def run_automl(X, y, name, input=None, output=CategoricalVector()):
telegram = TelegramLogger(
token=os.environ["TOKEN"],
channel="@autogoal_board",
name=name,
)
console = ConsoleLogger()
progress = ProgressLogger()
automl = AutoML(
search_iterations=1000,
metalearning_log=True,
search_kwargs=dict(search_timeout=2 * 60 * 60, pop_size=50),
errors="ignore",
input=input,
output=output,
cross_validation_steps=1,
)
automl.fit(X, y, logger=[telegram, console, progress])
class CustomLogger(Logger):
def error(self, e: Exception, solution):
if e and solution:
with open("haha_errors.log", "a") as fp:
fp.write(f"solution={repr(solution)}\nerror={repr(e)}\n\n")
def update_best(self, new_best, new_fn, *args):
with open("haha.log", "a") as fp:
fp.write(f"solution={repr(new_best)}\nfitness={new_fn}\n\n")
# Basic logging configuration.
logger = MemoryLogger()
loggers = [ProgressLogger(), ConsoleLogger(), logger]
if args.token:
from autogoal.contrib.telegram import TelegramLogger
telegram = TelegramLogger(
token=args.token,
name=f"HAHA",
channel=args.channel,
)
loggers.append(telegram)
# Finally, loading the HAHA dataset, running the `AutoML` instance,
# and printing the results.
X_train, y_train, X_test, y_test = haha.load(max_examples=args.examples)
from autogoal.contrib.keras import KerasSequenceClassifier
from autogoal.contrib.torch import BertTokenizeEmbedding
from autogoal.datasets import haha
from autogoal.kb import CategoricalVector, List, Sentence, Tuple
from autogoal.ml import AutoML
from autogoal.search import ConsoleLogger, ProgressLogger
classifier = AutoML(
input=List(Sentence()),
output=CategoricalVector(),
registry=[KerasSequenceClassifier, BertTokenizeEmbedding],
# search_kwargs=dict(memory_limit=4 * 1024 ** 3, evaluation_timeout=60),
search_kwargs=dict(memory_limit=0, evaluation_timeout=0),
)
Xtrain, Xtest, ytrain, ytest = haha.load(max_examples=10)
# embedding = BertEmbedding()
# tokens = embedding.run(Xtrain)
# classifier = KerasSequenceClassifier().sample()
# classifier.run((tokens, ytrain))
classifier.fit(Xtrain, ytrain, logger=[ConsoleLogger(), ProgressLogger()])
import numpy as np
from autogoal.ml import AutoML
from autogoal.contrib.keras import KerasImageClassifier, KerasImagePreprocessor
from autogoal.datasets import cifar10
from autogoal.kb import CategoricalVector, Tensor4
from autogoal.search import ConsoleLogger, ProgressLogger
automl = AutoML(
input=Tensor4(),
output=CategoricalVector(),
registry=[KerasImageClassifier],
# registry=[KerasImageClassifier, KerasImagePreprocessor],
cross_validation_steps=1,
search_kwargs=dict(
pop_size=20,
search_timeout=24 * 60 * 60,
evaluation_timeout=0,
memory_limit=0,
save=False,
),
search_iterations=1000,
validation_split=1 / 6)
Xtrain, ytrain, Xtest, ytest = cifar10.load()
X = np.vstack((Xtrain, Xtest))
y = np.hstack((ytrain, ytest))
automl.fit(X, y, logger=[ConsoleLogger(), ProgressLogger()])
print(solution)
except ValueError:
continue
# To evaluate how good a formula is, we simply feed the expression instance
# with a sequence of numbers from 1 to 9. If the expression requires more
# than 9 digits, it results in an error. The actual value of performing
# corresponding operations is done in the `__call__` method of the expression classes.
def evaluate(expr):
def stream():
for i in range(1, 10):
yield i
raise ValueError("Too many values asked")
return expr(stream())
# We will run 1000 iterations of each search strategy to compare their long-term performance.
search_rand = RandomSearch(grammar, evaluate, errors='ignore')
best_rand, best_fn_rand = search_rand.run(1000, logger=[ConsoleLogger(), ProgressLogger()])
search_pe = PESearch(grammar, evaluate, pop_size=10, errors='ignore')
best_pe, best_fn_pe = search_pe.run(1000, logger=[ConsoleLogger(), ProgressLogger()])
