Use 0-vector for OOV in StaticVectors (#336)

* Use 0-vector for OOV in StaticVectors

* chore: update mac os images on azure (#325)

* Bump version for all macos azure pipelines

Co-authored-by: Justin DuJardin <[email protected]>

Author: adrianeboyd

azure-pipelines.yml

@@ -13,7 +13,7 @@ jobs:
         imageName: 'ubuntu-16.04'
         python.version: '2.7.17'
       Python27Mac:
-        imageName: 'macos-10.13'
+        imageName: 'macos-10.15'
         python.version: '2.7'
       Python35Linux:
         imageName: 'ubuntu-16.04'
@@ -22,7 +22,7 @@ jobs:
         imageName: 'vs2017-win2016'
         python.version: '3.5'
       Python35Mac:
-        imageName: 'macos-10.13'
+        imageName: 'macos-10.15'
         python.version: '3.5'
       Python36Linux:
         imageName: 'ubuntu-16.04'
@@ -31,7 +31,7 @@ jobs:
         imageName: 'vs2017-win2016'
         python.version: '3.6'
       Python36Mac:
-        imageName: 'macos-10.13'
+        imageName: 'macos-10.15'
         python.version: '3.6'
       Python37Linux:
         imageName: 'ubuntu-16.04'
@@ -40,7 +40,7 @@ jobs:
         imageName: 'vs2017-win2016'
         python.version: '3.7'
       Python37Mac:
-        imageName: 'macos-10.13'
+        imageName: 'macos-10.15'
         python.version: '3.7'
       Python38Linux:
         imageName: 'ubuntu-16.04'
@@ -49,7 +49,7 @@ jobs:
         imageName: 'vs2017-win2016'
         python.version: '3.8'
       Python38Mac:
-        imageName: 'macos-10.13'
+        imageName: 'macos-10.15'
         python.version: '3.8'
     maxParallel: 4
   pool:

thinc/about.py

@@ -4,7 +4,7 @@
 # https://github.com/pypa/warehouse/blob/master/warehouse/__about__.py
 
 __name__ = "thinc"
-__version__ = "7.4.0.dev2"
+__version__ = "7.4.1"
 __summary__ = "Practical Machine Learning for NLP"
 __uri__ = "https://github.com/explosion/thinc"
 __author__ = "Matthew Honnibal"

thinc/neural/_classes/static_vectors.py

@@ -58,7 +58,10 @@ def begin_update(self, ids, drop=0.0):
         if ids.ndim >= 2:
             ids = self.ops.xp.ascontiguousarray(ids[:, self.column])
         vector_table = self.get_vectors()
-        vectors = vector_table[ids * (ids < vector_table.shape[0])]
+        # create a mask for OOV vectors
+        non_oov = ids < vector_table.shape[0]
+        # initially use row 0 for OOV vectors
+        vectors = vector_table[ids * non_oov]
         vectors = self.ops.xp.ascontiguousarray(vectors)
         assert vectors.shape[0] == ids.shape[0]
 
@@ -74,4 +77,6 @@ def finish_update(gradients, sgd=None):
         mask = self.ops.get_dropout_mask((dotted.shape[1],), drop)
         if mask is not None:
             dotted *= mask
+        # replace OOV vectors with zeros
+        dotted *= self.ops.xp.zeros(dotted.shape, dtype=bool) | non_oov[:, None]
         return dotted, finish_update