explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-07-07 21:33:13 +03:00
Code Issues Packages Projects Releases Wiki Activity

Fix GPU usage

Browse Source
This commit is contained in:
Matthew Honnibal 2017-10-31 02:33:34 +01:00
parent 7698903617
commit ce876c551e
1 changed files with 17 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
spacy/_ml.py
View File

@ -136,13 +136,14 @@ class PrecomputableAffine(Model):
def _add_padding(self, Yf): def _add_padding(self, Yf):
Yf_padded = self.ops.xp.vstack((self.pad, Yf)) Yf_padded = self.ops.xp.vstack((self.pad, Yf))
return Yf_padded[1:] return Yf_padded
def _backprop_padding(self, dY, ids): def _backprop_padding(self, dY, ids):
for i in range(ids.shape[0]): # (1, nF, nO, nP) += (nN, nF, nO, nP) where IDs (nN, nF) < 0
for j in range(ids.shape[1]): d_feats = dY[ids]
if ids[i, j] < 0: ids = ids.reshape((ids.shape[0], ids.shape[1], 1, 1))
self.d_pad[0, j] += dY[i, j] d_feats *= ids < 0
self.d_pad += d_feats.sum(axis=0, keepdims=True)
return dY, ids return dY, ids
@staticmethod @staticmethod
@ -157,12 +158,15 @@ class PrecomputableAffine(Model):
''' '''
if (model.W**2).sum() != 0.: if (model.W**2).sum() != 0.:
return return
model.ops.normal_init(model.W, model.nF * model.nI, inplace=True) ops = model.ops
xp = ops.xp
ops.normal_init(model.W, model.nF * model.nI, inplace=True)
ids = numpy.zeros((5000, model.nF), dtype='i') ids = ops.allocate((5000, model.nF), dtype='f')
ids += numpy.asarray(numpy.random.uniform(0, 1000, ids.shape), dtype='i') ids += xp.random.uniform(0, 1000, ids.shape)
tokvecs = numpy.zeros((5000, model.nI), dtype='f') ids = ops.asarray(ids, dtype='i')
tokvecs += numpy.random.normal(loc=0., scale=1., tokvecs = ops.allocate((5000, model.nI), dtype='f')
tokvecs += xp.random.normal(loc=0., scale=1.,
size=tokvecs.size).reshape(tokvecs.shape) size=tokvecs.size).reshape(tokvecs.shape)
def predict(ids, tokvecs): def predict(ids, tokvecs):
@ -185,10 +189,10 @@ class PrecomputableAffine(Model):
t_i = 0 t_i = 0
for t_i in range(t_max): for t_i in range(t_max):
acts1 = predict(ids, tokvecs) acts1 = predict(ids, tokvecs)
var = numpy.var(acts1) var = model.ops.xp.var(acts1)
mean = numpy.mean(acts1) mean = model.ops.xp.mean(acts1)
if abs(var - 1.0) >= tol_var: if abs(var - 1.0) >= tol_var:
model.W /= numpy.sqrt(var) model.W /= model.ops.xp.sqrt(var)
elif abs(mean) >= tol_mean: elif abs(mean) >= tol_mean:
model.b -= mean model.b -= mean
else: else: