towhee.functional.mixins.compile.CompileMixin

class towhee.functional.mixins.compile.CompileMixin

Bases: object

Mixin to just-in-time compile the Operator. More information about just-in-time compilation can be found at https://en.wikipedia.org/wiki/Just-in-time_compilation.

Examples:

1. compile the python function with numba:

>>> import numpy
>>> import towhee
>>> import time
>>> from towhee import register
>>> @register(name='inner_distance')
... def inner_distance(query, data):
...     dists = []
...     for vec in data:
...         dist = 0
...         for i in range(len(vec)):
...             dist += vec[i] * query[i]
...         dists.append(dist)
...     return dists
>>> data = [numpy.random.random((10000, 128)) for _ in range(10)]
>>> query = numpy.random.random(128)

>>> t1 = time.time()
>>> dc1 = (
...     towhee.dc['a'](data)
...     .runas_op['a', 'b'](func=lambda _: query)
...     .inner_distance[('b', 'a'), 'c']()
... )
>>> t2 = time.time()
>>> dc2 = (
...     towhee.dc['a'](data)
...     .config(jit='numba')
...     .runas_op['a', 'b'](func=lambda _: query)
...     .inner_distance[('b', 'a'), 'c']()
... )
>>> t3 = time.time()
>>> assert(t3-t2 < t2-t1)

2. compile the models with towhee.compiler:

Note

It will take some time to compile in the first time.

More information about towhee.compiler refer to https://github.com/towhee-io/towhee-compiler.

>>> import towhee
>>> dc1 = (towhee.dc(['hello world']) 
...     .set_jit('towhee')
...     .text_embedding.transformers(model_name='distilbert-base-cased')
... )

Methods

jit_resolve
set_jit

__init__() → None