eval limit as env and fix evaluate

app.py

@@ -16,7 +16,7 @@ from qwen_classifier.predict import predict_single  # Your existing function
 from qwen_classifier.evaluate import evaluate_batch  # Your existing function
 from qwen_classifier.globals import global_model, global_tokenizer
 from qwen_classifier.model import QwenClassifier
-from qwen_classifier.config import HF_REPO
+from qwen_classifier.config import HF_REPO, DEVICE
 
 
 app = FastAPI(title="Qwen Classifier")
@@ -70,7 +70,7 @@ async def load_model():
     
     model = QwenClassifier.from_pretrained(
         hf_repo,
-    )
+    ).to(DEVICE)
     global_tokenizer = AutoTokenizer.from_pretrained(hf_repo)
     print("Model loaded successfully!")
 

qwen_classifier/evaluate.py

@@ -15,6 +15,9 @@ from pathlib import Path
 from .config import TAG_NAMES, DEVICE, SPACE_URL, EVAL_LIMIT
 from .globals import global_model, global_tokenizer
 
+eval_limit = os.getenv("EVAL_LIM")
+if not eval_limit:
+    eval_limit == EVAL_LIMIT
 
 def _load_data(test_data_path):
     test_data_path = Path(__file__).parent / test_data_path
@@ -40,7 +43,7 @@ def _load_data(test_data_path):
                 raise ValueError("Empty zip archive - no files found")
                 
             # Process files with limit
-            for name in names[1:1+EVAL_LIMIT]:
+            for name in names[1:1+eval_limit]:
                 try:
                     with zip_file.open(name) as f:
                         content = f.read()
@@ -136,14 +139,14 @@ def _evaluate_local(test_data_path, hf_repo):
     with torch.no_grad():
         for batch in dataloader:
             print(f"EVALUATION RUNNING ON {global_model.device}")
-            batch = {k: v.to(global_model.device) for k, v in batch.items()}
+            batch = {k: v.to(DEVICE) for k, v in batch.items()}
             labels = batch["labels"].type(torch.float32)
 
             logits = global_model(batch["input_ids"], batch["attention_mask"])
 
             preds = torch.sigmoid(logits).cpu() > 0.5  # Keeps as PyTorch tensor
             preds = preds.float()  # Convert to 0.0/1.0 if needed
-            labels = labels.cpu().numpy()
+            labels = labels.cpu()
 
             all_preds.extend(preds)
             all_labels.extend(labels)
@@ -157,10 +160,10 @@ def _evaluate_local(test_data_path, hf_repo):
     val_f1_per_class = f1_score(all_labels, all_preds, average=None)
 
     metrics = {
-        'Accuracy':(100*val_acc).astype(int),
-        'Precision':(100*val_prec).astype(int),
-        'Recall':(100*val_rec).astype(int), 
-        'F1':(100*val_f1).astype(int),
+        'Accuracy':int(100*val_acc),
+        'Precision':int(100*val_prec),
+        'Recall':int(100*val_rec), 
+        'F1':int(100*val_f1),
         'Precision_per_class':(100*val_prec_per_class).astype(int),
         'Recall_per_class':(100*val_rec_per_class).astype(int),
         'F1_per_class':(100*val_f1_per_class).astype(int),