spaCy/spacy/serialize.pyx

323 lines
10 KiB
Cython

from libcpp.vector cimport vector
from libc.stdint cimport uint32_t
from libc.stdint cimport int64_t
from libc.stdint cimport int32_t
from libc.stdint cimport uint64_t
from preshed.maps cimport PreshMap
from murmurhash.mrmr cimport hash64
import numpy
cimport cython
ctypedef unsigned char uchar
# Format
# - Total number of bytes in message (32 bit int)
# - Words, terminating in an EOL symbol, huffman coded ~12 bits per word
# - Spaces ~1 bit per word
# - Parse: Huffman coded head offset / dep label / POS tag / entity IOB tag
# combo. ? bits per word. 40 * 80 * 40 * 12 = 1.5m symbol vocab
# Note that we're setting the most significant bits here first, when in practice
# we're actually wanting the last bit to be most significant (for Huffman coding,
# anyway).
cdef Code bit_append(Code code, bint bit) nogil:
cdef uint64_t one = 1
if bit:
code.bits |= one << code.length
else:
code.bits &= ~(one << code.length)
code.length += 1
return code
cdef class BitArray:
cdef bytes data
cdef unsigned char byte
cdef unsigned char bit_of_byte
cdef uint32_t i
def __init__(self):
self.data = b''
self.byte = 0
self.bit_of_byte = 0
self.i = 0
def __iter__(self):
cdef uchar byte, i
cdef uchar one = 1
start_byte = self.i // 8
if (self.i % 8) != 0:
for i in range(self.i % 8):
yield 1 if (self.data[start_byte] & (one << i)) else 0
start_byte += 1
for byte in self.data[start_byte:]:
for i in range(8):
yield 1 if byte & (one << i) else 0
for i in range(self.bit_of_byte):
yield 1 if self.byte & (one << i) else 0
def as_bytes(self):
if self.bit_of_byte != 0:
return self.data + chr(self.byte)
else:
return self.data
def append(self, bint bit):
cdef uint64_t one = 1
print 'append', bit
if bit:
self.byte |= one << self.bit_of_byte
else:
self.byte &= ~(one << self.bit_of_byte)
self.bit_of_byte += 1
if self.bit_of_byte == 8:
self.data += chr(self.byte)
self.byte = 0
self.bit_of_byte = 0
cdef int extend(self, uint64_t code, char n_bits) except -1:
cdef uint64_t one = 1
cdef unsigned char bit_of_code
for bit_of_code in range(n_bits):
if code & (one << bit_of_code):
self.byte |= one << self.bit_of_byte
else:
self.byte &= ~(one << self.bit_of_byte)
self.bit_of_byte += 1
if self.bit_of_byte == 8:
self.data += chr(self.byte)
self.byte = 0
self.bit_of_byte = 0
cdef class Serializer:
# Manage codecs, maintain consistent format for io
def __init__(self, Vocab vocab, model_dir):
self.vocab = vocab
self.lex = None
self.codecs = []
def __call__(self, doc_or_bits):
if isinstance(doc_or_bits, Doc):
return self.serialize(doc_or_bits)
elif isinstance(doc_or_bits, BitArray):
return self.deserialize(doc_or_bits)
else:
raise ValueError(doc_or_bits)
def train(self, doc):
array = doc.to_array(self.attrs)
for i, attr in enumerate(self.attrs):
for j in range(doc.length):
self.freqs[attr].inc(array[i, j], 1)
self.freqs[attr].inc(self.eol, 1)
def serialize(self, doc):
bits = BitArray()
array = doc.to_array(self.attrs)
for i, attr in enumerate(self.attrs, self.codecs):
codec.encode(array[i,], bits)
return bits
@cython.boundscheck(False)
def deserialize(self, bits):
cdef Doc doc = Doc(self.vocab)
biterator = iter(bits)
ids = self.codecs[0].decode(bits)
cdef int id_
cdef bint is_spacy
for id_ in ids:
is_spacy = biterator.next()
doc.push_back(vocab.lexemes.at(id_), is_spacy)
cdef int length = doc.length
cdef int i
cdef attr_t value
cdef attr_id_t attr_id
cdef attr_t[:] values
cdef TokenC* tokens = doc.data
for codec in vocab.codecs[1:]:
values = codec.decode(biterator)
attr_id = codec.id
if attr_id == HEAD:
for i in range(length):
tokens[i].head = values[i]
elif attr_id == TAG:
for i in range(length):
tokens[i].tag = values[i]
elif attr_id == DEP:
for i in range(length):
tokens[i].dep = values[i]
elif attr_id == ENT_IOB:
for i in range(length):
tokens[i].ent_iob = values[i]
elif attr_id == ENT_TYPE:
for i in range(length):
tokens[i].ent_type = values[i]
return doc
def lex_codec(self):
cdef Address mem
cdef int i
cdef float[:] cv_probs
mem = Address(len(self), sizeof(float))
probs = <float*>mem.ptr
for i in range(len(self.vocab)):
probs[i] = <float>c_exp(self.lexemes[i].prob)
cv_probs = <float[:len(self)]>probs
return HuffmanCodec(cv_probs, 0, id=ID)
cdef class HuffmanCodec:
"""Create a Huffman code table, and use it to pack and unpack sequences into
byte strings. Emphasis is on efficiency, so API is quite strict:
Messages will be encoded/decoded as indices that refer to the probability sequence.
