locally tested

tasks/audio.py

@@ -10,13 +10,21 @@ from torch.utils.data import DataLoader
 from .utils.evaluation import AudioEvaluationRequest
 from .utils.emissions import tracker, clean_emissions_data, get_space_info
 from .utils.data import FFTDataset
-from .utils.models import DualEncoder, CNNKan
+from .utils.models import DualEncoder, CNNKan, CNNKanFeaturesEncoder
 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
+import xgboost as xgb
+from tqdm import tqdm
+from sklearn.metrics import accuracy_score, classification_report, roc_auc_score
+from sklearn.model_selection import train_test_split
+import warnings
+import pandas as pd
+
+warnings.filterwarnings("ignore")
 
 
 
@@ -28,6 +36,28 @@ router = APIRouter()
 DESCRIPTION = "Conformer"
 ROUTE = "/audio"
 
+def create_dataframe(ds, save_name='test'):
+    data = []
+    # Iterate over the dataset
+    pbar = tqdm(enumerate(ds))
+    for i, batch in pbar:
+        label = batch['label']
+        features = batch['audio']['features']
+
+        # Flatten the nested dictionary structure
+        feature_dict = {'label': label}
+        for k, v in features.items():
+            if isinstance(v, dict):
+                for sub_k, sub_v in v.items():
+                    feature_dict[f"{k}_{sub_k}"] = sub_v[0].item()  # Aggregate (e.g., mean)
+        data.append(feature_dict)
+    # Convert to DataFrame
+    df = pd.DataFrame(data)
+    print(os.getcwd())
+    df.to_csv(f"tasks/utils/dfs/{save_name}.csv", index=False)
+    X = df.drop(columns=['label'])
+    y = df['label']
+    return X, y
 
 @router.post(ROUTE, tags=["Audio Task"],
              description=DESCRIPTION)
@@ -69,10 +99,13 @@ async def evaluate_audio(request: AudioEvaluationRequest):
     model_args = Container(**yaml.safe_load(open(args_path, 'r'))['CNNEncoder'])
     model_args_f = Container(**yaml.safe_load(open(args_path, 'r'))['CNNEncoder_f'])
     conformer_args = Container(**yaml.safe_load(open(args_path, 'r'))['Conformer'])
+    boost_args = Container(**yaml.safe_load(open(args_path, 'r'))['XGBoost'])
     kan_args = Container(**yaml.safe_load(open(args_path, 'r'))['KAN'])
 
-    test_dataset = FFTDataset(test_dataset)
-    test_dl = DataLoader(test_dataset, batch_size=data_args.batch_size, collate_fn=collate_fn)
+    test_dataset = FFTDataset(test_dataset, features=False)
+    test_dl = DataLoader(test_dataset, batch_size=data_args.batch_size)
+
+    # Watchlist to monitor performance on train and validation data
 
     model = CNNKan(model_args, conformer_args, kan_args.get_dict())
     model = model.to(device)

tasks/models/frugal_2025-02-01/frugal_kan_features_2.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d3fbc9f7a73a40a99863fbbf70e244598d2594e451f01737812b553e354541c2
-size 614523
+oid sha256:f6e20edd364e79003a08cfd4221ec8fc312c16898b1e4871159a8fd1864b791e
+size 1876605

tasks/run.py

@@ -172,8 +172,8 @@ print(num_xgb_features)
 # model = DualEncoder(model_args, model_args_f, conformer_args)
 # model = FasterKAN([18000,64,64,16,1])
 # model = CNNKan(model_args, conformer_args, kan_args.get_dict())
-# model = CNNKanFeaturesEncoder(xgb_model, model_args, kan_args.get_dict())
-model = CNNFeaturesEncoder(xgb_model,model_args)
+model = CNNKanFeaturesEncoder(xgb_model, model_args, kan_args.get_dict())
+# model = CNNFeaturesEncoder(xgb_model,model_args)
 # model.kan.speed()
 # model = KanEncoder(kan_args.get_dict())
 model = model.to(local_rank)

tasks/utils/config.yaml

@@ -5,7 +5,7 @@ Data:
   dataset: "FFTDataset"
   data_dir: None
   model_name: "CNNEncoder"
-  batch_size: 32
+  batch_size: 4
   num_epochs: 10
   exp_num: 2
   max_len_spectra: 4096

tasks/utils/train.py

@@ -226,14 +226,14 @@ class Trainer(object):
     
     def train_batch(self, batch, batch_idx, device):
         x, fft, y = batch['audio']['array'], batch['audio']['fft_mag'], batch['label']
-        features = batch['audio']['features']
+        # features = batch['audio']['features']
         # cwt = batch['audio']['cwt_mag']
         x = x.to(device).float()
         fft = fft.to(device).float()
         # cwt = cwt.to(device).float()
         y = y.to(device).float()
         x_fft = torch.cat((x.unsqueeze(dim=1), fft.unsqueeze(dim=1)), dim=1)
-        y_pred = self.model(x_fft, features).squeeze()
+        y_pred = self.model(x_fft).squeeze()
         loss = self.criterion(y_pred, y)
         loss.backward()
         self.optimizer.step()
@@ -267,7 +267,7 @@ class Trainer(object):
 
     def eval_batch(self, batch, batch_idx, device):
         x, fft, y = batch['audio']['array'], batch['audio']['fft_mag'], batch['label']
-        features = batch['audio']['features']
+        # features = batch['audio']['features']
 
         # features = batch['audio']['features_arr'].to(device).float()
         x = x.to(device).float()
@@ -275,7 +275,7 @@ class Trainer(object):
         x_fft = torch.cat((x.unsqueeze(dim=1), fft.unsqueeze(dim=1)), dim=1)
         y = y.to(device).float()
         with torch.no_grad():
-            y_pred = self.model(x_fft, features).squeeze()
+            y_pred = self.model(x_fft).squeeze()
         loss = self.criterion(y_pred.squeeze(), y)
         probs = torch.sigmoid(y_pred)
         cls_pred = (probs > 0.5).float()