update requirements

tasks/run.py

@@ -91,8 +91,8 @@ test_dl = DataLoader(test_ds,batch_size=data_args.batch_size, collate_fn=collate
 
 # 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(model_args, mlp_args, kan_args.get_dict())
+# model = CNNKan(model_args, conformer_args, kan_args.get_dict())
+model = CNNKanFeaturesEncoder(model_args, mlp_args, kan_args.get_dict())
 # model.kan.speed()
 # model = KanEncoder(kan_args.get_dict())
 model = model.to(local_rank)
@@ -127,7 +127,7 @@ trainer = Trainer(model=model, optimizer=optimizer,
                            exp_num=datetime_dir, log_path=data_args.log_dir,
                             range_update=None,
                            accumulation_step=1, max_iter=np.inf,
-                           exp_name=f"frugal_kan_{exp_num}")
+                           exp_name=f"frugal_kan_features_{exp_num}")
 fit_res = trainer.fit(num_epochs=100, device=local_rank,
                         early_stopping=10, only_p=False, best='loss', conf=True)
 output_filename = f'{data_args.log_dir}/{datetime_dir}/{model_name}_frugal_{exp_num}.json'

tasks/utils/config.yaml

@@ -32,7 +32,7 @@ CNNEncoder:
 
 MLP:
   input_dim: 6
-  hidden_dims: [16,32]
+  hidden_dims: [16,32,6]
   dropout: 0.2
 
 KAN:

tasks/utils/data.py

@@ -89,8 +89,8 @@ class FFTDataset(IterableDataset):
             fft_data = fft(audio_data)
             magnitude = torch.abs(fft_data)
             phase = torch.angle(fft_data)
-            # features = compute_all_features(audio_data, sample_rate=self.target_sample_rate)
-            # features_arr = torch.tensor([v for _, v in features['frequency_domain'].items()])
+            features = compute_all_features(audio_data, sample_rate=self.target_sample_rate)
+            features_arr = torch.tensor([v for _, v in features['frequency_domain'].items()])
             magnitude_centered = fftshift(magnitude)
             phase_centered = fftshift(phase)
             # cwt = features['cwt_power']
@@ -103,7 +103,7 @@ class FFTDataset(IterableDataset):
             # item['audio']['cwt_mag'] = torch.nan_to_num(cwt, 0)
             item['audio']['array'] = torch.nan_to_num(audio_data, 0)
             # item['audio']['features'] = features
-            # item['audio']['features_arr'] = torch.nan_to_num(features_arr, 0)
+            item['audio']['features_arr'] = torch.nan_to_num(features_arr, 0)
             yield item