def from_bytes(self, bytes_data: bytes) -> "DocBin":
"""Deserialize the DocBin's annotations from a bytestring.
bytes_data (bytes): The data to load from.
RETURNS (DocBin): The loaded DocBin.
DOCS: https://spacy.io/api/docbin#from_bytes
"""
try:
msg = srsly.msgpack_loads(zlib.decompress(bytes_data))
except zlib.error:
raise ValueError(Errors.E1014)
self.attrs = msg["attrs"]
self.strings = set(msg["strings"])
lengths = numpy.frombuffer(msg["lengths"], dtype="int32")
flat_spaces = numpy.frombuffer(msg["spaces"], dtype=bool)
flat_tokens = numpy.frombuffer(msg["tokens"], dtype="uint64")
shape = (flat_tokens.size // len(self.attrs), len(self.attrs))
flat_tokens = flat_tokens.reshape(shape)
flat_spaces = flat_spaces.reshape((flat_spaces.size, 1))
self.tokens = NumpyOps().unflatten(flat_tokens, lengths)
self.spaces = NumpyOps().unflatten(flat_spaces, lengths)
self.cats = msg["cats"]
self.span_groups = msg.get("span_groups", [b"" for _ in lengths])
self.flags = msg.get("flags", [{} for _ in lengths])
if "user_data" in msg:
self.user_data = list(msg["user_data"])
else:
self.user_data = [None] * len(self)
for tokens in self.tokens:
assert len(
tokens.shape) == 2, tokens.shape # this should never happen
return self
def test_truncate_alignment_from_end(sequences, max_length, align, mask_from_end):
# print("Max length", max_length)
# print("Sequences", sequences)
# print("Mask", mask_from_end)
ops = NumpyOps()
truncated = _truncate_alignment(align, mask_from_end)
# print(truncated.dataXd.shape, truncated.lengths.sum())
# print("Before", list(map(list, ops.unflatten(align.dataXd, align.lengths))))
# print("After", list(map(list, ops.unflatten(truncated.dataXd, truncated.lengths))))
# Check that the number of tokens hasn't changed. We still need to have
# alignment for every token.
assert truncated.lengths.shape[0] == align.lengths.shape[0]
start = 0
for i, seq in enumerate(sequences):
end = start + len(seq)
# Get the alignment for this sequence of tokens. Each length in the
# alignment indicates the number of wordpiece tokens, so we need to
# check that the sum of the lengths doesn't exceed the maximum.
wp_indices = truncated[start:end]
assert wp_indices.lengths.sum() <= max_length
# We're truncating from the end, so we shouldn't see different values
# except at the end of the sequence.
seen_zero = False
before = align[start:end]
for length_now, length_before in zip(wp_indices.lengths, before.lengths):
if seen_zero:
assert length_now == 0, wp_indices.lengths
elif length_now == 0:
seen_zero = True
else:
length_now == length_before
def get_expected_predict(input_data, Ws, bs):
numpy_ops = NumpyOps()
X = input_data
for i, (W, b) in enumerate(zip(Ws, bs)):
X = numpy_ops.asarray(X)
if i > 0:
X *= X > 0
X = numpy.tensordot(X, W, axes=[[1], [1]]) + b
return X
def test_apply_alignment(nested_align, X_cols):
ops = NumpyOps()
align = get_ragged(ops, nested_align)
X_shape = (align.data.max() + 1, X_cols)
X = ops.alloc2f(*X_shape)
Y, get_dX = apply_alignment(ops, align, X)
assert isinstance(Y, Ragged)
assert Y.data.shape[0] == align.data.shape[0]
assert Y.lengths.shape[0] == len(nested_align)
dX = get_dX(Y)
assert dX.shape == X.shape
def set_backend(name, gpu_id):
global CONFIG
if name == "jax":
set_current_ops(JaxOps())
CONFIG = CONFIG.replace("PyTorch", "")
else:
if gpu_id == -1:
set_current_ops(NumpyOps())
else:
set_current_ops(CupyOps())
CONFIG = CONFIG.replace("LSTM.v1", "PyTorchLSTM.v1")
def set_backend(name, gpu_id):
global CONFIG
if name == "generic":
set_current_ops(Ops())
else:
if gpu_id == -1:
set_current_ops(NumpyOps(use_blis=True))
else:
set_current_ops(CupyOps())
if name == "pytorch":
import torch
torch.set_num_threads(1)
CONFIG = CONFIG.replace("LSTM.v1", "PyTorchLSTM.v1")
def test_large_seq2col_gpu_against_cpu(nW):
cupy_ops = CupyOps()
numpy_ops = NumpyOps()
