Merge pull request #35 from minggnim/bugfix-metrics

Bugfix metrics

notebooks/02_b_multitask_model_training_example.ipynb

@@ -189,7 +189,7 @@
     "tokenizer_dir = output_dir / 'tokenizer'\n",
     "model_file = output_dir / 'mtl.bin'\n",
     "\n",
-    "model,_,_ = trainer.load_checkpoint('./drive/MyDrive/chatbot/chkpt/chkpt9.pt', model)\n",
+    "model,_,_ = trainer.load_checkpoint('../chkpt/chkpt9.pt', model)\n",
     "pretrained_tokenizer.save_pretrained(tokenizer_dir)\n",
     "AutoModelForMTL.save_model(model, model_file)"
    ]

setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = nlp-models
-version = 2.3.1
+version = 2.3.2
 author= Ming Gao
 author_email = ming_gao@outlook.com
 url = https://github.com/minggnim/nlp-models

src/bert_classifier/data.py

@@ -3,14 +3,20 @@
 '''
 import torch
 from torch.utils.data import Dataset
-from .bert import bert_encoder
 
 
 class CustomDataset(Dataset):
     '''
     Class to construct torch Dataset from dataframe
     '''
-    def __init__(self, dataframe, data_field, label_field, tokenizer, max_len, multi_label):
+    def __init__(self,
+                 dataframe,
+                 data_field,
+                 label_field,
+                 tokenizer,
+                 max_len,
+                 multi_label=False
+                 ):
         self.max_len = max_len
         self.data = dataframe
         self.tokenizer = tokenizer

src/bert_classifier/train.py

@@ -23,12 +23,12 @@ def loss_fn(outputs, targets):
     return torch.nn.CrossEntropyLoss()(outputs, targets)
 
 
-def optimizer(model, learning_rate):
-    optimizer = torch.optim.AdamW(
+def optimizer_obj(model, learning_rate):
+    opt = torch.optim.AdamW(
         params=model.parameters(),
         lr=learning_rate
     )
-    return optimizer
+    return opt
 
 
 def custom_trainer(model, optimizer, train_dataloader, test_dataloader, epochs=5, device='cpu'):

src/multi_task_model/metrics.py

@@ -11,8 +11,8 @@
 def hamming_distance(
     outputs,
     targets,
-    multi_label: bool = False, 
-    num_labels: Optional[int] = None, 
+    num_labels: int,
+    multi_label: bool = False,
     device: torch.device = torch.device('cpu'),
     average: Literal['micro', 'macro', 'weighted', 'none'] = 'macro'
 ):

src/multi_task_model/trainer.py

@@ -15,7 +15,7 @@
 class Configs:
     epochs: int = 1
     optimizer_class = torch.optim.AdamW
-    optimizer_params: Dict[str, float] = field(default_factory = lambda: ({"lr": 2e-5})) 
+    optimizer_params: Dict[str, float] = field(default_factory=lambda: ({"lr": 2e-5}))
     weight_decay: float = 0.01
     scheduler: str = 'WarmupLinear'
     warmup_steps: int = 10000
@@ -25,14 +25,14 @@ class Configs:
 
 
 class Trainer:
-    def __init__(self, 
+    def __init__(self,
                  model: torch.nn.Module,
                  train_dataloader: DataLoader,
                  test_dataloader: DataLoader,
                  configs: Configs,
-                 metrics = accuracy,
+                 metrics=accuracy,
                  device: torch.device = torch.device('cpu'),
-                 chkpt_dir = Path('../chkpt')
+                 chkpt_dir=Path('../chkpt')
                  ):
         self.model = model
         self.train_dataloader = train_dataloader
@@ -54,19 +54,19 @@ def set_params(self):
                 param.requires_grad = False
 
         self.optimizer = self.get_optimizer(
-            list(self.model.named_parameters()), 
-            self.configs.optimizer_class, 
-            self.configs.optimizer_params, 
+            list(self.model.named_parameters()),
+            self.configs.optimizer_class,
+            self.configs.optimizer_params,
             self.configs.weight_decay)
 
         self.scheduler = self.get_scheduler(
-            self.optimizer, 
-            self.configs.scheduler, 
-            self.configs.warmup_steps, 
-            len(self.train_dataloader)*self.configs.epochs)
+            self.optimizer,
+            self.configs.scheduler,
+            self.configs.warmup_steps,
+            len(self.train_dataloader) * self.configs.epochs)
 
     def train(self):
-        epochs = tqdm(range(1, self.configs.epochs+1), leave = True, desc="Training...")
+        epochs = tqdm(range(1, self.configs.epochs + 1), leave=True, desc="Training...")
         for epoch in epochs:
             self.model.train()
             epochs.set_description(f"EPOCH {epoch} / {self.configs.epochs} | training...")
@@ -105,10 +105,12 @@ def train_one_epoch(self, epoch):
 
             total_train_loss += loss.item()
             total_train_acc += self.metrics(
-                outputs[0], labels, 
-                self.configs.multi_label, 
-                self.configs.num_labels, 
-                self.device).item()
+                outputs[0],
+                labels,
+                self.configs.num_labels,
+                self.configs.multi_label,
+                self.device
+            ).item()
 
             batches.set_description(f"Train Loss Step: {loss.item():.2f}")
 
@@ -141,10 +143,12 @@ def validate_one_epoch(self, epoch):
 
         avg_val_loss = val_loss.mean().item()
         avg_val_acc = self.metrics(
-            val_outputs, val_targets, 
-            self.configs.multi_label, 
-            self.configs.num_labels, 
-            self.device).item()
+            val_outputs,
+            val_targets,
+            self.configs.num_labels,
+            self.configs.multi_label,
+            self.device
+        ).item()
 
         self.logger(epoch, avg_val_acc, avg_val_loss, 'test')
 
@@ -211,14 +215,14 @@ def save_checkpoint(self, epoch):
 
     def schedule_cold_start(self):
         self.scheduler = self.get_scheduler(
-            self.optimizer, 
-            self.configs.scheduler, 
-            0, 
-            len(self.train_dataloader)*self.configs.epochs)
+            self.optimizer,
+            self.configs.scheduler,
+            0,
+            len(self.train_dataloader) * self.configs.epochs)
 
     def print_per_epoch(self, epoch):
         print(f"\n\n{'-'*30}EPOCH {epoch}/{self.configs.epochs}{'-'*30}")
-        epoch -= 1 
+        epoch -= 1
         train_loss = self.train_logs[epoch][f'epoch_{epoch}']['loss']
         train_acc = self.train_logs[epoch][f'epoch_{epoch}']['accuracy']
         val_loss = self.val_logs[epoch][f'epoch_{epoch}']['loss']