zhuohan123 / terapipe Goto Github PK

Dear Sir,

I notice that in your paper, there is an equation:

t_{fwd} (i, j) = t_{fwd}(i, 0) + t_{ctx}(i, j),

Let j=0, the second term will become zero.

But then we will have t_{ctx}(i, 0) = a_0 + a_1*i =0, which means that we don't need a_0 and a_1 at all. Or it means a_0, a_1 and i have a relation while a_0 and a_1 are constant and i is variable. This is wrong unless a_0 and a_1 are zero.

I am not sure if there is any restrictions on using this formula.

thanks for sharing the code.

i have a question about the attention module in "transformer_models.py"

in your code , assume that there are 10 tokens in setence, i think the 7th token can only calculate with 1-6th tokens, as writen in your paper.

but when i read the GPT2 paper and , then read "GENERATING WIKIPEDIA BY SUMMARIZING LONG SEQUENCES" find the attention module is not calculate in that way as your paper.

Where can I find relevant information about the GPT2's attention calculation method？ i only find the attention with left and right tokens.

thank you

I notice that in this repository, it's still using the uniform way to split the input sequence.

Could you please tell us when you will update the Dynamic Programming Version as mentioned in the paper?