# Use array with a large enough batch to require multiple
# CUDA grids.
batch_size = 128 * 128 * 2 # threads per block * blocks * 2
X = numpy_ops.xp.random.randn(batch_size * 2).astype("float32").reshape(
-1, 2)
X_gpu = cupy_ops.asarray2f(X)
# Use somewhat interesting sequence lengths.
lengths = numpy_ops.asarray1i([1, 4, 2, 1] * (batch_size // 8))
lengths_gpu = cupy_ops.asarray1i(lengths)
cols = numpy_ops.seq2col(X, nW=nW, lengths=lengths)
cols_gpu = cupy_ops.seq2col(X_gpu, nW=nW, lengths=lengths_gpu)
assert_allclose(cols, cols_gpu.get())
def test_list2padded():
ops = NumpyOps()
seqs = [numpy.zeros((5, 4)), numpy.zeros((8, 4)), numpy.zeros((2, 4))]
padded = ops.list2padded(seqs)
arr = padded.data
size_at_t = padded.size_at_t
assert arr.shape == (8, 3, 4)
assert size_at_t[0] == 3
assert size_at_t[1] == 3
assert size_at_t[2] == 2
assert size_at_t[3] == 2
assert size_at_t[4] == 2
assert size_at_t[5] == 1
assert size_at_t[6] == 1
assert size_at_t[7] == 1
unpadded = ops.padded2list(padded)
assert unpadded[0].shape == (5, 4)
assert unpadded[1].shape == (8, 4)
assert unpadded[2].shape == (2, 4)
def from_batch_encoding(cls,
token_data: BatchEncoding) -> "WordpieceBatch":
assert isinstance(token_data, BatchEncoding) or isinstance(
token_data, dict)
pad_token = token_data.get("pad_token", "[PAD]")
lengths = [
len([tok for tok in tokens if tok != pad_token])
for tokens in token_data["input_texts"]
]
numpy_ops = NumpyOps()
return cls(
strings=token_data["input_texts"],
input_ids=numpy_ops.asarray2i(token_data["input_ids"]),
attention_mask=numpy_ops.asarray2f(token_data["attention_mask"]),
lengths=lengths,
token_type_ids=(numpy_ops.asarray2i(token_data["token_type_ids"])
if "token_type_ids" in token_data else None),
)
def test_partials_from_config():
"""Test that functions registered with partial applications are handled
correctly (e.g. initializers)."""
name = "uniform_init.v1"
cfg = {"test": {"@initializers": name, "lo": -0.2}}
func = my_registry.make_from_config(cfg)["test"]
assert hasattr(func, "__call__")
# The partial will still have lo as an arg, just with default
assert len(inspect.signature(func).parameters) == 4
# Make sure returned partial function has correct value set
assert inspect.signature(func).parameters["lo"].default == -0.2
# Actually call the function and verify
func(NumpyOps(), (2, 3))
# Make sure validation still works
bad_cfg = {"test": {"@initializers": name, "lo": [0.5]}}
with pytest.raises(ConfigValidationError):
my_registry.make_from_config(bad_cfg)
bad_cfg = {"test": {"@initializers": name, "lo": -0.2, "other": 10}}
with pytest.raises(ConfigValidationError):
my_registry.make_from_config(bad_cfg)
def get_model(W_values, b_values):
model = Linear(W_values.shape[0], W_values.shape[1], ops=NumpyOps())
model.initialize()
model.set_param("W", W_values)
model.set_param("b", b_values)
return model
assert arr.shape == (8, 3, 4)
assert size_at_t[0] == 3
assert size_at_t[1] == 3
assert size_at_t[2] == 2
assert size_at_t[3] == 2
assert size_at_t[4] == 2
assert size_at_t[5] == 1
assert size_at_t[6] == 1
assert size_at_t[7] == 1
unpadded = ops.padded2list(padded)
assert unpadded[0].shape == (5, 4)
assert unpadded[1].shape == (8, 4)
assert unpadded[2].shape == (2, 4)
@pytest.mark.parametrize("ops", [Ops(), NumpyOps()])
@pytest.mark.parametrize("nO,nI", [(1, 2), (2, 2), (100, 200), (9, 6)])
def test_LSTM_init_with_sizes(ops, nO, nI):
model = with_padded(LSTM(nO, nI, depth=1)).initialize()
for node in model.walk():
model.ops = ops
# Check no unallocated params.
assert node.has_param("LSTM") is not None
assert node.has_param("HC0") is not None
for node in model.walk():
