paths fix

tasks/audio.py

@@ -14,6 +14,11 @@ from .utils.models import DualEncoder
 from .utils.train import Trainer
 from .utils.data_utils import collate_fn, Container
 import yaml
+import asyncio
+from huggingface_hub import login
+from collections import OrderedDict
+
+
 
 from dotenv import load_dotenv
 load_dotenv()
@@ -60,7 +65,7 @@ async def evaluate_audio(request: AudioEvaluationRequest):
     # Update the code below to replace the random baseline by your model inference within the inference pass where the energy consumption and emissions are tracked.
     #--------------------------------------------------------------------------------------------
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    args_path = 'utils/config.yaml'
+    args_path = 'tasks/utils/config.yaml'
     data_args = Container(**yaml.safe_load(open(args_path, 'r'))['Data'])
     model_args = Container(**yaml.safe_load(open(args_path, 'r'))['CNNEncoder'])
     model_args_f = Container(**yaml.safe_load(open(args_path, 'r'))['CNNEncoder_f'])
@@ -71,7 +76,13 @@ async def evaluate_audio(request: AudioEvaluationRequest):
 
     model = DualEncoder(model_args, model_args_f, conformer_args)
     model = model.to(device)
-    missing, unexpected = model.load_state_dict(torch.load(model_args.checkpoint_path))
+    state_dict = torch.load(model_args.checkpoint_path)
+    new_state_dict = OrderedDict()
+    for key, value in state_dict.items():
+        if key.startswith('module.'):
+            key = key[7:]
+        new_state_dict[key] = value
+    missing, unexpected = model.load_state_dict(new_state_dict)
 
     loss_fn = torch.nn.BCEWithLogitsLoss()
     optimizer = torch.optim.Adam(model.parameters(), lr=5e-4)
@@ -83,9 +94,9 @@ async def evaluate_audio(request: AudioEvaluationRequest):
                       range_update=None,
                       accumulation_step=1, max_iter=np.inf,
                       exp_name=f"frugal_cnnencoder_inference")
-    predictions, acc = trainer.predict(test_dl, device=device)
+    predictions, true_labels, acc = trainer.predict(test_dl, device=device)
     # Make random predictions (placeholder for actual model inference)
-    true_labels = test_dataset["label"]
+    print("accuracy: ", acc)
 
     #--------------------------------------------------------------------------------------------
     # YOUR MODEL INFERENCE STOPS HERE
@@ -115,4 +126,17 @@ async def evaluate_audio(request: AudioEvaluationRequest):
         }
     }
     
-    return results 
+    return results
+
+# if __name__ == "__main__":
+#     with open("../logs//token.txt", "r") as f:
+#         api_key = f.read()
+#     login(api_key)
+#     # Create a sample request object
+#     sample_request = AudioEvaluationRequest(
+#         dataset_name="rfcx/frugalai",  # Replace with actual dataset name
+#         test_size=0.2,  # Example values
+#         test_seed=42
+#     )
+#
+#     asyncio.run(evaluate_audio(sample_request))

tasks/utils/config.yaml

@@ -28,7 +28,7 @@ CNNEncoder:
   activation: "silu"
   sine_w0: 1.0
   avg_output: True
-  checkpoint_path: 'models/frugal_2025-01-10/frugal_cnnencoder_2.pth'
+  checkpoint_path: 'tasks/models/frugal_2025-01-10/frugal_cnnencoder_2.pth'
 
 CNNEncoder_f:
   # Model

tasks/utils/train.py

@@ -274,7 +274,8 @@ class Trainer(object):
         total = 0
         all_accs = 0
         predictions = []
-        pbar = tqdm(self.val_dl)
+        true_labels = []
+        pbar = tqdm(test_dataloader)
         for i,batch in enumerate(pbar):
             x, fft, y = batch['audio']['array'], batch['audio']['fft'], batch['label']
             x = x.to(device).float()
@@ -286,7 +287,11 @@ class Trainer(object):
             probs = torch.sigmoid(y_pred)
             cls_pred = (probs > 0.5).float()
             acc = (cls_pred == y).sum()
-            predictions.append(cls_pred)
+            predictions.append(cls_pred.cpu().numpy())
+            true_labels.append(y.cpu().numpy())
             all_accs += acc
             total += len(y)
-        return predictions, all_accs/total
+            pbar.set_description("acc: {:.4f}".format(acc))
+            if i > self.max_iter:
+                break
+        return predictions, true_labels, all_accs/total