Don't fix random seeds on import

spacy/cli/evaluate.py

@@ -3,8 +3,6 @@ from __future__ import unicode_literals, division, print_function
 
 import plac
 from timeit import default_timer as timer
-import random
-import numpy.random
 
 from ..gold import GoldCorpus
 from ..util import prints
@@ -12,10 +10,6 @@ from .. import util
 from .. import displacy
 
 
-random.seed(0)
-numpy.random.seed(0)
-
-
 @plac.annotations(
     model=("model name or path", "positional", None, str),
     data_path=("location of JSON-formatted evaluation data", "positional",
@@ -31,6 +25,8 @@ def evaluate(model, data_path, gpu_id=-1, gold_preproc=False, displacy_path=None
     Evaluate a model. To render a sample of parses in a HTML file, set an
     output directory as the displacy_path argument.
     """
+
+    util.fix_random_seed()
     if gpu_id >= 0:
         util.use_gpu(gpu_id)
     util.set_env_log(False)

spacy/cli/train.py

@@ -6,8 +6,6 @@ from pathlib import Path
 import tqdm
 from thinc.neural._classes.model import Model
 from timeit import default_timer as timer
-import random
-import numpy.random
 
 from ..gold import GoldCorpus, minibatch
 from ..util import prints
@@ -16,9 +14,6 @@ from .. import about
 from .. import displacy
 from ..compat import json_dumps
 
-random.seed(0)
-numpy.random.seed(0)
-
 
 @plac.annotations(
     lang=("model language", "positional", None, str),
@@ -45,6 +40,7 @@ def train(lang, output_dir, train_data, dev_data, n_iter=30, n_sents=0,
     """
     Train a model. Expects data in spaCy's JSON format.
     """
+    util.fix_random_seed()
     util.set_env_log(True)
     n_sents = n_sents or None
     output_path = util.ensure_path(output_dir)

spacy/util.py

@@ -17,6 +17,7 @@ from thinc.neural._classes.model import Model
 import functools
 import cytoolz
 import itertools
+import numpy as np
 
 from .symbols import ORTH
 from .compat import cupy, CudaStream, path2str, basestring_, input_, unicode_
@@ -623,3 +624,8 @@ def use_gpu(gpu_id):
     Model.ops = CupyOps()
     Model.Ops = CupyOps
     return device
+
+
+def fix_random_seed(seed=0):
+    random.seed(0)
+    np.random.seed(0)