Refactor non-parameter kernel function and enhance VAE training script

  - Renamed _non_para_kernel to _non_para_kernel_t4 for clarity
  - Added device selection for GPU/CPU in train_vae_nogcn function
  - Improved folder creation logic with warning for existing directories

Author: Starlitnightly

dynamo/external/celldancer/utilities.py

@@ -27,18 +27,18 @@ def __call__(self, *args, **kwargs):
         joblib.parallel.BatchCompletionCallBack = old_batch_callback
         tqdm_object.close()
 
-def _non_para_kernel(X,Y,down_sample_idx):
+def _non_para_kernel_t4(X,Y,down_sample_idx):
     # (no first cls),pseudotime r square calculation
     # this version has downsampling section
     # TO DO WHEN ONLY USING ONE GENE, WILL CAUSL PROBLEM WHEN COMBINING
     # Usage: Gene pseudotime fitting and r square (moved to utilities)
     # input: X,Y
     # return: estimator, r_square
-    # example: 
+    # example:
     # X = pd.DataFrame(np.arange(100)*np.pi/100)
     # Y = pd.DataFrame(np.sin(X)+np.random.normal(loc = 0, scale = 0.5, size = (100,1)))
     # estimator,r_square=non_para_kernel(X,Y)
-    
+
     # X2=pd.DataFrame(np.random.randint(0,100,size=[200,1]))
     # Y2=pd.DataFrame(np.random.normal(9,5,size=[200]))
     # X = pd.DataFrame(np.arange(100)*np.pi/100)

dynamo/external/latentvelo/trainer_nogcn.py

@@ -11,12 +11,15 @@
 
 def train_vae_nogcn(model, adata, epochs = 50, learning_rate = 1e-2, batch_size = 200, grad_clip = 1, shuffle=True, test=0.1, name = '', optimizer='adam', random_seed=42):
 
+    # Set device (GPU if available, otherwise CPU)
+    device = th.device('cuda' if th.cuda.is_available() else 'cpu')
+
     results_folder = './' + name + '/'
     if not os.path.exists(results_folder):
         os.mkdir(results_folder)
     else:
         print('Warning, folder already exists. This may overwrite a previous fit.')
-    
+
     if optimizer == 'adam':
         optimizer = th.optim.Adam(model.parameters(), lr = learning_rate)
     elif optimizer == 'adamW':