def predictor_from_file(config_path, model_path):
with open(config_path, "rb") as f:
config_dict = json.load(f)
if config_dict["model_type"] == "neural_bandit":
# initialize the pytorch model and put it in `eval` mode
model = embed_dnn.EmbedDnn(**config_dict["model_spec"])
model.load_state_dict(torch.load(model_path))
model.eval()
else:
with open(model_path, "rb") as f:
model = pickle.load(f)
# initialize transforms
transforms = {}
for feature_name, transform_spec in config_dict["transforms"].items():
if transform_spec is None:
# id lists don't have transforms
transform = None
elif transform_spec["name"] == "StandardScaler":
transform = preprocessing.StandardScaler()
transform.mean_ = np.array(transform_spec["mean"])
transform.scale_ = np.array(transform_spec["scale"])
transform.var_ = np.array(transform_spec["var"])
elif transform_spec["name"] == "OneHotEncoder":
transform = preprocessing.OneHotEncoder()
transform.sparse = transform_spec["sparse"]
transform.categories_ = np.array(transform_spec["categories"])
else:
raise Exception(
f"Don't know how to load transform_spec of type {transform_spec['name']}"
)
transforms[feature_name] = transform
# initialize imputers
imputers = {}
for feature_name, imputer_spec in config_dict["imputers"].items():
if imputer_spec is None:
# categoricals & id lists don't have imputers
imputer = None
else:
imputer = SimpleImputer()
imputer.set_params(**imputer_spec["parameters"])
imputer.statistics_ = np.array(imputer_spec["statistics"])
imputers[feature_name] = imputer
return BanditPredictor(
feature_config=config_dict["feature_config"],
float_feature_order=config_dict["float_feature_order"],
id_feature_order=config_dict["id_feature_order"],
id_feature_str_to_int_map=config_dict["id_feature_str_to_int_map"],
transforms=transforms,
imputers=imputers,
model=model,
model_type=config_dict["model_type"],
reward_type=config_dict["reward_type"],
model_spec=config_dict["model_spec"],
dense_features_to_use=config_dict["dense_features_to_use"],
)
def predict_catkit_demo(images):
"""Return a prediction of adsorption energies for structures generated with
CatKitDemo.
Parameters
----------
images : list
List of atoms objects representing adsorbate-surface structures.
model : str
Path and filename of Catlearn model pickle.
"""
model_ref = {'H': 'H2', 'O': 'H2O, H2', 'C': 'CH4, H2'}
# Make list of strings showing the references.
display_ref = []
for atoms in images:
try:
initial_state = [
model_ref[s] for s in ase.atoms.string2symbols(
atoms.info['key_value_pairs']['species'])
]
except KeyError:
return {}
display_ref.append('*, ' + ', '.join(list(np.unique(initial_state))))
images = autogen_info(images)
gen = FeatureGenerator(nprocs=1)
train_fpv = default_fingerprinters(gen, 'adsorbates')
train_fpv = [
gen.mean_chemisorbed_atoms, gen.count_chemisorbed_fragment,
gen.count_ads_atoms, gen.count_ads_bonds, gen.ads_av, gen.ads_sum,
gen.bulk, gen.term, gen.strain, gen.mean_surf_ligands, gen.mean_site,
gen.median_site, gen.max_site, gen.min_site, gen.sum_site,
gen.generalized_cn, gen.en_difference_ads, gen.en_difference_chemi,
gen.en_difference_active, gen.db_size, gen.delta_energy
]
matrix = gen.return_vec(images, train_fpv)
feature_index = np.load(clean_index_name)
clean_feature_mean = np.load(clean_mean)
impute = SimpleImputer(strategy='mean')
impute.statistics_ = clean_feature_mean
new_data = impute.transform(matrix[:, feature_index])
prediction = gp.predict(new_data,
get_validation_error=False,
get_training_error=False,
uncertainty=True)
output = {
'mean': list(prediction['prediction']),
'uncertainty': list(prediction['uncertainty']),
'references': display_ref
}
return output
def predictor_from_file(config_path, net_path):
with open(config_path, "rb") as f:
config_dict = json.load(f)
# initialize the pytorch model and put it in `eval` mode
net = embed_dnn.EmbedDnn(**config_dict["net_spec"])
net.load_state_dict(torch.load(net_path))
net.eval()
# initialize transforms
transforms = {}
for feature_name, transform_spec in config_dict["transforms"].items():
if transform_spec is None:
# id lists don't have transforms
transform = None
elif transform_spec["name"] == "StandardScaler":
transform = preprocessing.StandardScaler()
transform.mean_ = np.array(transform_spec["mean"])
transform.scale_ = np.array(transform_spec["scale"])
transform.var_ = np.array(transform_spec["var"])
elif transform_spec["name"] == "OneHotEncoder":
transform = preprocessing.OneHotEncoder()
transform.sparse = transform_spec["sparse"]
transform.categories_ = np.array(transform_spec["categories"])
else:
raise Exception(
f"Don't know how to load transform_spec of type {transform_spec['name']}"
)
transforms[feature_name] = transform
# initialize imputers
imputers = {}
for feature_name, imputer_spec in config_dict["imputers"].items():
if imputer_spec is None:
# id lists don't have imputers
imputer = None
else:
imputer = SimpleImputer()
imputer.set_params(**imputer_spec["parameters"])
imputer.statistics_ = np.array(imputer_spec["statistics"])
imputers[feature_name] = imputer
return BanditPredictor(
experiment_specific_params=config_dict["experiment_specific_params"],
float_feature_order=config_dict["float_feature_order"],
id_feature_order=config_dict["id_feature_order"],
transforms=transforms,
imputers=imputers,
net=net,
net_spec=config_dict["net_spec"],
)
