Efficient GPU support for Entity Linking pipe

[svlandeg]

The new Entity Linking code has not been tested/adapted to GPU yet, cf Matt's comment here.

[zainbnv]

Yes I tried with GPU, but following error occurred:

020-03-05 03:50:46,682 - ERROR - main - Error updating batch:Unsupported type <class 'numpy.ndarray'> 2020-03-05 03:50:46,694 - ERROR - main - Error updating batch:Unsupported type <class 'numpy.ndarray'> 2020-03-05 03:50:46,696 - ERROR - main - Error updating batch:Unsupported type <class 'numpy.ndarray'> 2020-03-05 03:50:46,707 - ERROR - main - Error updating batch:Unsupported type <class 'numpy.ndarray'> 2020-03-05 03:50:46,714 - ERROR - main - Error updating batch:Unsupported type <class 'numpy.ndarray'> 2020-03-05 03:50:46,723 - ERROR - main - Error updating batch:Unsupported type <class 'numpy.ndarray'> Traceback (most recent call last): File "./all_wiki/wikidata_train_entity_linker.py", line 194, in plac.call(main) File "/home/administrator/anaconda3/lib/python3.7/site-packages/plac_core.py", line 328, in call cmd, result = parser.consume(arglist) File "/home/administrator/anaconda3/lib/python3.7/site-packages/plac_core.py", line 207, in consume return cmd, self.func(*(args + varargs + extraopts), **kwargs) File "./all_wiki/wikidata_train_entity_linker.py", line 146, in main logging.info("Epoch {}, train loss {}".format(itn, round(losses["entity_linker"] / batchnr, 2))) TypeError: type cupy.core.core.ndarray doesn't define round method

[svlandeg]

Update: the core EL code (i.e. not the wikipedia scripts which are outside the main library) has been tested and now runs on GPU (https://github.com/explosion/spaCy/pull/5236), but probably not very efficiently yet, so keeping this issue for now.

The wikipedia scripts is a separate issue - cf #5270. These scripts are not supported as part of the core library, as they really only showcase how the EL functionality can be used, but need adaptation to specific projects.