Work on serialization for models

spacy/_ml.py

@@ -1,3 +1,4 @@
+import ujson
 from thinc.api import add, layerize, chain, clone, concatenate, with_flatten
 from thinc.neural import Model, Maxout, Softmax, Affine
 from thinc.neural._classes.hash_embed import HashEmbed
@@ -15,7 +16,45 @@ from thinc.neural._classes.affine import _set_dimensions_if_needed
 from .attrs import ID, LOWER, PREFIX, SUFFIX, SHAPE, TAG, DEP
 from .tokens.doc import Doc
 
+import dill
 import numpy
+import io
+
+
+def model_to_bytes(model):
+    weights = []
+    metas = []
+    queue = [model]
+    i = 0
+    for layer in queue:
+        if hasattr(layer, '_mem'):
+            weights.append(layer._mem.weights)
+            metas.append(layer._mem._offsets)
+            i += 1
+        if hasattr(layer, '_layers'):
+            queue.extend(layer._layers)
+    data = {'metas': metas, 'weights': weights}
+    # TODO: Replace the pickle here with something else
+    return dill.dumps(data)
+
+
+def model_from_bytes(model, bytes_data):
+    # TODO: Replace the pickle here with something else
+    data = dill.loads(bytes_data)
+    metas = data['metas']
+    weights = data['weights']
+    queue = [model]
+    i = 0
+    for layer in queue:
+        if hasattr(layer, '_mem'):
+            params = weights[i]
+            flat_mem = layer._mem._mem.ravel()
+            flat_params = params.ravel()
+            flat_mem[:flat_params.size] = flat_params
+            layer._mem._offsets.update(metas[i])
+            i += 1
+        if hasattr(layer, '_layers'):
+            queue.extend(layer._layers)
  
 
 def _init_for_precomputed(W, ops):

spacy/pipeline.pyx

@@ -9,6 +9,7 @@ import numpy
 cimport numpy as np
 import cytoolz
 import util
+import ujson
 
 from thinc.api import add, layerize, chain, clone, concatenate, with_flatten
 from thinc.neural import Model, Maxout, Softmax, Affine
@@ -35,6 +36,7 @@ from .syntax import nonproj
 
 from .attrs import ID, LOWER, PREFIX, SUFFIX, SHAPE, TAG, DEP, POS
 from ._ml import rebatch, Tok2Vec, flatten, get_col, doc2feats
+from ._ml import model_to_bytes, model_from_bytes
 from .parts_of_speech import X
 
 
@@ -148,7 +150,6 @@ class TokenVectorEncoder(object):
         if self.model is True:
             self.model = self.Model()
 
-
     def use_params(self, params):
         """Replace weights of models in the pipeline with those provided in the
         params dictionary.
@@ -158,6 +159,39 @@ class TokenVectorEncoder(object):
         with self.model.use_params(params):
             yield
 
+    def to_bytes(self, **exclude):
+        data = {
+            'model': self.model,
+            'vocab': self.vocab
+        }
+        return util.to_bytes(data, exclude)
+
+    def from_bytes(self, bytes_data, **exclude):
+        data = ujson.loads(bytes_data)
+        if 'model' not in exclude:
+            util.model_from_bytes(self.model, data['model'])
+        if 'vocab' not in exclude:
+            self.vocab.from_bytes(data['vocab'])
+        return self
+
+    def to_disk(self, path, **exclude):
+        path = util.ensure_path(path)
+        if not path.exists():
+            path.mkdir()
+        if 'vocab' not in exclude:
+            self.vocab.to_disk(path / 'vocab')
+        if 'model' not in exclude:
+            with (path / 'model.bin').open('wb') as file_:
+                file_.write(util.model_to_bytes(self.model))
+
+    def from_disk(self, path, **exclude):
+        path = util.ensure_path(path)
+        if 'vocab' not in exclude:
+            self.vocab.from_disk(path / 'vocab')
+        if 'model.bin' not in exclude:
+            with (path / 'model.bin').open('rb') as file_:
+                util.model_from_bytes(self.model, file_.read())
+
 
 class NeuralTagger(object):
     name = 'nn_tagger'

spacy/tests/test_misc.py

@@ -2,12 +2,38 @@
 from __future__ import unicode_literals
 
 from ..util import ensure_path
+from .._ml import model_to_bytes, model_from_bytes
 
 from pathlib import Path
 import pytest
+from thinc.neural import Maxout, Softmax
+from thinc.api import chain
 
 
 @pytest.mark.parametrize('text', ['hello/world', 'hello world'])
 def test_util_ensure_path_succeeds(text):
     path = ensure_path(text)
     assert isinstance(path, Path)
+
+
+def test_simple_model_roundtrip_bytes():
+    model = Maxout(5, 10, pieces=2)
+    model.b += 1
+    data = model_to_bytes(model)
+    model.b -= 1
+    model_from_bytes(model, data)
+    assert model.b[0, 0] == 1
+
+
+def test_multi_model_roundtrip_bytes():
+    model = chain(Maxout(5, 10, pieces=2), Maxout(2, 3))
+    model._layers[0].b += 1
+    model._layers[1].b += 2
+    data = model_to_bytes(model)
+    model._layers[0].b -= 1
+    model._layers[1].b -= 2
+    model_from_bytes(model, data)
+    assert model._layers[0].b[0, 0] == 1
+    assert model._layers[1].b[0, 0] == 2
+
+

spacy/util.py

@@ -408,6 +408,18 @@ def get_raw_input(description, default=False):
     return user_input
 
 
+def to_bytes(unserialized, exclude):
+    serialized = {}
+    for key, value in unserialized.items():
+        if key in exclude:
+            continue
+        elif hasattr(value, 'to_bytes'):
+            serialized[key] = value.to_bytes()
+        else:
+            serialized[key] = ujson.dumps(value)
+    return ujson.dumps(serialized)
+
+
 def print_table(data, title=None):
     """Print data in table format.
 

spacy/vocab.pyx

@@ -56,15 +56,7 @@ cdef class Vocab:
         if strings:
             for string in strings:
                 self.strings.add(string)
-        # Load strings in a special order, so that we have an onset number for
+        for name in tag_map.keys():
-        # the vocabulary. This way, when words are added in order, the orth ID
-        # is the frequency rank of the word, plus a certain offset. The structural
-        # strings are loaded first, because the vocab is open-class, and these
-        # symbols are closed class.
-        # TODO: Actually this has turned out to be a pain in the ass...
-        # It means the data is invalidated when we add a symbol :(
-        # Need to rethink this.
-        for name in symbols.NAMES + list(sorted(tag_map.keys())):
             if name:
                 self.strings.add(name)
         self.lex_attr_getters = lex_attr_getters