Hello.
In the training phase, there's reset_mems for resetting memory.
When the new music (sequence) begins, it blocks attention to memory of previous music by setting reset_mems=True.
https://github.com/amazon-research/transformer-gan/blob/1ccc9f251c1b1d054c1acc8be36c1da7bf8cf11c/model/mem_transformer.py#L529
However, since memory length (1024) is longer than batch length (128), new music sequence can do attention to previous music after one block.
So, it seems that new sequence can do attention to 'prev sequence memory (896) + cur sequece memory (128)'.
I am wondering if this is intentional.
Thank you for sharing code!
Hello.
In the training phase, there's reset_mems for resetting memory.
When the new music (sequence) begins, it blocks attention to memory of previous music by setting reset_mems=True.
https://github.com/amazon-research/transformer-gan/blob/1ccc9f251c1b1d054c1acc8be36c1da7bf8cf11c/model/mem_transformer.py#L529
However, since memory length (1024) is longer than batch length (128), new music sequence can do attention to previous music after one block.
So, it seems that new sequence can do attention to 'prev sequence memory (896) + cur sequece memory (128)'.
I am wondering if this is intentional.
Thank you for sharing code!