Merge pull request #398 from PNRxA/dev

numpy to use CPU when using CUDA

utils/generic_utils.py

@@ -157,7 +157,7 @@ def check_update(model, grad_clip, ignore_stopnet=False):
         grad_norm = torch.nn.utils.clip_grad_norm_([param for name, param in model.named_parameters() if 'stopnet' not in name], grad_clip)
     else:
         grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
-    if np.isinf(grad_norm):
+    if np.isinf(grad_norm.cpu()):
         print(" | > Gradient is INF !!")
         skip_flag = True
     return grad_norm, skip_flag