gradient of bias is nan at first iteration for 'char-rnn-classification'

[yinhang-e5b0b9e888aa]

torch.version is '0.4.0a0+60a16e5'

Some changes of pytorch has broken the code in tutorial?

10000 10% (0m 14s) nan Auttenberg / Polish ✗ (German)
15000 15% (0m 21s) nan Demas / Polish ✗ (Greek)
20000 20% (0m 28s) nan Leveque / Polish ✗ (French)
25000 25% (0m 35s) nan Vale / Polish ✗ (English)
30000 30% (0m 42s) nan Savchak / Polish ✗ (Czech)
35000 35% (0m 49s) nan Ton / Polish ✗ (Vietnamese)
40000 40% (0m 56s) nan Franco / Polish ✗ (Portuguese)
45000 45% (1m 4s) nan Langbroek / Polish ✗ (Dutch)
50000 50% (1m 11s) nan Chu / Polish ✗ (Korean)
55000 55% (1m 19s) nan Egr / Polish ✗ (Czech)
60000 60% (1m 26s) nan Maloof / Polish ✗ (Arabic)
65000 65% (1m 33s) nan Morra / Polish ✗ (Italian)
70000 70% (1m 40s) nan Zellweger / Polish ✗ (German)
75000 75% (1m 47s) nan Suarez / Polish ✗ (Spanish)
80000 80% (1m 54s) nan Assaf / Polish ✗ (Arabic)
85000 85% (2m 1s) nan Morimoto / Polish ✗ (Japanese)
90000 90% (2m 8s) nan Jez / Polish ✓
95000 95% (2m 15s) nan Borovsky / Polish ✗ (Czech)
100000 100% (2m 22s) nan Sugawara / Polish ✗ (Japanese)