def __init__(self, metadata, session=None):
super(LogicModel, self).__init__(metadata)
# freiburg
perf_fn = str(Path(__file__).parents[4] / "perf_matrix.csv")
feat_fn = str(Path(__file__).parents[4] / "meta_features.csv")
af = AFCsvFacade(perf_fn=perf_fn, feat_fn=feat_fn)
config = {
'StandardScaler': False,
'fgroup_all': True,
'imputer_strategy': 'mean',
'pca': False,
'selector': 'PairwiseClassifier',
'classifier': 'RandomForest',
'rf:bootstrap': False,
'rf:criterion': 'gini',
'rf:max_depth': 132,
'rf:max_features': 'log2',
'rf:min_samples_leaf': 3,
'rf:min_samples_split': 3,
'rf:n_estimators': 68
}
LOGGER.info('--------- Fitting AF model ----------')
# fit AF using a loaded configuration on all data!
af.fit(config=config)
LOGGER.info('--------- Model.metadata ----------')
LOGGER.info('path: %s', self.metadata.get_dataset_name())
LOGGER.info('shape: %s', self.metadata.get_tensor_size(0))
LOGGER.info('size: %s', self.metadata.size())
LOGGER.info('num_class: %s', self.metadata.get_output_size())
test_metadata_filename = self.metadata.get_dataset_name().replace(
'train', 'test') + '/metadata.textproto'
self.num_test = [
int(line.split(':')[1])
for line in open(test_metadata_filename, 'r').readlines()[:3]
if 'sample_count' in line
][0]
LOGGER.info('num_test: %d', self.num_test)
self.timers = {
'train': skeleton.utils.Timer(),
'test': skeleton.utils.Timer()
}
self.info = {
'dataset': {
'path': self.metadata.get_dataset_name(),
'shape': self.metadata.get_tensor_size(0),
'size': self.metadata.size(),
'num_class': self.metadata.get_output_size()
},
'loop': {
'epoch': 0,
'test': 0,
'best_score': 0.0
},
'condition': {
'first': {
'train': True,
'valid': True,
'test': True
}
},
'terminate': False
}
# TODO: adaptive logic for hyper parameter
# 支持导入配置文件，可修改config_path
# config_path = 'AutoDL_sample_code_submission/configs/configs_deepwis/config_with_chucky_freiburg_params_and_dataloading.yaml'
# default_path = 'AutoDL_sample_code_submission/configs/config_deepblue.yaml'
# print('Path: ', config_path)
# get feature vector containing num_channels, num_classes, num_train, resolution_0
train_metadata_filename = self.metadata.get_dataset_name(
) + '/metadata.textproto'
num_train = \
[int(line.split(':')[1]) for line in open(train_metadata_filename, 'r').readlines()[:3] if
'sample_count' in line][0]
LOGGER.info('num_test: %d', num_train)
_, resolution_0, _, num_channels = self.metadata.get_tensor_shape()
num_classes = self.info['dataset']['num_class']
feature = [num_channels, num_classes, num_train, resolution_0]
pred_config_name = af.predict(
np.array(feature))["pseudo_instance"][0][0]
# pred_config_name = AFCsvFacade.load_and_predict(vec=np.array(feature), load_fn=model_fn)
LOGGER.info("AF suggesting to use config: {}".format(pred_config_name))
config_path = Path(__file__).parents[
4] / "configs" / "effnet_optimized_per_dataset_new_cs_new_data_03_14" / pred_config_name
config_path = config_path.with_suffix(".yaml")
try:
with config_path.open() as in_stream:
model_config = yaml.safe_load(in_stream)
LOGGER.info("Using config: {}".format(config_path))
except:
with config_path.with_name("default.yaml").open() as in_stream:
model_config = yaml.safe_load(in_stream)
LOGGER.info("Using default config.")