# Check param sizes.
if node.has_param("LSTM"):
params = node.get_param("LSTM")
assert params.shape == (
((nO * 4 * nI)) + (nO * 4) + (nO * 4 * nO + nO * 4),
)
def get_input(nr_batch, nr_in):
ops = NumpyOps()
return ops.alloc2f(nr_batch, nr_in)
def get_ops():
return NumpyOps()
def set_backend(name, gpu_id):
if gpu_id == -1:
set_current_ops(NumpyOps())
else:
set_current_ops(CupyOps())
CONFIG = CONFIG.replace("LSTM.v1", "PyTorchLSTM.v1")
def test_initializer_func_setup(init_func):
ops = NumpyOps()
data = numpy.ndarray([1, 2, 3, 4], dtype="f")
result = init_func(ops, data.shape)
assert not numpy.array_equal(data, result)
def test_initializer_from_config(name, kwargs):
"""Test that initializers are loaded and configured correctly from registry
(as partials)."""
cfg = {"test": {"@initializers": name, **kwargs}}
func = registry.make_from_config(cfg)["test"]
func(NumpyOps(), (1, 2, 3, 4))
from typing import List
import numpy
import pytest
from numpy.testing import assert_almost_equal
from spacy.vocab import Vocab
from thinc.api import NumpyOps, Model, data_validation
from thinc.types import Array2d, Ragged
from spacy.lang.en import English
from spacy.ml import FeatureExtractor, StaticVectors
from spacy.ml._character_embed import CharacterEmbed
from spacy.tokens import Doc
OPS = NumpyOps()
texts = ["These are 4 words", "Here just three"]
l0 = [[1, 2], [3, 4], [5, 6], [7, 8]]
l1 = [[9, 8], [7, 6], [5, 4]]
list_floats = [OPS.xp.asarray(l0, dtype="f"), OPS.xp.asarray(l1, dtype="f")]
list_ints = [OPS.xp.asarray(l0, dtype="i"), OPS.xp.asarray(l1, dtype="i")]
array = OPS.xp.asarray(l1, dtype="f")
ragged = Ragged(array, OPS.xp.asarray([2, 1], dtype="i"))
def get_docs():
vocab = Vocab()
for t in texts:
for word in t.split():
hash_id = vocab.strings.add(word)
vector = numpy.random.uniform(-1, 1, (7, ))
def ops():
return NumpyOps()
with Model.define_operators({">>": chain}):
model = model1 >> model2 >> model3
assert len(model.layers) == 2
assert len(model.layers[0].layers) == 2
def test_chain_right_branch(model1, model2, model3):
# Previously we 'flattened' these nested calls. We might opt to do so
# again, especially for the operators.
merge1 = chain(model1, model2)
merge2 = chain(merge1, model3)
assert len(merge1.layers) == 2
assert len(merge2.layers) == 2
@pytest.mark.parametrize("ops", [NumpyOps(), NumpyOps(use_blis=True)])
def test_chain(ops):
data = numpy.asarray([[1, 2, 3, 4]], dtype="f")
model = chain(Linear(1), Dropout(), Linear(1))
model.ops = ops
model.initialize(data, data)
Y, backprop = model(data, is_train=True)
backprop(Y)
# Layers with and without nO/nI
model = chain(Linear(1), Dropout(), Linear(1, 1))
model.initialize(data, data)
# Setting dim on model
model = chain(Linear(1), Dropout(), Linear(1))
model.set_dim("nO", 1)
model.initialize(data, None)
model = chain(Linear(1, 1), Dropout(), Linear(1, 1))
def sgd():
return SGD(0.001, ops=NumpyOps())
from hypothesis.strategies import composite, integers
from numpy.testing import assert_allclose
from thinc.api import NumpyOps, CupyOps, Ops, get_ops
from thinc.api import get_current_ops, use_ops
from thinc.api import fix_random_seed
from thinc.api import LSTM
import inspect
from .. import strategies
from ..strategies import ndarrays_of_shape
MAX_EXAMPLES = 10
VANILLA_OPS = Ops(numpy)
NUMPY_OPS = NumpyOps()
BLIS_OPS = NumpyOps(use_blis=True)
CPU_OPS = [NUMPY_OPS, VANILLA_OPS]
XP_OPS = [NUMPY_OPS]
if CupyOps.xp is not None:
XP_OPS.append(CupyOps())
ALL_OPS = XP_OPS + [VANILLA_OPS]
@pytest.mark.parametrize("op", [NumpyOps, CupyOps])
def test_ops_consistency(op):
"""Test that specific ops don't define any methods that are not on the
Ops base class and that all ops methods define the exact same arguments."""
attrs = [m for m in dir(op) if not m.startswith("_")]
for attr in attrs:
assert hasattr(Ops, attr)