# And here are the results.
print(best_rand, best_fn_rand)
print(best_pe, best_fn_pe)
from autogoal.contrib.keras import KerasClassifier
from autogoal.datasets import cars
from autogoal.kb import CategoricalVector, MatrixContinuousDense
from autogoal.ml import AutoML
from autogoal.search import ConsoleLogger, ProgressLogger
classifier = AutoML(
input=MatrixContinuousDense(),
registry=[KerasClassifier],
search_kwargs=dict(memory_limit=0, evaluation_timeout=0),
)
X, y = cars.load()
classifier.fit(X, y, logger=[ConsoleLogger(), ProgressLogger()])
# with a sequence of numbers from 1 to 9. If the expression requires more
# than 9 digits, it results in an error. The actual value of performing
# corresponding operations is done in the `__call__` method of the expression classes.
def evaluate(expr):
def stream():
for i in range(1, 10):
yield i
raise ValueError("Too many values asked")
return expr(stream())
# We will run 1000 iterations of each search strategy to compare their long-term performance.
search_rand = RandomSearch(grammar, evaluate, errors='ignore')
best_rand, best_fn_rand = search_rand.run(
1000, logger=[ConsoleLogger(), ProgressLogger()])
search_pe = PESearch(grammar, evaluate, pop_size=10, errors='ignore')
best_pe, best_fn_pe = search_pe.run(1000,
logger=[ConsoleLogger(),
ProgressLogger()])
# And here are the results.
print(best_rand, best_fn_rand)
print(best_pe, best_fn_pe)
try:
with open(fR"{path}/binary_X",
'rb') as xfd, open(fR"{path}/binary_Y", 'rb') as yfd:
X = pickle.load(xfd)
y = pickle.load(yfd)
return X, y
except Exception as e:
#TODO: implement corpus reading from directories
print(e)
pass
if __name__ == "__main__":
g = generate_cfg(SklearnNLPClassifier)
X, y = load_movie_reviews(100)
# X, y = load_corpus("examples/Revolico")
def fitness(pipeline):
pipeline.fit(X, y)
score = pipeline.score(X, y)
return score
search = RandomSearch(g,
fitness,
random_state=0,
errors='warn',
evaluation_timeout=100)
result = search.run(50, logger=ProgressLogger())
# care of train/test splitting, fitting a pipeline in the training set and computing # the accuracy on the test set. fitness_fn = movie_reviews.make_fn(examples=100) # ### Random search # # The `RandomSearch` strategy simply calls `grammar.sample()` a bunch of times # and stores the best performing pipeline. It has no intelligence whatsoever, # but it serves as a good baseline implementation. # # We will run it for a total of `1000` fitness evaluations, or equivalently, a total # of `1000` different random pipelines. To see what's actually going on we will use # the wonderfull `enlighten` library through our implementation `EnlightenLogger`. logger = ProgressLogger() random_search = RandomSearch(grammar, fitness_fn, random_state=0) best_rand, fn_rand = random_search.run(1000, logger=logger) # !!! note # For reproducibility purposes we can pass a fixed random seed in `random_state`. # # ### Evolutionary Search # # Random search is fun, but to search with purpose, we need a more intelligent sampling # strategy. The `PESearch` (short for Probabilistic Evolutionary Search, phew), does just that. # It starts with a random sampling strategy, but as it evaluates more pipelines, it modifies # an probabilistic sampling model so that pipelines similar to the best ones found are more # commonly sampled. #
from autogoal.contrib.sklearn import SklearnClassifier
from autogoal.grammar import generate_cfg
from autogoal.search import RandomSearch, ProgressLogger
from sklearn.datasets import make_classification
g = generate_cfg(SklearnClassifier)
X, y = make_classification()
print(g)
def fitness(pipeline):
pipeline.fit(X, y)
return pipeline.score(X, y)
search = RandomSearch(g, fitness, random_state=0, errors='warn')
search.run(1000, logger=ProgressLogger())
# care of train/test splitting, fitting a pipeline in the training set and computing # the accuracy on the test set. fitness_fn = movie_reviews.make_fn(examples=100) # ### Random search # # The `RandomSearch` strategy simply calls `grammar.sample()` a bunch of times # and stores the best performing pipeline. It has no intelligence whatsoever, # but it serves as a good baseline implementation. # # We will run it for a total of `1000` fitness evaluations, or equivalently, a total # of `1000` different random pipelines. To see what's actually going on we will use # the wonderfull `enlighten` library through our implementation `EnlightenLogger`. logger = ProgressLogger(log_solutions=True) random_search = RandomSearch(grammar, fitness_fn, random_state=0) best_rand, fn_rand = random_search.run(1000, logger=logger) # !!! note # For reproducibility purposes we can pass a fixed random seed in `random_state`. # # ### Evolutionary Search # # Random search is fun, but to search with purpose, we need a more intelligent sampling # strategy. The `PESearch` (short for Probabilistic Evolutionary Search, phew), does just that. # It starts with a random sampling strategy, but as it evaluates more pipelines, it modifies # an probabilistic sampling model so that pipelines similar to the best ones found are more # commonly sampled. #