fix(flow_de): remove dead code and fix stale NaN debug prints

Physics_Informed_Neural_Network_Diffusion_Equation_Sijil_Jose/flow_de/flow_de.py

@@ -93,45 +93,10 @@ def forward(self, t, xt):
             print('  qVectorField: vt.shape', vt.shape)
             print('  qVectorField: vt', vt)
 
-        # sum over arguments of exponential, that is,
-        # over the d-dimensions of each element in x0,
-        # so that we get the product of d normal densities.
-
-        '''
-        v2 = (vt*vt).sum(dim=-1)
-        vv = torch.where(v2 < 207, v2, 207)        ## why not use the logsumexp for greater stability
-        if torch.isnan(vv).any():
-            raise ValueError("vv contains at least one NAN")
-
-        if debug:
-            print('  qVectorField: vv.shape', vv.shape)
-
-        # compute unnormalized probability densities.
-        pt = torch.exp(-vv/2)
-        if torch.isnan(pt).any():
-            raise ValueError("pt contains at least one NAN")
-
-        # pt.shape: (N, M)
-        if debug:
-            print('  qVectorField: pt.shape', pt.shape)
-
-        # sum over the M d-dimensional Gaussian densities
-        ptsum = pt.sum(dim=-1)
-        # ptsum.shape: (N, )
-        if torch.isnan(ptsum).any():
-            raise ValueError("ptsum contains at least one NAN")
-
-        # protect sum against divide by zero
-        pt_sum = torch.where(ptsum < 1.e-44, 1, ptsum).unsqueeze(-1)
-        # pt_sum.shape: (N, 1)
-        if torch.isnan(ptsum).any():
-            raise ValueError("ptsum contains at least one NAN")
-
-        # compute weights
-        wt = pt / pt_sum
-        '''
-
-
+        # Compute squared distances from each collocation point xt to every
+        # MC sample x0, summed over the d feature dimensions.
+        # Uses the log-sum-exp trick for numerical stability (avoids exp overflow
+        # that occurred in the old implementation which used torch.where clamping).
         v2 = (vt * vt).sum(dim=-1)
         log_weights = -0.5 * v2
         log_weights = log_weights - torch.logsumexp(log_weights, dim=-1, keepdim=True)
@@ -140,11 +105,11 @@ def forward(self, t, xt):
 
         # wt.shape: (N, M)
         if torch.isnan(wt).any():
-            print('pt')
-            print(pt)
+            print('v2')
+            print(v2)
             print()
-            print('pt_sum')
-            print(pt_sum)
+            print('log_weights')
+            print(log_weights)
             print()
             print('wt')
             print(wt)