Vertically align the norm outputs for the log_norms in the constituen…

…cy parser --log_norms. Also, count the zeros (or close to zero) terms from the tensors as a signal of how many neurons are dying

stanza/models/constituency/lstm_model.py

@@ -667,9 +667,15 @@ def get_norms(self):
             lines.append("reduce_linear:")
             for c_idx, c_open in enumerate(self.constituent_opens):
                 lines.append("  %s weight %.6g bias %.6g" % (c_open, torch.norm(self.reduce_linear_weight[c_idx]).item(), torch.norm(self.reduce_linear_bias[c_idx]).item()))
+        max_name_len = max(len(name) for name, param in self.named_parameters() if param.requires_grad and name not in skip)
+        max_norm_len = max(len("%.6g" % torch.norm(param).item()) for name, param in self.named_parameters() if param.requires_grad and name not in skip)
+        format_string = "%-" + str(max_name_len) + "s   norm %" + str(max_norm_len) + "s  zeros %d / %d"
+        print(format_string)
         for name, param in self.named_parameters():
             if param.requires_grad and name not in skip:
-                lines.append("%s %.6g" % (name, torch.norm(param).item()))
+                zeros = torch.sum(param.abs() < 0.000001).item()
+                norm = "%.6g" % torch.norm(param).item()
+                lines.append(format_string % (name, norm, zeros, param.nelement()))
         return lines
 
     def log_norms(self):