Set loss-scale to 1 to prevent using dynamic scaling

train.py

@@ -80,7 +80,8 @@
 parser.add_argument('--seed', default=123456, type=int, help='Seed to generators')
 parser.add_argument('--opt-level', type=str)
 parser.add_argument('--keep-batchnorm-fp32', type=str, default=None)
-parser.add_argument('--loss-scale', type=str, default=None)
+parser.add_argument('--loss-scale', default=1,
+ help='Loss scaling used by Apex. Default is 1 due to warp-ctc not supporting scaling of gradients')
 
 torch.manual_seed(123456)
 torch.cuda.manual_seed_all(123456)