def from_keras(
*,
model_creator,
logs_dir,
resources_per_trial,
name,
):
"""
Create an AutoEstimator for tensorflow keras.
:param model_creator: Tensorflow keras model creator function.
:param logs_dir: Local directory to save logs and results.
:param resources_per_trial: Dict. resources for each trial. e.g. {"cpu": 2}.
:param name: Name of the auto estimator.
:return: an AutoEstimator object.
"""
from zoo.automl.model import ModelBuilder
from zoo.automl.search import SearchEngineFactory
model_builder = ModelBuilder.from_tfkeras(model_creator=model_creator)
searcher = SearchEngineFactory.create_engine(
backend="ray",
logs_dir=logs_dir,
resources_per_trial=resources_per_trial,
name=name)
return AutoEstimator(model_builder=model_builder, searcher=searcher)
def test_fit_evaluate(self):
modelBuilder_keras = ModelBuilder.from_tfkeras(model_creator_keras)
model = modelBuilder_keras.build(config={
"lr": 1e-2,
"batch_size": 32,
"metric": "mse"
})
val_result = model.fit_eval(x=self.data["x"],
y=self.data["y"],
validation_data=(self.data["val_x"], self.data["val_y"]),
epochs=20)
assert val_result is not None
def test_fit_evaluate(self):
modelBuilder = ModelBuilder.from_pytorch(
model_creator=model_creator_pytorch,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator)
model = modelBuilder.build(config={
"lr": 1e-2,
"batch_size": 32,
})
val_result = model.fit_eval(x=self.data["x"],
y=self.data["y"],
validation_data=(self.data["val_x"],
self.data["val_y"]),
epochs=20)
assert val_result is not None
def from_torch(
*,
model_creator,
optimizer,
loss,
logs_dir,
resources_per_trial,
name,
):
"""
Create an AutoEstimator for torch.
:param model_creator: PyTorch model creator function.
:param optimizer: PyTorch optimizer creator function or pytorch optimizer name (string).
Note that you should specify learning rate search space with key as "lr" or LR_NAME
(from zoo.orca.automl.pytorch_utils import LR_NAME) if input optimizer name.
Without learning rate search space specified, the default learning rate value of 1e-3
will be used for all estimators.
:param loss: PyTorch loss instance or PyTorch loss creator function
or pytorch loss name (string).
:param logs_dir: Local directory to save logs and results.
:param resources_per_trial: Dict. resources for each trial. e.g. {"cpu": 2}.
:param name: Name of the auto estimator.
:return: an AutoEstimator object.
"""
from zoo.orca.automl.pytorch_utils import validate_pytorch_loss, \
validate_pytorch_optim
from zoo.automl.model import ModelBuilder
from zoo.automl.search import SearchEngineFactory
loss = validate_pytorch_loss(loss)
optimizer = validate_pytorch_optim(optimizer)
model_builder = ModelBuilder.from_pytorch(model_creator=model_creator,
optimizer_creator=optimizer,
loss_creator=loss)
searcher = SearchEngineFactory.create_engine(
backend="ray",
logs_dir=logs_dir,
resources_per_trial=resources_per_trial,
name=name)
return AutoEstimator(model_builder=model_builder, searcher=searcher)
def test_predict(self):
modelBuilder = ModelBuilder.from_pytorch(
model_creator=model_creator_pytorch,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator)
model = modelBuilder.build(config={
"lr": 1e-2,
"batch_size": 32,
})
model.fit_eval(x=self.data["x"],
y=self.data["y"],
validation_data=(self.data["val_x"],
self.data["val_y"]),
epochs=20)
pred = model.predict(x=self.data["val_x"])
try:
import onnx
import onnxruntime
pred_onnx = model.predict_with_onnx(x=self.data["val_x"])
np.testing.assert_almost_equal(pred, pred_onnx)
except ImportError:
pass
def prepare_searcher(data,
model_creator=linear_model_creator,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator,
