModelMetricAdapter",
+    "__name_space__": "mrc_ner"
+  },
+
+  "optimizer": {
+    "__type__": "MRCOptimizer",
+    "__name_space__": "mrc_ner",
+    "lr": 0.00002,
+    "eps": 0.00000001,
+    "weight_decay": 0.01
+  },
+
+  "lr_scheduler": {
+    "__type__": "MRCLrScheduler",
+    "__name_space__": "mrc_ner",
+    "max_lr": 0.00002,
+    "final_div_factor": 10000,
+    "total_steps": null,
+  },
+  "grad_rescaled": null,
+
+  "process_group_parameter": {
+    "__type__": "ProcessGroupParameter",
+    "__name_space__": "__easytext__",
+    "host": "127.0.0.1",
+    "port": "2345",
+    "backend": "nccl"
+  },
+
+  "distributed_data_parallel_parameter": {
+    "__type__": "DistributedDataParallelParameter",
+    "__name_space__": "__easytext__",
+    "find_unused_parameters": false
+  },
+
+  "num_epoch": 500,
+  "patient": 100,
+  "num_check_point_keep": 1,
+
+  "devices": ["cpu"],
+  "serialize_dir": "/Users/panxu/MyProjects/github/easytext/data/mrc_ner/serialize",
+  "train_batch_size": 4,
+  "test_batch_size": 8
+}

mrc/data/bert_model_collate.py

@@ -19,7 +19,10 @@
 
 from mrc.data import MRCModelInputs
 
+from easytext.component.register import ComponentRegister
 
+
+@ComponentRegister.register(name_space="mrc_ner")
 class BertModelCollate:
     """
     Bert Model Collate

mrc/data/msra_flat_ner_dataset.py

@@ -15,9 +15,10 @@
 
 from easytext.data import Instance
 from easytext.data.dataset import Dataset
-from easytext.data.tokenizer import ZhTokenizer
+from easytext.component.register import ComponentRegister
 
 
+@ComponentRegister.register(name_space="mrc_ner")
 class MSRAFlatNerDataset(Dataset):
     """
     MSRA flat ner dataset

mrc/launcher.py

@@ -33,11 +33,20 @@
 from easytext.distributed import ProcessGroupParameter
 from easytext.utils.json_util import json2str
 from easytext.component.register import Registry
+from easytext.utils.bert_tokenizer import bert_tokenizer
 
+from mrc.data import MSRAFlatNerDataset
+from mrc.data import BertModelCollate
+from mrc.models import MRCNer
+from mrc.loss import MRCBCELoss
+from mrc.metric import MrcModelMetricAdapter
+from mrc.optimizer import MRCOptimizer
+from mrc.optimizer import MRCLrScheduler
 
-class MrcLauncher(Launcher):
+
+class MrcNerLauncher(Launcher):
     """
-    ner 训练的启动器
+    mrc ner 训练的启动器
     """
 
     NEW_TRAIN = 0
@@ -63,7 +72,7 @@ def _preprocess(self):
         logging.info(f"config:\n{self.config}\n")
 
         serialize_dir = self.config.serialize_dir
-        if self._train_type == MrcLauncher.NEW_TRAIN:
+        if self._train_type == MrcNerLauncher.NEW_TRAIN:
             # 清理 serialize dir
             if os.path.isdir(serialize_dir):
                 shutil.rmtree(serialize_dir)
@@ -131,5 +140,5 @@ def _start(self, rank: Optional[int], world_size: int, device: torch.device) ->
         logging.fatal("--config 参数为空!")
         exit(-1)
     logging.info(f"config file path: {parsed_args.config}")
-    ner_launcher = MrcLauncher(config_file_path=parsed_args.config, train_type=MrcLauncher.NEW_TRAIN)
+    ner_launcher = MrcNerLauncher(config_file_path=parsed_args.config, train_type=MrcNerLauncher.NEW_TRAIN)
     ner_launcher()

mrc/loss/__init__.py

@@ -10,3 +10,5 @@
 Authors: PanXu
 Date:    2021/10/27 14:11:00
 """
+
+from .mrc_bce_loss import MRCBCELoss

mrc/loss/mrc_bce_loss.py

@@ -18,7 +18,10 @@
 
 from mrc.models import MRCNerOutput
 
+from easytext.component.register import ComponentRegister
 
+
+@ComponentRegister.register(name_space="mrc_ner")
 class MRCBCELoss:
     """
     基于 bce 的 los

mrc/metric/mrc_metric.py

@@ -25,7 +25,7 @@
 from mrc.metric import MRCF1Metric
 
 
-@ComponentRegister.register(name_space="mrc")
+@ComponentRegister.register(name_space="mrc_ner")
 class MrcModelMetricAdapter(ModelMetricAdapter):
     """
     Ner Model Metric Adapter

mrc/models/mrc_ner.py

@@ -25,6 +25,7 @@
 from easytext.utils.seed_util import set_seed
 
 from mrc.models import MRCNerOutput
+from easytext.component.register import ComponentRegister
 
 
 class MultiNonLinearClassifier(Module):
@@ -45,6 +46,7 @@ def forward(self, input_features):
         return features_output2
 
 
+@ComponentRegister.register(name_space="mrc_ner")
 class MRCNer(Module):
     """
     基于 MRC 的 ner 模型

mrc/optimizer/__init__.py

@@ -12,3 +12,4 @@
 """
 
 from .mrc_optimizer import MRCOptimizer
+from .mrc_lr_scheduler import MRCLrScheduler

mrc/optimizer/mrc_lr_scheduler.py

@@ -15,8 +15,10 @@
 
 from easytext.model import Model
 from easytext.optimizer import LRSchedulerFactory
+from easytext.component.register import ComponentRegister
 
 
+@ComponentRegister.register(name_space="mrc_ner")
 class MRCLrScheduler(LRSchedulerFactory):
 
     def __init__(self, max_lr: float, final_div_factor: float, total_steps: int = None, pct_start: float = 0):

mrc/optimizer/mrc_optimizer.py

@@ -12,10 +12,12 @@
 """
 from easytext.model import Model
 from easytext.optimizer import OptimizerFactory
+from easytext.component.register import ComponentRegister
 
 from transformers import AdamW
 
 
+@ComponentRegister.register(name_space="mrc_ner")
 class MRCOptimizer(OptimizerFactory):
 
     def __init__(self, lr: float, eps: float, weight_decay: float):