def test_ort_gradient_optimizers_optimal_use_ort(self):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X, y = make_regression( # pylint: disable=W0632
100,
n_features=10,
bias=2,
random_state=0)
X = X.astype(numpy.float32)
y = y.astype(numpy.float32)
X_train, _, y_train, __ = train_test_split(X, y)
reg = LinearRegression()
reg.fit(X_train, y_train)
reg.coef_ = reg.coef_.reshape((1, -1))
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearRegressor'})
onx_loss = add_loss_output(onx)
inits = ['intercept', 'coef']
train_session = OrtGradientOptimizer(
onx_loss,
inits,
max_iter=10,
learning_rate=LearningRateSGD(learning_rate='optimal'))
self.assertRaise(lambda: train_session.get_state(), AttributeError)
train_session.fit(X_train, y_train, use_numpy=False)
state_tensors = train_session.get_state()
self.assertEqual(len(state_tensors), 2)
r = repr(train_session)
self.assertIn("OrtGradientOptimizer(model_onnx=", r)
self.assertIn("learning_rate='optimal'", r)
losses = train_session.train_losses_
self.assertGreater(len(losses), 1)
self.assertFalse(any(map(numpy.isnan, losses)))
def test_ort_gradient_optimizers_use_numpy_w_l1(self):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X, y = make_regression( # pylint: disable=W0632
100,
n_features=10,
bias=2,
random_state=0)
X = X.astype(numpy.float32)
y = y.astype(numpy.float32)
w = (numpy.random.rand(y.shape[0]) + 1).astype(X.dtype)
X_train, _, y_train, __, w_train, ___ = train_test_split(X, y, w)
reg = LinearRegression()
reg.fit(X_train, y_train, sample_weight=w_train)
reg.coef_ = reg.coef_.reshape((1, -1))
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearRegressor'})
set_model_props(onx, {'info': 'unit test'})
onx_loss = add_loss_output(onx, weight_name='weight', score_name='l1')
inits = ['intercept', 'coef']
train_session = OrtGradientOptimizer(onx_loss,
inits,
learning_rate=1e-3)
self.assertRaise(lambda: train_session.get_state(), AttributeError)
train_session.fit(X_train, y_train, w_train, use_numpy=True)
state_tensors = train_session.get_state()
self.assertEqual(len(state_tensors), 2)
r = repr(train_session)
self.assertIn("OrtGradientOptimizer(model_onnx=", r)
self.assertIn("learning_rate='invscaling'", r)
losses = train_session.train_losses_
self.assertGreater(len(losses), 1)
self.assertFalse(any(map(numpy.isnan, losses)))
def wtest_ort_gradient_optimizers_binary(self, use_weight=False):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X = numpy.arange(60).astype(numpy.float32).reshape((-1, 3))
y = numpy.arange(X.shape[0]).astype(numpy.float32).reshape(
(-1, 1)) > 10
X = X.astype(numpy.float32)
y = y.astype(numpy.int64)
y[0, 0] = 0
y[-1, 0] = 1
w = (numpy.random.rand(X.shape[0]) + 1).astype(numpy.float32)
X_train, _, y_train, __, w_train, ___ = train_test_split(X, y, w)
reg = SGDClassifier(loss='log')
reg.fit(X_train, y_train.ravel())
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearClassifier'},
options={'zipmap': False})
onx_loss = add_loss_output(
onx,
'log',
output_index=1,
weight_name='weight' if use_weight else None)
inits = ['intercept', 'coef']
inputs = onx_loss.graph.input
self.assertEqual(len(inputs), 3 if use_weight else 2)
dt = inputs[1].type.tensor_type.elem_type
self.assertEqual(TensorProto.INT64, dt) # pylint: disable=E1101
train_session = OrtGradientOptimizer(onx_loss,
inits,
learning_rate=1e9)
self.assertRaise(lambda: train_session.get_state(), AttributeError)
if use_weight:
train_session.fit(X_train,
y_train.reshape((-1, 1)),
w_train.reshape((-1, 1)),
use_numpy=False)
else:
train_session.fit(X_train,
y_train.reshape((-1, 1)),
use_numpy=False)
state_tensors = train_session.get_state()
self.assertEqual(len(state_tensors), 2)
r = repr(train_session)
self.assertIn("OrtGradientOptimizer(model_onnx=", r)
