batch_vals.append(vals)
# The dtype here matches what thinc is expecting -- which differs per
# platform (by int definition). This should be fixed once the problem
# is fixed on Thinc's side.
lengths = self.ops.asarray([arr.shape[0] for arr in batch_keys],
dtype=numpy.int_)
batch_keys = self.ops.xp.concatenate(batch_keys)
batch_vals = self.ops.asarray(self.ops.xp.concatenate(batch_vals),
dtype="f")
return (batch_keys, batch_vals, lengths), None
@describe.on_data(_set_dimensions_if_needed,
lambda model, X, y: model.init_weights(model))
@describe.attributes(
nI=Dimension("Input size"),
nF=Dimension("Number of features"),
nO=Dimension("Output size"),
nP=Dimension("Maxout pieces"),
W=Synapses("Weights matrix", lambda obj: (obj.nF, obj.nO, obj.nP, obj.nI)),
b=Biases("Bias vector", lambda obj: (obj.nO, obj.nP)),
pad=Synapses(
"Pad",
lambda obj: (1, obj.nF, obj.nO, obj.nP),
lambda M, ops: ops.normal_init(M, 1.0),
),
d_W=Gradient("W"),
d_pad=Gradient("pad"),
d_b=Gradient("b"),
)
class PrecomputableAffine(Model):
bigrams = [ops.ngrams(2, doc_unis) for doc_unis in unigrams]
keys = [ops.xp.concatenate(feats) for feats in zip(unigrams, bigrams)]
keys, vals = zip(*[ops.xp.unique(k, return_counts=True) for k in keys])
# The dtype here matches what thinc is expecting -- which differs per
# platform (by int definition). This should be fixed once the problem
# is fixed on Thinc's side.
lengths = ops.asarray([arr.shape[0] for arr in keys], dtype=numpy.int_)
keys = ops.xp.concatenate(keys)
vals = ops.asarray(ops.xp.concatenate(vals), dtype='f')
return (keys, vals, lengths), None
@describe.on_data(_set_dimensions_if_needed,
lambda model, X, y: model.init_weights(model))
@describe.attributes(
nI=Dimension("Input size"),
nF=Dimension("Number of features"),
nO=Dimension("Output size"),
nP=Dimension("Maxout pieces"),
W=Synapses("Weights matrix",
lambda obj: (obj.nF, obj.nO, obj.nP, obj.nI)),
b=Biases("Bias vector",
lambda obj: (obj.nO, obj.nP)),
pad=Synapses("Pad",
lambda obj: (1, obj.nF, obj.nO, obj.nP),
lambda M, ops: ops.normal_init(M, 1.)),
d_W=Gradient("W"),
d_pad=Gradient("pad"),
d_b=Gradient("b"))
class PrecomputableAffine(Model):
def __init__(self, nO=None, nI=None, nF=None, nP=None, **kwargs):
@layerize
def _preprocess_doc(docs, drop=0.):
keys = [doc.to_array([LOWER]) for doc in docs]
ops = Model.ops
# The dtype here matches what thinc is expecting -- which differs per
# platform (by int definition). This should be fixed once the problem
# is fixed on Thinc's side.
lengths = ops.asarray([arr.shape[0] for arr in keys], dtype=numpy.int_)
keys = ops.xp.concatenate(keys)
vals = ops.allocate(keys.shape[0]) + 1
return (keys, vals, lengths), None
@describe.on_data(_set_dimensions_if_needed,
lambda model, X, y: model.init_weights(model))
@describe.attributes(nI=Dimension("Input size"),
nF=Dimension("Number of features"),
nO=Dimension("Output size"),
nP=Dimension("Maxout pieces"),
W=Synapses("Weights matrix", lambda obj:
(obj.nF, obj.nO, obj.nP, obj.nI)),
b=Biases("Bias vector", lambda obj: (obj.nO, obj.nP)),
pad=Synapses("Pad", lambda obj:
(1, obj.nF, obj.nO, obj.nP),
lambda M, ops: ops.normal_init(M, 1.)),
d_W=Gradient("W"),
d_pad=Gradient("pad"),
d_b=Gradient("b"))
class PrecomputableAffine(Model):
def __init__(self, nO=None, nI=None, nF=None, nP=None, **kwargs):
Model.__init__(self, **kwargs)
arr = numpy.zeros((len(doc) + 1, ), dtype='uint64')
for token in doc:
arr[token.i] = token.orth
arr[len(doc)] = 0
seqs.append(arr)
return seqs, None
from thinc import describe
from thinc.describe import Dimension, Synapses, Gradient
from thinc.neural._lsuv import LSUVinit
@describe.on_data(LSUVinit)
@describe.attributes(
nM=Dimension("Vector dimensions"),
nO=Dimension("Size of output"),
W=Synapses("A projection matrix, to change vector dimensionality",
lambda obj: (obj.nO, obj.nM),
lambda W, ops: ops.xavier_uniform_init(W)),
d_W=Gradient("W"),
)
class SpacyVectors(Model):
ops = NumpyOps()
name = 'spacy-vectors'
def __init__(self, nlp, nO):
Model.__init__(self)
self._id_map = {0: 0}
self.nO = nO
self.nM = nlp.vocab.vectors_length