def test_randomly_replace_elements():
for val in (0.0, 0.5, 5):
for p in (0.1, 0.2, 0.5, 0.75, 0.99):
X = np.random.normal(size=(1024, 2048)).astype(np.float32)
Xd = op.to_gpu(X)
Xr, M = op.randomly_replace_elements(X, p, val)
assert (Xr is X)
assert_almost_equal((X == val).mean(),
p,
decimal=2,
err_msg="val: %.1f p: %.1f" % (val, p))
assert_almost_equal(M.mean(),
1 - p,
decimal=2,
err_msg="M val: %.1f p: %.1f" % (val, p))
Xrd, Md = op.randomly_replace_elements(Xd, p, val)
assert (Xrd is Xd)
assert_almost_equal(op.to_cpu(op.mean(Xd == val)),
p,
decimal=2,
err_msg="val: %.1f p: %.1f (gpu)" % (val, p))
assert_almost_equal(op.to_cpu(op.mean(Md)),
1 - p,
decimal=2,
err_msg="M val: %.1f p: %.1f (gpu)" % (val, p))
def test_csrmm_bug():
''' the 2nd call might crash'''
from scipy.sparse import csr_matrix
W = np.random.normal(size=(5, 3)).astype(np.float32, order="c")
X = np.random.laplace(size=(6, 3)).astype(np.float32)
X[X<0.1] = 0
X = csr_matrix(X, dtype=np.float32)
Xd = GPUCSRArray(X)
Wd = op.to_gpu(W)
Cd = op.dot(Xd, Wd, False, True, out=None, stream=op.streams[0])
op.add_dot(Cd, Xd, Wd, True, False, alpha=-0.3, beta=1.0, stream=op.streams[0])
op.mean(Cd, axis=0, stream=op.streams[1])
def bprop(self, delta, momentum=0.0):
op.streams[2].synchronize() # make sure layer above is done
self.dfunc(delta, self.A, self.Z, stream=op.streams[0])
op.streams[0].synchronize()
op.add_dot(delta,
self.X,
self.dW,
True,
False,
alpha=1.0 / delta.shape[0],
beta=momentum,
stream=op.streams[0])
m = op.mean(delta, axis=0, stream=op.streams[1])
op.add_vec(self.db, 1.0, m, beta=momentum, stream=op.streams[1])
if self.l2_penalty > 0:
op.add_vec(self.dW, self.l2_penalty, self.W, stream=op.streams[0])
if not self.is_input_layer:
if self.dropout > 0.0 and self.activation not in ("relu",
"sigmoid"):
return op.dot(delta, self.W) * self.M
else:
return op.dot(delta, self.W)
else:
return 0.0
def test_randomly_replace_elements():
for val in (0.0, 0.5, 5):
for p in (0.1, 0.2, 0.5, 0.75, 0.99):
X = np.random.normal(size=(1024, 2048)).astype(np.float32)
Xd = op.to_gpu(X)
Xr, M = op.randomly_replace_elements(X, p, val)
assert(Xr is X)
assert_almost_equal((X == val).mean(), p, decimal=2,
err_msg="val: %.1f p: %.1f" % (val, p))
assert_almost_equal(M.mean(), 1-p, decimal=2,
err_msg="M val: %.1f p: %.1f" % (val, p))
Xrd, Md = op.randomly_replace_elements(Xd, p, val)
assert(Xrd is Xd)
assert_almost_equal(op.to_cpu(op.mean(Xd == val)), p, decimal=2,
err_msg="val: %.1f p: %.1f (gpu)" % (val, p))
assert_almost_equal(op.to_cpu(op.mean(Md)), 1-p, decimal=2,
err_msg="M val: %.1f p: %.1f (gpu)" % (val, p))
def test_csrmm_bug():
''' the 2nd call might crash'''
from scipy.sparse import csr_matrix
W = np.random.normal(size=(5, 3)).astype(np.float32, order="c")
X = np.random.laplace(size=(6, 3)).astype(np.float32)
X[X < 0.1] = 0
X = csr_matrix(X, dtype=np.float32)
Xd = GPUCSRArray(X)
Wd = op.to_gpu(W)
Cd = op.dot(Xd, Wd, False, True, out=None, stream=op.streams[0])
op.add_dot(Cd,
Xd,
Wd,
True,
False,
alpha=-0.3,
beta=1.0,
stream=op.streams[0])
op.mean(Cd, axis=0, stream=op.streams[1])
def _get_score(self, target, pred):
'''Calculates the quality of predictions of a model.
Like sklearn, we follow the convention that higher return values are
better than lower return values.'''
if self.output == "softmax":
# convert from 1hot
if len(target.shape) != 1 and target.shape[1] != 1:
target = op.to_cpu(op.argmax(target, 1))
if len(pred.shape) != 1 and pred.shape[1] != 1:
pred = op.to_cpu(op.argmax(pred, 1))
acc = op.to_cpu(op.mean(pred == target))
return float(acc)
elif self.output == "sigmoid":
# Note: this is meant for multitask learning, but for e.g.
# using sigmoid+squarederror as multiclass problem, this will
# give the wrong result!
return op.to_cpu(op.mean(target == (pred > 0.5)))
elif self.output == "linear":
return -op.to_cpu(op.mean((target - pred)**2)) # negate by convention
else:
raise NotImplementedError()
def _get_score(self, target, pred):
'''Calculates the quality of predictions of a model.
Like sklearn, we follow the convention that higher return values are
better than lower return values.'''
if self.output == "softmax":
# convert from 1hot
if len(target.shape) != 1 and target.shape[1] != 1:
target = op.to_cpu(op.argmax(target, 1))
if len(pred.shape) != 1 and pred.shape[1] != 1:
pred = op.to_cpu(op.argmax(pred, 1))
acc = op.to_cpu(op.mean(pred == target))
return float(acc)
elif self.output == "sigmoid":
# Note: this is meant for multitask learning, but for e.g.
# using sigmoid+squarederror as multiclass problem, this will
# give the wrong result!
return op.to_cpu(op.mean(target == (pred > 0.5)))
elif self.output == "linear":
return -op.to_cpu(op.mean(
(target - pred)**2)) # negate by convention
else:
raise NotImplementedError()
def bprop(self, delta, momentum=0.0):
op.streams[2].synchronize() # make sure layer above is done
self.dfunc(delta, self.A, self.Z, stream=op.streams[0])
op.streams[0].synchronize()
op.add_dot(delta, self.X, self.dW, True, False,
alpha=1.0/delta.shape[0], beta=momentum, stream=op.streams[0])
m = op.mean(delta, axis=0, stream=op.streams[1])
op.add_vec(self.db, 1.0, m, beta=momentum, stream=op.streams[1])
if self.l2_penalty > 0:
op.add_vec(self.dW, self.l2_penalty, self.W, stream=op.streams[0])
if not self.is_input_layer:
return op.dot(delta, self.W, stream=op.streams[2])
else:
return 0.0