self.assertIn("learning_rate='invscaling'", r)
losses = train_session.train_losses_
self.assertGreater(len(losses), 1)
if any(map(numpy.isnan, losses)):
raise AssertionError(losses)
def test_ort_gradient_optimizers_use_numpy_nan_w(self):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X, y = make_regression( # pylint: disable=W0632
100,
n_features=10,
bias=2,
random_state=0)
X = X.astype(numpy.float32)
y = y.astype(numpy.float32)
w = (numpy.random.rand(y.shape[0]) + 1).astype(X.dtype)
X_train, _, y_train, __, w_train, ___ = train_test_split(X, y, w)
reg = LinearRegression()
reg.fit(X_train, y_train, w_train)
reg.coef_ = reg.coef_.reshape((1, -1))
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearRegressor'})
set_model_props(onx, {'info': 'unit test'})
onx_loss = add_loss_output(onx, weight_name='weight')
inits = ['intercept', 'coef']
train_session = OrtGradientOptimizer(onx_loss,
inits,
learning_rate=1e3)
self.assertRaise(
lambda: train_session.fit(
X_train, y_train, w_train, use_numpy=True), ConvergenceError)
def wtest_ort_gradient_optimizers_reg(self, use_weight=False):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X = numpy.arange(60).astype(numpy.float32).reshape((-1, 3))
y = numpy.arange(X.shape[0]).astype(numpy.float32).reshape((-1, 1))
y[0, 0] += 1
y[-1, 0] += 1
w = (numpy.random.rand(X.shape[0]) + 1).astype(numpy.float32)
X_train, _, y_train, __, w_train, ___ = train_test_split(X, y, w)
reg = SGDRegressor()
reg.fit(X_train, y_train.ravel())
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearRegressor'})
onx_loss = add_loss_output(
onx, 'squared_error', weight_name='weight' if use_weight else None)
inits = ['intercept', 'coef']
inputs = onx_loss.graph.input
self.assertEqual(len(inputs), 3 if use_weight else 2)
train_session = OrtGradientOptimizer(onx_loss,
inits,
learning_rate=1e9)
self.assertRaise(lambda: train_session.get_state(), AttributeError)
if use_weight:
self.assertRaise(
lambda: train_session.fit(X_train,
y_train.reshape((-1, 1)),
w_train.reshape((-1, 1)),
use_numpy=False), ConvergenceError)
else:
self.assertRaise(
lambda: train_session.fit(
X_train, y_train.reshape((-1, 1)), use_numpy=False),
ConvergenceError)
state_tensors = train_session.get_state()
self.assertEqual(len(state_tensors), 2)
r = repr(train_session)
self.assertIn("OrtGradientOptimizer(model_onnx=", r)
self.assertIn("learning_rate='invscaling'", r)
losses = train_session.train_losses_
self.assertGreater(len(losses), 1)
if any(map(numpy.isnan, losses)):
raise AssertionError(losses)
def test_ort_gradient_optimizers_use_numpy_saved(self):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X, y = make_regression( # pylint: disable=W0632
100,
n_features=10,
bias=2,
random_state=0)
X = X.astype(numpy.float32)
y = y.astype(numpy.float32)
X_train, _, y_train, __ = train_test_split(X, y)
reg = LinearRegression()
reg.fit(X_train, y_train)
reg.coef_ = reg.coef_.reshape((1, -1))
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearRegressor'})
set_model_props(onx, {'info': 'unit test'})
onx_loss = add_loss_output(onx)
inits = ['intercept', 'coef']
temp = get_temp_folder(__file__, "temp_OrtGradientOptimizer")
filename = os.path.join(temp, "saved.onnx")
train_session = OrtGradientOptimizer(onx_loss,
inits,
learning_rate=1e-3,
saved_gradient=filename)
self.assertRaise(lambda: train_session.get_state(), AttributeError)
train_session.fit(X_train, y_train, use_numpy=True)
state_tensors = train_session.get_state()
self.assertEqual(len(state_tensors), 2)
r = repr(train_session)
self.assertIn("OrtGradientOptimizer(model_onnx=", r)
self.assertIn("learning_rate='invscaling'", r)
losses = train_session.train_losses_
self.assertGreater(len(losses), 1)
self.assertFalse(any(map(numpy.isnan, losses)))
self.assertExists(filename)
