NER training is very slow

[damianoporta]

Hello,
i am training a new NER model with this code: https:/explosion/spaCy/blob/master/examples/training/train_new_entity_type.py
i have noticed that the training is very very slow.
Doing a test with 2500 documents (500-800 tokens long) i only can see few iterations after 8 hours.

Ok, i am running it via CPU but my pc is good.

• Intel® Core™ i7-6700K CPU @ 4.00GHz × 8
• GeForce GTX 1070/PCIe/SSE2
• 32 gb ram
• SSD
  (i can try with GPU, is now possible to enable it?)
  is really the training of a new model that slow? I have around 50k documents, it will took forever.
  Can i optimize it somehow?

Thanks!

Your Environment

Platform           Linux-4.4.0-112-generic-x86_64-with-Ubuntu-16.04-xenial
Models             en, it, en_core_web_md, en_core_web_sm
spaCy version      2.0.6.dev0     
Location           /home/damiano/lavoro/python/parser/.env/lib/python3.5/site-packages/spacy
Python version     3.5.2  

[r-wheeler]

Re: how to enable a gpu, previous thread here

[honnibal]

That code doesn't use minibatching, in order to keep the example simple. For larger training tasks, try using the spacy train command.

[damianoporta]

@honnibal Ok, can I convert the TRAIN_DATA in the spacy's json format somehow?

[honnibal]

The easiest way at the moment would be to put the data into BILUO format and use the spacy convert command.

[damianoporta]

@honnibal i am testing the same code with minibatches and it works good now (with CPU).

[honnibal]

Great!

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.