If I fine-tune run_seq2seq.py with a specific model from the hub and a specific dataset in 4 epochs it finishes correctly.
I save a checkpoint at the third epoch.
But if I load the checkpoint at the third epoch and try to continue until the forth I get an out of memory error.
Is it necessary to make adjustments in the arguments of run_seq2seq.py, when starting from a checkpoint, to obtain similar results?