Checking efficiency of dnadna
There has been concerned in the past concerning the efficiency of dnadna. See
@embray did a lot of profiling and made its best to make it fast, yet I'm still wondering why during training :
- GPU usage is low (every 10s the GPU goes to >50% but is 0% otherwise)
- It takes ~5s / batch no matter the speed of the hard drive
So maybe it's just the results of the size of our input data and the size of the network, but it might be worth trying to compare with a basic script, to make sure that dnadna did not introduce some slowdown.