feature_transformer=None,
recipe=SimpleRecipe(),
name="demo"):
modelBuilder = ModelBuilder.from_pytorch(model_creator=model_creator,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator)
searcher = SearchEngineFactory.create_engine(backend="ray",
logs_dir="~/zoo_automl_logs",
resources_per_trial={"cpu": 2},
name=name)
search_space = recipe.search_space(feature_transformer.get_feature_list())\
if feature_transformer else None
searcher.compile(data=data,
model_create_func=modelBuilder,
recipe=recipe,
feature_transformers=feature_transformer,
search_space=search_space)
return searcher
def test_evaluate(self):
modelBuilder = ModelBuilder.from_pytorch(
model_creator=model_creator_pytorch,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator)
model = modelBuilder.build(config={
"lr": 1e-2,
"batch_size": 32,
})
model.fit_eval(x=self.data["x"],
y=self.data["y"],
validation_data=(self.data["val_x"],
self.data["val_y"]),
epochs=20)
mse_eval = model.evaluate(x=self.data["val_x"], y=self.data["val_y"])
try:
import onnx
import onnxruntime
mse_eval_onnx = model.evaluate_with_onnx(x=self.data["val_x"],
y=self.data["val_y"])
np.testing.assert_almost_equal(mse_eval, mse_eval_onnx)
except ImportError:
pass
# incremental training test
model.fit_eval(x=self.data["x"],
y=self.data["y"],
validation_data=(self.data["val_x"],
self.data["val_y"]),
epochs=20)
mse_eval = model.evaluate(x=self.data["val_x"], y=self.data["val_y"])
try:
import onnx
import onnxruntime
mse_eval_onnx = model.evaluate_with_onnx(x=self.data["val_x"],
y=self.data["val_y"])
np.testing.assert_almost_equal(mse_eval, mse_eval_onnx)
except ImportError:
pass
def get_linear_data(a, b, size):
x = np.arange(0, 10, 10 / size, dtype=np.float32)
y = a*x + b
return x, y
train_x, train_y = get_linear_data(2, 5, 1000)
df = pd.DataFrame({'x': train_x, 'y': train_y})
val_x, val_y = get_linear_data(2, 5, 400)
val_df = pd.DataFrame({'x': val_x, 'y': val_y})
return df, val_df
if __name__ == "__main__":
# 1. the way to enable auto tuning model from creators.
init_orca_context(init_ray_on_spark=True)
modelBuilder = ModelBuilder.from_pytorch(model_creator=model_creator,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator)
searcher = SearchEngineFactory.create_engine(backend="ray",
logs_dir="~/zoo_automl_logs",
resources_per_trial={"cpu": 2},
name="demo")
# pass input data, modelbuilder and recipe into searcher.compile. Note that if user doesn't pass
# feature transformer, the default identity feature transformer will be used.
df, val_df = get_data()
searcher.compile(df,
modelBuilder,
recipe=SimpleRecipe(),
feature_cols=["x"],
target_col="y",
def get_linear_data(a, b, size):
x = np.arange(0, 10, 10 / size, dtype=np.float32)
y = a * x + b
return x, y
train_x, train_y = get_linear_data(2, 5, 1000)
val_x, val_y = get_linear_data(2, 5, 400)
data = {'x': train_x, 'y': train_y, 'val_x': val_x, 'val_y': val_y}
return data
if __name__ == "__main__":
# 1. the way to enable auto tuning model from creators.
init_orca_context(init_ray_on_spark=True)
modelBuilder = ModelBuilder.from_pytorch(
model_creator=model_creator,
optimizer_creator=optimizer_creator,
loss_creator=loss_creator)
searcher = SearchEngineFactory.create_engine(
backend="ray",
logs_dir="~/zoo_automl_logs",
resources_per_trial={"cpu": 2},
name="demo")
# pass input data, modelbuilder and recipe into searcher.compile. Note that if user doesn't pass
# feature transformer, the default identity feature transformer will be used.
data = get_data()
searcher.compile(data=data,
model_create_func=modelBuilder,
recipe=SimpleRecipe())