pointerNetwork | custom layer in keras to implement a pointer | Machine Learning library

I want to implement Q&A systems with attention mechanism. I have two inputs; context and query which shapes are (batch_size, context_seq_len, embd_size) and (batch_size, query_seq_len, embd_size).
I am following the below paper. Machine Comprehension Using Match-LSTM and Answer Pointer. https://arxiv.org/abs/1608.07905

Then, I want to obtain a attention matrix which shape is (batch_size, context_seq_len, query_seq_len, embd_size). In the thesis, they calculate values for each row (it means each context word, G_i, alpha_i in the paper).

My code is below and it is running. But I am not sure my way is good or not. For example, I use for loop for generating sequence data (for i in range(T):). And to obtain each row, I use in-place operator like G[:,i,:,:], embd_context[:,i,:].clone() is a good manner in pytorch? If not, where should I change the code?

And if you notice other points, let me know. I am a new in this field and pytorch. Sorry for my ambiguous question.