Increasing the number of worker threads/core to be used about tagger HOT 3 CLOSED

@kewlcoder : As tensorflow/theano computes everything on the symbolic graph, each time the graph is fed with data, the calculation has been done on gpu/cpu. For smaller batch, transferring time of data to gpu/cpu bus bottleneck the overall computation time. So if you have smaller batch single thread will be best. Even sometimes codes run faster on CPU than a titan X gpu for this reason.

from tagger.

Hi,
Actually in my experience, using a single thread was always faster than using multi-threading. The reason for this is that the tagger is not batched, I guess.

from tagger.

Yes for me the tagger was faster on CPU than on GPU (I had K40 at the time). The tagger could definitely be batched, but I did not do it as the datasets were small enough and training can already be done in a few hours.

from tagger.