def get_prediction(model_path, unit_converter, molecules, labels,
extra_features):
model = load_model(model_path)
dataset = load_dataset(model_path, molecules, labels, extra_features)
data_loader = DataLoaderReactionNetwork(dataset,
batch_size=100,
shuffle=False)
feature_names = ["atom", "bond", "global"]
# evaluate
predictions = evaluate(model, feature_names, data_loader)
# in case some entry fail
if len(predictions) != len(dataset.failed):
pred = []
idx = 0
for failed in dataset.failed:
if failed:
pred.append(None)
else:
pred.append(predictions[idx] * unit_converter)
idx += 1
predictions = pred
else:
predictions = np.asarray(predictions) * unit_converter
return predictions
def main(
model_name="bdncm/20200808",
sdf_file="~/Applications/db_access/mol_builder/struct_rxn_ntwk_rgrn_qc.sdf",
label_file="~/Applications/db_access/mol_builder/label_rxn_ntwk_rgrn_qc.yaml",
feature_file="~/Applications/db_access/mol_builder/feature_rxn_ntwk_rgrn_qc.yaml",
error_file="~/Applications/db_access/mol_builder/post_analysis/evaluation_error.tsv",
charge_file="~/Applications/db_access/mol_builder/post_analysis/charges.tsv",
):
seed_torch()
dataset = load_dataset(model_name, sdf_file, label_file, feature_file)
# trainset, valset, testset = train_validation_test_split(
# dataset, validation=0.1, test=0.1
# )
trainset, valset, testset = train_validation_test_split_selected_bond_in_train(
dataset,
validation=0.1,
test=0.1,
selected_bond_type=(("H", "H"), ("H", "F"), ("F", "F")),
)
# data_loader = DataLoaderReactionNetwork(trainset, batch_size=100, shuffle=False)
# data_loader = DataLoaderReactionNetwork(valset, batch_size=100, shuffle=False)
data_loader = DataLoaderReactionNetwork(testset,
batch_size=100,
shuffle=False)
model = load_model(model_name)
# make predictions
feature_names = ["atom", "bond", "global"]
ids, targets, predictions, errors, species = evaluate(
model, feature_names, data_loader)
# sort by error
ids, targets, predictions, errors, species = zip(*sorted(
zip(ids, targets, predictions, errors, species), key=lambda x: x[3]))
df = pd.DataFrame({
"identifier": ids,
"target": targets,
"prediction": predictions,
"error": errors,
"species": species,
})
df.to_csv(to_path(error_file), sep="\t", index=False)
# charges
df = get_charges(label_file, feature_file)
df.to_csv(to_path(charge_file), sep="\t", index=False)
def main(
model_name="bdncm/20200808",
sdf_file="~/Applications/db_access/mol_builder/struct_rxn_ntwk_rgrn_qc.sdf",
label_file="~/Applications/db_access/mol_builder/label_rxn_ntwk_rgrn_qc.yaml",
feature_file="~/Applications/db_access/mol_builder/feature_rxn_ntwk_rgrn_qc.yaml",
feat_filename="~/Applications/db_access/mol_builder/post_analysis/feats.tsv",
meta_filename="~/Applications/db_access/mol_builder/post_analysis/feats_metadata.tsv",
):
seed_torch()
dataset = load_dataset(model_name, sdf_file, label_file, feature_file)
_, _, testset = train_validation_test_split(dataset,
validation=0.1,
test=0.1)
data_loader = DataLoaderReactionNetwork(testset,
batch_size=100,
shuffle=False)
# data_loader = DataLoaderReactionNetwork(dataset, batch_size=100, shuffle=False)
model = load_model(model_name)
# make predictions
feature_names = ["atom", "bond", "global"]
ids, targets, predictions, errors, species, features = evaluate(
model, feature_names, data_loader, compute_features=True)
df = pd.DataFrame(features)
df.to_csv(to_path(feat_filename), sep="\t", header=False, index=False)
# metadata
charges = get_charges(label_file, feature_file)
rct_charges = []
prdt1_charges = []
prdt2_charges = []
for i in ids:
c = charges[charges["identifier"] == i].to_dict("records")[0]
rct_charges.append(c["charge"])
prdt1_charges.append(c["product1 charge"])
prdt2_charges.append(c["product2 charge"])
df = pd.DataFrame({
"identifier": ids,
"target": targets,
"prediction": predictions,
"error": errors,
"species": species,
"reactant charge": rct_charges,
"product1 charge": prdt1_charges,
"product2 charge": prdt2_charges,
})
df.to_csv(to_path(meta_filename), sep="\t", index=False)
def main(
model_name="mesd/20200808",
sdf_file="/Users/mjwen/Applications/db_access/mol_builder/post_analysis/lbdc/struct.sdf",
label_file="/Users/mjwen/Applications/db_access/mol_builder/post_analysis/lbdc/label.yaml",
feature_file="/Users/mjwen/Applications/db_access/mol_builder/post_analysis/lbdc/feature.yaml",
feat_meta_prefix=f"~/Applications/db_access/mol_builder/post_analysis/lbdc",
):
seed_torch()
dataset = load_dataset(model_name, sdf_file, label_file, feature_file)
data_loader = DataLoaderReactionNetwork(dataset, batch_size=100, shuffle=False)
model = load_model(model_name, pretrained=False)
# make predictions
feature_names = ["atom", "bond", "global"]
ids, targets, predictions, broken_bonds, species, features = evaluate(
model, feature_names, data_loader
)
# write to file
for idx, ft in features.items():
fname = to_path(feat_meta_prefix).joinpath(f"feats_layer{idx}.tsv")
df = pd.DataFrame(ft)
df.to_csv(fname, sep="\t", header=False, index=False)
df = pd.DataFrame(
{
"identifier": ids,
"target": targets,
"prediction": predictions,
"broken_bonds": broken_bonds,
"species": species,
}
)
to_path(feat_meta_prefix).joinpath("feats_metadata.tsv")
df.to_csv(fname, sep="\t", index=False)