# It follows :epkg:`scikit-learn` API (see `SGDRegressor
# <https://scikit-learn.org/stable/modules/
# generated/sklearn.linear_model.SGDRegressor.html>`_.
# The gradient graph is not available at this stage.
train_session = OrtGradientOptimizer(onx_train,
list(weights),
device=device,
verbose=1,
learning_rate=1e-2,
warm_start=False,
max_iter=200,
batch_size=10,
saved_gradient="saved_gradient.onnx")
train_session.fit(X, y)
######################################
# And the trained coefficient are...
state_tensors = train_session.get_state()
pprint(["trained coefficients:", state_tensors])
print("last_losses:", train_session.train_losses_[-5:])
min_length = min(len(train_session.train_losses_), len(lr.loss_curve_))
df = DataFrame({
'ort losses': train_session.train_losses_[:min_length],
'skl losses': lr.loss_curve_[:min_length]
})
df.plot(title="Train loss against iterations")
def benchmark(N=1000, n_features=100, hidden_layer_sizes="50,10", max_iter=1000,
learning_rate_init=1e-4, batch_size=100, run_skl=True,
device='cpu', opset=14):
"""
Compares :epkg:`onnxruntime-training` to :epkg:`scikit-learn` for
training. Training algorithm is SGD.
:param N: number of observations to train on
:param n_features: number of features
:param hidden_layer_sizes: hidden layer sizes, comma separated values
:param max_iter: number of iterations
:param learning_rate_init: initial learning rate
:param batch_size: batch size
:param run_skl: train scikit-learn in the same condition (True) or
just walk through one iterator with *scikit-learn*
:param device: `'cpu'` or `'cuda'`
:param opset: opset to choose for the conversion
"""
N = int(N)
n_features = int(n_features)
max_iter = int(max_iter)
learning_rate_init = float(learning_rate_init)
batch_size = int(batch_size)
run_skl = run_skl in (1, True, '1', 'True')
print("N=%d" % N)
print("n_features=%d" % n_features)
print(f"hidden_layer_sizes={hidden_layer_sizes!r}")
print("max_iter=%d" % max_iter)
print(f"learning_rate_init={learning_rate_init:f}")
print("batch_size=%d" % batch_size)
print(f"run_skl={run_skl!r}")
print(f"opset={opset!r}")
print(f"device={device!r}")
print('------------------')
if not isinstance(hidden_layer_sizes, tuple):
hidden_layer_sizes = tuple(map(int, hidden_layer_sizes.split(",")))
X, y = make_regression(N, n_features=n_features, bias=2)
X = X.astype(numpy.float32)
y = y.astype(numpy.float32)
X_train, X_test, y_train, y_test = train_test_split(X, y)
nn = MLPRegressor(hidden_layer_sizes=hidden_layer_sizes,
max_iter=max_iter if run_skl else 1,
solver='sgd', learning_rate_init=learning_rate_init,
n_iter_no_change=max_iter, batch_size=batch_size)
begin = time.perf_counter()
with warnings.catch_warnings():
warnings.simplefilter('ignore')
nn.fit(X_train, y_train)
dur_skl = time.perf_counter() - begin
print("time_skl=%r, mean_squared_error=%r" % (
dur_skl, mean_squared_error(y_train, nn.predict(X_train))))
# conversion to ONNX
onx = to_onnx(nn, X_train[:1].astype(numpy.float32), target_opset=opset)
# add loss
onx_train = add_loss_output(onx)
# list of weights
weights = get_train_initializer(onx)
print('weights:', list(sorted(weights)))
# training
print(f"device={device!r} get_device()={get_device()!r}")
