重置了trainer into utils

Author: 刘宇

utils/__init__.py

@@ -0,0 +1 @@
+from trainer import Trainer

utils/trainer.py

@@ -0,0 +1,50 @@
+import torch.nn as nn
+import torch
+from sklearn.model_selection import StratifiedKFold, KFold
+import transformers
+from tqdm import tqdm
+transformers.logging.set_verbosity_error()
+
+
+class Train(object):
+    def __init__(self, model: nn.Module, epochs=20, lr=1e-5, weight_decay=0, 
+                 show_batch=50, use_cuda=True,compute_metrics=None):
+        self.model = model
+        self.device = torch.device(
+            "cuda:0" if torch.cuda.is_available() else "cpu")
+        self.model.to(self.device)
+        self.epochs = epochs
+        self.lr = lr
+        self.show_batch = show_batch
+        self.weight_decay = weight_decay
+        self.optimizer = torch.optim.AdamW(
+            self.model.parameters(), lr=self.lr, weight_decay=self.weight_decay)
+        self.compute_metrics = compute_metrics
+    def train(self, dataset_train, dataset_eval=None):
+        for epoch in range(self.epochs):
+            self.model.train()
+            for idx, batch in tqdm(enumerate(dataset_train), total=len(dataset_train)):
+                batch = {k: v.to(self.device) for k, v in batch.items()}
+                score = self.compute_metrics(batch)
+                loss = self.model(**batch)["loss"]
+                self.optimizer.zero_grad()
+                loss.backward()
+                self.optimizer.step()
+                if idx % self.show_batch == 0:
+                    print(
+                        'Epoch [{}/{}],batch:{} Loss: {:.4f}'.format(self.epochs, epoch + 1, idx, loss.item()))
+            with torch.no_grad():  # 评估时禁止计算梯度
+                self.evaluation(dataset_eval, epoch)
+
+    def evaluation(self, dataset_eval, epoch):
+        print("evaluation.....")
+        self.model.eval()
+        score_list = []
+        for idx, batch in tqdm(enumerate(dataset_eval), total=len(dataset_eval)):
+            batch = {k: v.to(self.device) for k, v in batch.items()}
+            score = self.compute_metrics(batch)
+            score_list.append(score)
+        score = sum(score_list) / len(score_list) * 100
+        print(
+            'Epoch [{}/{}], score: {:.4f} %'.format(self.epochs, epoch + 1, score))
+   

深度学习/nlp/MRC_阅读理解/main.py

@@ -57,19 +57,23 @@ def forward(self, input_ids, attention_mask, token_type_ids, start_positions, en
 
 
 import sys
+from os.path import dirname as rn
+import os
 
-sys.path.append("..")
-from trainer import Trainer
-
+path = rn(rn(rn(rn(__file__))))
+print(path)
+sys.path.append(path)
+from utils.trainer import Trainer
+from utils import Trainer
 if __name__ == '__main__':
-    # train_dataset = get_squad_dataset(data_dir="./data/cmrc2018_public", filename="train.json")
-    # dev_dataset = get_squad_dataset(data_dir="./data/cmrc2018_public", filename="dev.json")
+    train_dataset = get_squad_dataset(data_dir="./data/cmrc2018_public", filename="train.json")
+    dev_dataset = get_squad_dataset(data_dir="./data/cmrc2018_public", filename="dev.json")
     import pickle
 
     # pickle.dump(train_dataset, open("train.pt", "wb"))
     # pickle.dump(dev_dataset, open("dev.pt", "wb"))
-    train_dataset = pickle.load(open("train.pt", "rb"))
-    dev_dataset = pickle.load(open("dev.pt", "rb"))
+    # train_dataset = pickle.load(open("train.pt", "rb"))
+    # dev_dataset = pickle.load(open("dev.pt", "rb"))
     batch_size = 8
     train_dataloader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
     dev_dataloader = DataLoader(dataset=dev_dataset, batch_size=batch_size, shuffle=True)

深度学习/nlp/__init__.py.py renamed to 深度学习/nlp/__init__.py

@@ -69,6 +69,4 @@ def f(batch):
     model = BertForSequenceClassification.from_pretrained(check_point, num_labels=16)
 
 
-# train
-# # 'dac_predictions.npy'
-train(model, data_loader, dev_loader, test_loader, args)
+    train(model, data_loader, dev_loader, test_loader, args)

深度学习/nlp/trainer.py

@@ -86,7 +86,7 @@ def train(model, train_iter, dev_iter, test_iter, args):
                 break
         if flag:
             break
-    test(config, model, test_iter, args)
+    test(model, test_iter, args)
 
 
 def test(config, model, test_iter, args):