def use_gpu(gpu_id):
try:
import cupy.cuda.device
except ImportError:
return None
from thinc.neural.ops import CupyOps
device = cupy.cuda.device.Device(gpu_id)
device.use()
Model.ops = CupyOps()
Model.Ops = CupyOps
return device
def use_gpu(gpu_id):
try:
import cupy.cuda.device
except ImportError:
return None
from thinc.neural.ops import CupyOps
device = cupy.cuda.device.Device(gpu_id)
device.use()
Model.ops = CupyOps()
Model.Ops = CupyOps
return device
def set_active_gpu(gpu_id):
import cupy.cuda.device
device = cupy.cuda.device.Device(gpu_id)
device.use()
try:
import torch
torch.cuda.set_device(gpu_id)
torch.set_default_tensor_type("torch.cuda.FloatTensor")
except ImportError:
pass
return device
def set_active_gpu(gpu_id: int) -> "cupy.cuda.Device": # pragma: no cover
"""Set the current GPU device for cupy and torch (if available)."""
import cupy.cuda.device
device = cupy.cuda.device.Device(gpu_id)
device.use()
try:
import torch
torch.cuda.set_device(gpu_id)
torch.set_default_tensor_type("torch.cuda.FloatTensor")
except ImportError:
pass
return device
def to_gpu(self, device_num):
import cupy.cuda.device
device = cupy.cuda.device.Device(device_num)
device.use()
queue = [self]
for layer in queue:
layer.ops = CupyOps()
layer.Ops = CupyOps
if hasattr(layer, '_mem'):
layer._mem._mem = self.ops.xp.asarray(layer._mem._mem)
layer._mem.ops = layer.ops
if hasattr(layer, '_layers'):
queue.extend(layer._layers)
return device
def to_gpu(self, device_num):
import cupy.cuda.device
device = cupy.cuda.device.Device(device_num)
device.use()
queue = [self]
for layer in queue:
layer.ops = CupyOps()
layer.Ops = CupyOps
if hasattr(layer, u'_mem'):
layer._mem._mem = self.ops.xp.asarray(layer._mem._mem)
layer._mem.ops = layer.ops
if hasattr(layer, u'_layers'):
queue.extend(layer._layers)
return device
def to_gpu(self, gpu_id: int) -> None: # pragma: no cover
"""Transfer the model to a given GPU device."""
import cupy.cuda.device
device = cupy.cuda.device.Device(gpu_id)
with device.use():
self._to_ops(CupyOps())
def main(vectors,
gpu_id=-1,
n_neighbors=100,
batch_size=1024,
cutoff=0,
start=0,
end=None):
"""
Step 6: Precompute nearest-neighbor queries (optional)
Precompute nearest-neighbor queries for every entry in the vocab to make
Sense2Vec.most_similar faster. The --cutoff option lets you define the
number of earliest rows to limit the neighbors to. For instance, if cutoff
is 100000, no word will have a nearest neighbor outside of the top 100k
vectors.
"""
if gpu_id == -1:
xp = numpy
else:
import cupy as xp
import cupy.cuda.device
cupy.take_along_axis = take_along_axis
device = cupy.cuda.device.Device(gpu_id)
device.use()
vectors_dir = Path(vectors)
vectors_file = vectors_dir / "vectors"
if not vectors_dir.is_dir() or not vectors_file.exists():
err = "Are you passing in the exported sense2vec directory containing a vectors file?"
msg.fail(f"Can't load vectors from {vectors}", err, exits=1)
with msg.loading(f"Loading vectors from {vectors}"):
vectors = xp.load(str(vectors_file))
msg.good(
f"Loaded {vectors.shape[0]:,} vectors with dimension {vectors.shape[1]}"
)
norms = xp.linalg.norm(vectors, axis=1, keepdims=True)
norms[norms == 0] = 1
# Normalize to unit norm
vectors /= norms
if cutoff < 1:
cutoff = vectors.shape[0]
if end is None:
end = vectors.shape[0]
mean = float(norms.mean())
var = float(norms.var())
msg.good(f"Normalized (mean {mean:,.2f}, variance {var:,.2f})")
msg.info(
f"Finding {n_neighbors:,} neighbors among {cutoff:,} most frequent")
n = min(n_neighbors, vectors.shape[0])
subset = vectors[:cutoff]
best_rows = xp.zeros((end - start, n), dtype="i")
scores = xp.zeros((end - start, n), dtype="f")
for i in tqdm.tqdm(list(range(start, end, batch_size))):
size = min(batch_size, end - i)
batch = vectors[i:i + size]
sims = xp.dot(batch, subset.T)
# Set self-similarities to -inf, so that we don't return them.
for j in range(size):
if i + j < sims.shape[1]:
sims[j, i + j] = -xp.inf
# This used to use argpartition, to do a partial sort...But this ended
# up being a ratsnest of terrible numpy crap. Just sorting the whole
# list isn't really slower, and it's much simpler to read.
ranks = xp.argsort(sims, axis=1)
batch_rows = ranks[:, -n:]
# Reverse
batch_rows = batch_rows[:, ::-1]
batch_scores = xp.take_along_axis(sims, batch_rows, axis=1)
best_rows[i:i + size] = batch_rows
scores[i:i + size] = batch_scores
msg.info("Saving output")
if not isinstance(best_rows, numpy.ndarray):
best_rows = best_rows.get()
if not isinstance(scores, numpy.ndarray):
scores = scores.get()
output = {
"indices": best_rows,
"scores": scores.astype("float16"),
"start": start,
"end": end,
"cutoff": cutoff,
}
output_file = vectors_dir / "cache"
with msg.loading("Saving output..."):
srsly.write_msgpack(output_file, output)
msg.good(f"Saved cache to {output_file}")