#######################################
# The training session.
train_session = OrtGradientOptimizer(
onx_train, list(weights), device=device, verbose=0,
learning_rate=learning_rate_init,
warm_start=False, max_iter=max_iter, batch_size=batch_size)
begin = time.perf_counter()
train_session.fit(X, y)
dur_ort = time.perf_counter() - begin
print("time_skl=%r, mean_squared_error=%r" % (
dur_skl, mean_squared_error(y_train, nn.predict(X_train))))
print("time_ort=%r, last_trained_error=%r" % (
dur_ort, train_session.train_losses_[-1]))
def wtest_ort_gradient_optimizers_binary(self, use_weight=False):
from onnxcustom.utils.orttraining_helper import add_loss_output
from onnxcustom.training.optimizers import OrtGradientOptimizer
X, y = make_classification( # pylint: disable=W0632
100, n_features=10, random_state=0)
X = X.astype(numpy.float32)
y = y.astype(numpy.int64)
w = (numpy.random.rand(X.shape[0]) + 1).astype(numpy.float32)
X_train, _, y_train, __, w_train, ___ = train_test_split(X, y, w)
reg = SGDClassifier(loss='log')
reg.fit(X_train, y_train)
onx = to_onnx(reg,
X_train,
target_opset=opset,
black_op={'LinearClassifier'},
options={'zipmap': False})
onx2 = load_onnx(BytesIO(onx.SerializeToString()))
set_model_props(onx, {'info': 'unit test'})
onx_loss = add_loss_output(
onx,
'log',
output_index=1,
weight_name='weight' if use_weight else None)
inits = ['intercept', 'coef']
inputs = onx_loss.graph.input
self.assertEqual(len(inputs), 3 if use_weight else 2)
dt = inputs[1].type.tensor_type.elem_type
self.assertEqual(TensorProto.INT64, dt) # pylint: disable=E1101
train_session = OrtGradientOptimizer(onx_loss,
inits,
learning_rate=1e-3)
self.assertRaise(lambda: train_session.get_state(), AttributeError)
if use_weight:
train_session.fit(X_train,
y_train.reshape((-1, 1)),
w_train.reshape((-1, 1)),
use_numpy=False)
else:
train_session.fit(X_train,
y_train.reshape((-1, 1)),
use_numpy=False)
state_tensors = train_session.get_state()
self.assertEqual(len(state_tensors), 2)
r = repr(train_session)
self.assertIn("OrtGradientOptimizer(model_onnx=", r)
self.assertIn("learning_rate='invscaling'", r)
losses = train_session.train_losses_
self.assertGreater(len(losses), 1)
self.assertFalse(any(map(numpy.isnan, losses)))
# get_trained_weight
trained_onnx = train_session.get_trained_onnx(model=onx2)
sess = InferenceSession(onx2.SerializeToString(),
providers=['CPUExecutionProvider'])
got1 = sess.run(None, {'X': X_train})
sess = InferenceSession(trained_onnx.SerializeToString(),
providers=['CPUExecutionProvider'])
got2 = sess.run(None, {'X': X_train})
self.assertEqual(len(got1), len(got2))
self.assertEqual(got1[0].shape, got2[0].shape)
# state
state = train_session.get_state()
self.assertIsInstance(state, dict)
train_session.set_state(state)