For instance, the sequence [5, 10, 8] indicates the 5th most frequent item,
the 10th most frequent item, the 8th most frequent item. The codec will add
the EOL symbol to your message. An exception will be raised if you include
the EOL symbol in your message.
Arguments:
probs (float[:]): A descending-sorted sequence of probabilities/weights.
Must include a weight for an EOL symbol.
eol (uint32_t): The index of the weight of the EOL symbol.
"""
def __init__(self, float[:] probs, uint32_t eol, id=0):
self.id = id
self.eol = eol
self.codes.resize(len(probs))
for i in range(len(self.codes)):
self.codes[i].bits = 0
self.codes[i].length = 0
populate_nodes(self.nodes, probs)
cdef Code path
path.bits = 0
path.length = 0
assign_codes(self.nodes, self.codes, len(self.nodes) - 1, path)
def encode(self, uint32_t[:] sequence, BitArray bits=None):
if bits is None:
bits = BitArray()
for i in sequence:
bits.extend(self.codes[i].bits, self.codes[i].length)
bits.extend(self.codes[self.eol].bits, self.codes[self.eol].length)
return bits
def decode(self, bits):
node = self.nodes.back()
symbols = []
for bit in bits:
branch = node.right if bit else node.left
if branch >= 0:
node = self.nodes.at(branch)
else:
symbol = -(branch + 1)
if symbol == self.eol:
return symbols
else:
symbols.append(symbol)
node = self.nodes.back()
return symbols
property strings:
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)
def __get__(self):
output = []
cdef int i, j
cdef bytes string
cdef Code code
for i in range(self.codes.size()):
code = self.codes[i]
string = b'{0:b}'.format(code.bits).rjust(code.length, '0')
string = string[::-1]
output.append(string)
return output
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)
cdef int populate_nodes(vector[Node]& nodes, float[:] probs) except -1:
assert len(probs) >= 3
cdef int size = len(probs)
cdef int i = size - 1
cdef int j = 0
while i >= 0 or (j+1) < nodes.size():
if i < 0:
_cover_two_nodes(nodes, j)
j += 2
elif j >= nodes.size():
_cover_two_words(nodes, i, i-1, probs[i] + probs[i-1])
i -= 2
elif i >= 1 and (j == nodes.size() or probs[i-1] < nodes[j].prob):
_cover_two_words(nodes, i, i-1, probs[i] + probs[i-1])
i -= 2
elif (j+1) < nodes.size() and nodes[j+1].prob < probs[i]:
_cover_two_nodes(nodes, j)
j += 2
else:
_cover_one_word_one_node(nodes, j, i, probs[i])
i -= 1
j += 1
return 0
cdef int _cover_two_nodes(vector[Node]& nodes, int j) nogil:
cdef Node node
node.left = j
node.right = j+1
node.prob = nodes[j].prob + nodes[j+1].prob
nodes.push_back(node)
cdef int _cover_one_word_one_node(vector[Node]& nodes, int j, int id_, float prob) nogil:
cdef Node node
# Encode leaves as negative integers, where the integer is the index of the
# word in the vocabulary.
cdef int64_t leaf_id = - <int64_t>(id_ + 1)
cdef float new_prob = prob + nodes[j].prob
if prob < nodes[j].prob:
node.left = leaf_id
node.right = j
node.prob = new_prob
else:
node.left = j
node.right = leaf_id
node.prob = new_prob
nodes.push_back(node)
cdef int _cover_two_words(vector[Node]& nodes, int id1, int id2, float prob) nogil:
cdef Node node
node.left = -(id1+1)
node.right = -(id2+1)
node.prob = prob
nodes.push_back(node)
cdef int assign_codes(vector[Node]& nodes, vector[Code]& codes, int i, Code path) except -1:
cdef Code left_path = bit_append(path, 0)
cdef Code right_path = bit_append(path, 1)
# Assign down left branch
if nodes[i].left >= 0:
assign_codes(nodes, codes, nodes[i].left, left_path)
else:
# Leaf on left
id_ = -(nodes[i].left + 1)
codes[id_] = left_path
# Assign down right branch
if nodes[i].right >= 0:
assign_codes(nodes, codes, nodes[i].right, right_path)
else:
# Leaf on right
id_ = -(nodes[i].right + 1)
codes[id_] = right_path