self.hyper_params = model_config["autocv"]
skip_valid_after_test = min(
10, max(3, int(self.info['dataset']['size'] // 1000)))
self.hyper_params["conditions"][
"skip_valid_after_test"] = skip_valid_after_test
self.checkpoints = []
LOGGER.info('[init] build')
self.build()
LOGGER.info('[init] session')
self.dataloaders = {'train': None, 'valid': None, 'test': None}
self.is_skip_valid = True
self.changed_model = False
self.save_res = False
self.preds = []
self.results = []
# self.last_best_score=0
# self.is_fuse=False
LOGGER.info('[init] done')
from autofolio.facade.af_csv_facade import AFCsvFacade
__author__ = "Marius Lindauer"
__license__ = "BSD"
__version__ = "2.1.0"
perf_fn = "/home/lindauer/git/AutoDLComp19/experiments/new_config_space_after_March_4/eval_new_cs_old_data_approx_600/perf_matrix.csv"
feat_fn = "/home/lindauer/git/AutoDLComp19/src/meta_features/non-nn/old_data/meta_features.csv"
# will be created (or overwritten) by AutoFolio
old_model_fn = "af_model.pkl"
new_model_fn = "af_model_final.pkl"
config_fn = "config.pkl"
af = AFCsvFacade(perf_fn=perf_fn, feat_fn=feat_fn)
#with open(old_model_fn, "br") as fp:
#_, _, _, _, config = pickle.load(fp)
#with open(config_fn, "bw") as fp:
# pickle.dump(config, fp)
with open(config_fn, "br") as fp:
config = pickle.load(fp)
print(config)
# evaluate configuration using a 10-fold cross validation
score = af.cross_validation(config=config)
import numpy as np from autofolio.facade.af_csv_facade import AFCsvFacade __author__ = "Marius Lindauer" __license__ = "BSD" __version__ = "2.1.0" perf_fn = "perf_matrix.csv" feat_fn = "meta_features.csv" # will be created (or overwritten) by AutoFolio model_fn = "af_model.pkl" af = AFCsvFacade(perf_fn=perf_fn, feat_fn=feat_fn) # fit AutoFolio; will use default hyperparameters of AutoFolio af.fit() # tune AutoFolio's hyperparameter configuration for 4 seconds config = af.tune(wallclock_limit=360) # evaluate configuration using a 10-fold cross validation score = af.cross_validation(config=config) # re-fit AutoFolio using the (hopefully) better configuration # and save model to disk af.fit(config=config, save_fn=model_fn) # load AutoFolio model and
import numpy as np from autofolio.facade.af_csv_facade import AFCsvFacade __author__ = "Marius Lindauer" __license__ = "BSD" __version__ = "2.1.0" # will be created (or overwritten) by AutoFolio model_fn = "af_model_final.pkl" # load AutoFolio model and # get predictions for new meta-feature vector pred = AFCsvFacade.load_and_predict(vec=np.array([3.,102., 6149., -1.,]), load_fn=model_fn) print(pred)
import numpy as np from autofolio.facade.af_csv_facade import AFCsvFacade __author__ = "Marius Lindauer" __license__ = "BSD" __version__ = "2.1.0" perf_fn = "perf.csv" feat_fn = "feats.csv" # will be created (or overwritten) by AutoFolio model_fn = "af_model.pkl" af = AFCsvFacade(perf_fn=perf_fn, feat_fn=feat_fn) # load AutoFolio model and # get predictions for new meta-feature vector pred = AFCsvFacade.load_and_predict(vec=np.array([1.]), load_fn=model_fn) print(pred)
import pickle
import numpy as np
from autofolio.facade.af_csv_facade import AFCsvFacade
__author__ = "Marius Lindauer"
__license__ = "BSD"
__version__ = "2.1.0"
perf_fn = "perf_matrix.csv"
feat_fn = "meta_features.csv"
af = AFCsvFacade(perf_fn=perf_fn, feat_fn=feat_fn)
config = {
'StandardScaler': False,
'fgroup_all': True,
'imputer_strategy': 'mean',
'pca': False,
'selector': 'PairwiseClassifier',
'classifier': 'RandomForest',
'rf:bootstrap': False,
'rf:criterion': 'gini',
'rf:max_depth': 132,
'rf:max_features': 'log2',
'rf:min_samples_leaf': 3,
'rf:min_samples_split': 3,
'rf:n_estimators': 68
}
# fit AF using a loaded configuration on all data!