v2.0.0, update Autograd
- Now Autograd is tape-based, (previously it was fully recursive)
- Bug fixes and optimization
- Simplify
nn.Embedding - Simplify autograd
reverse_broadcastmethod - Fix
wheregrad_fn method
- Simplify
- Add Seq2Seq Transformer example
- Other
- Add
nn.LogSoftmax - Edit
nn.Linear bias initialization - Оther fixes and improvements
- Add
New autograd version enables transformer training, overcoming previous limitations due to slow backpropagation speed.