How do we implement BPTT from scratch using python for a simple RNN

**y(t+1)=0.5y(t)+0.5x(t)**

Generally, we donâ€™t implement BPTT (Back-Propagation Through Time) from scratch since it can be a little overwhelming to calculate all the gradients through all time-steps.

However, if you want to know how, check this:

**Building RNN BPTT from scratch using numpy**

Or if you just want to learn how to write an RNN from scratch without using PyTorchâ€™s `nn.RNN()`

and perform gradient descent without using `optimizer.step()`

, you can check this:

NLP from scratch: Classifying names with Character-level RNN

1 Like

yaa thanks a lotâ€¦But could you explain what those deltas are in the below screenshot â€¦Remaining things i was able to understandâ€¦

You mean this? â‰ś

It just means that `'LHS symbol/function' is defined by 'RHS expression'`

.