def fit(self, X, y, X_va=None, y_va=None, skip_output=-1):
y, y_va = self._check_y_shape(y, y_va)
assert X.shape[0] == y.shape[0], "X and y have the a different number of samples"
assert X.shape[1] == self.n_inputs, "X doesn't have the right number of features"
if X_va is not None:
assert X_va.shape[0] == y_va.shape[0], "X_va and y_va have the a different number of samples"
assert X_va.shape[1] == X.shape[1], "X_va doesn't have the right number of features"
# PyCUDA, doesn't allow shuffling via indexing (`X[idx]`) so to have
# a random split, we just shuffle the data (if we're allowed to)
if X_va is None and self.fraction_validation_set > 0.0:
if self.shuffle_data:
idx = np.arange(X.shape[0])
Xn, yn = op.shuffle_rows(X, y, idx=idx)
else:
warnings.warn("using first part of X as validation set without shuffling first")
vi = int(y.shape[0]*self.fraction_validation_set)
X, X_va, y, y_va = X[vi:], X[:vi], y[vi:], y[:vi]
oldverbose = self.verbose
try:
# generate storage for shuffling now
if self.shuffle_data:
idx = np.arange(X.shape[0])
Xn, yn = op.shuffle_rows(X, y, idx=idx)
t0 = time.time()
for i in range(self.current_epoch, self.max_iter):
if self.shuffle_data:
Xn, yn = op.shuffle_rows(X, y, output=(Xn, yn), idx=idx)
else:
Xn, yn = X, y
err = float(self.partial_fit(Xn, yn, encode_labels=False))
if oldverbose and skip_output > 0:
self.verbose = (self.current_epoch % skip_output) == 0
# always show first epoch to measure times
if self.current_epoch == 1:
self.verbose = True
#self.track_progress(t0, err, Xn, yn, X_va, y_va)
for cb in self._epoch_callbacks:
cb(self, t0, err, Xn, yn, X_va, y_va)
if self.snapshot_interval is not None \
and (self.snapshot_interval % self.current_epoch == 0) \
and self.current_epoch > 0:
_create_snapshot(self)
# early stopping checks
if self._no_improvement_since >= self.convergence_iter_tol and self.early_stopping:
for i, l in enumerate(self.layers):
l.W = self._best_params[0][i]
l.b = self._best_params[1][i]
self.current_epoch -= self._no_improvement_since
break
finally:
self.verbose = oldverbose
def test_swaprows():
n = 1270
X = 5.0*np.random.randn(n, 1000).astype(np.float32)
ytemp = np.array(range(X.shape[0]))[:, None]
y = np.hstack((ytemp, ytemp, ytemp)).astype(np.float32)
idx = list(range(X.shape[0]))
idx = np.array(idx, dtype=np.int32)
np.random.shuffle(idx)
Xd = op.to_gpu(X)
yd = op.to_gpu(y)
Xoutd = gpuarray.empty_like(Xd)
youtd = gpuarray.empty_like(yd)
op.shuffle_rows(Xd, yd, (Xoutd, youtd), idx)
X2 = op.to_cpu(Xoutd)
y2 = op.to_cpu(youtd)
assert_allclose(X[idx], X2)
assert_allclose(y[idx], y2)
def test_swaprows():
n = 1270
X = 5.0 * np.random.randn(n, 1000).astype(np.float32)
ytemp = np.array(range(X.shape[0]))[:, None]
y = np.hstack((ytemp, ytemp, ytemp)).astype(np.float32)
idx = list(range(X.shape[0]))
idx = np.array(idx, dtype=np.int32)
np.random.shuffle(idx)
Xd = op.to_gpu(X)
yd = op.to_gpu(y)
Xoutd = gpuarray.empty_like(Xd)
youtd = gpuarray.empty_like(yd)
op.shuffle_rows(Xd, yd, (Xoutd, youtd), idx)
X2 = op.to_cpu(Xoutd)
y2 = op.to_cpu(youtd)
assert_allclose(X[idx], X2)
assert_allclose(y[idx], y2)
def fit(self, X, y, X_va=None, y_va=None, skip_output=-1):
y, y_va = self._check_y_shape(y, y_va)
assert X.shape[0] == y.shape[
0], "X and y have the a different number of samples"
assert X.shape[
1] == self.n_inputs, "X doesn't have the right number of features"
if X_va is not None:
assert X_va.shape[0] == y_va.shape[
0], "X_va and y_va have the a different number of samples"
assert X_va.shape[1] == X.shape[
1], "X_va doesn't have the right number of features"
# PyCUDA, doesn't allow shuffling via indexing (`X[idx]`) so to have
# a random split, we just shuffle the data (if we're allowed to)
if X_va is None and self.fraction_validation_set > 0.0:
if self.shuffle_data:
idx = np.arange(X.shape[0])
Xn, yn = op.shuffle_rows(X, y, idx=idx)
else:
warnings.warn(
"using first part of X as validation set without shuffling first"
)
vi = int(y.shape[0] * self.fraction_validation_set)
X, X_va, y, y_va = X[vi:], X[:vi], y[vi:], y[:vi]
oldverbose = self.verbose
try:
# generate storage for shuffling now
if self.shuffle_data:
idx = np.arange(X.shape[0])
Xn, yn = op.shuffle_rows(X, y, idx=idx)
t0 = time.time()
for i in range(self.current_epoch, self.max_iter):
if self.shuffle_data:
Xn, yn = op.shuffle_rows(X, y, output=(Xn, yn), idx=idx)
else:
Xn, yn = X, y
err = float(self.partial_fit(Xn, yn, encode_labels=False))
if oldverbose and skip_output > 0:
self.verbose = (self.current_epoch % skip_output) == 0
# always show first epoch to measure times
if self.current_epoch == 1:
self.verbose = True
#self.track_progress(t0, err, Xn, yn, X_va, y_va)
for cb in self._epoch_callbacks:
cb(self, t0, err, Xn, yn, X_va, y_va)
if self.snapshot_interval is not None \
and (self.snapshot_interval % self.current_epoch == 0) \
and self.current_epoch > 0:
_create_snapshot(self)
# early stopping checks
if self._no_improvement_since >= self.convergence_iter_tol and self.early_stopping:
for i, l in enumerate(self.layers):
l.W = self._best_params[0][i]
l.b = self._best_params[1][i]
self.current_epoch -= self._no_improvement_since
break
finally:
self.verbose = oldverbose
