def evaluate_generated_graphs(generated_graphs, termination, nlls, start_time,
ts_properties, generation_batch_idx):
""" Computes molecular properties for input set of generated graphs, saves
results to CSV, and writes `generated_mols` to disk as a SMILES file.
Properties are expensive to calculate, so only done when
`gen_batch_idx` == 0 (i.e. for the first batch of generated molecules).
Args:
generated_graphs (list) : Contains `GenerationGraph`s.
termination (torch.Tensor) : Molecular termination details; contains 1 at
index if graph from `generated_mols` was "properly" terminated, 0
otherwise.
nlls (torch.Tensor) : Contains final NLL of each item in `generated_mols`.
start_time (time) : Program start time.
ts_properties (dict) : Contains training set properties.
gen_batch_idx (int) : Generation batch index.
"""
epoch_key = util.get_last_epoch()
if generation_batch_idx == 0:
# calculate molecular properties of generated set
prop_dict = get_molecular_properties(molecules=generated_graphs,
epoch_key=epoch_key,
termination=termination)
else:
prop_dict = {} # initialize the property dictionary
# add a few additional properties to the propery dictionary
prop_dict[(epoch_key, "final_nll")] = nlls
prop_dict[(epoch_key, "run_time")] = round(time.time() - start_time, 2)
# output evaluation metrics to CSV
output = C.job_dir
# calculate validity list now, so as not to write to CSV in previous step
epoch_id = epoch_key[6:] + "_" + str(generation_batch_idx)
fraction_valid, validity_tensor = util.write_molecules(
molecules=generated_graphs, final_nlls=nlls, epoch=epoch_id)
# add these validity properties to the property dictionary
prop_dict[(epoch_key, "fraction_valid")] = fraction_valid
prop_dict[(epoch_key, "validity_tensor")] = validity_tensor
# write these properties to disk, only for the first generation batch
if generation_batch_idx == 0:
util.properties_to_csv(prop_dict=prop_dict,
csv_filename=f"{output}generation.csv",
epoch_key=epoch_key,
append=True)
# join ts properties with prop_dict for plotting
merged_properties = {**prop_dict, **ts_properties}
# plot properties for this epoch
plot_filename = f"{output}generation/features{epoch_key[6:]}.png"
plot_molecular_properties(properties_dict=merged_properties,
plot_filename=plot_filename)
def create_output_files(self) -> None:
"""
Creates output files (with appropriate headers) for new (i.e. non-restart) jobs.
If restart a job, all new output will be appended to existing output files.
"""
if not self.constants.restart:
print("* Touching output files.", flush=True)
# begin writing `generation.log` file
csv_path_and_filename = self.constants.job_dir + "generation.log"
util.properties_to_csv(prop_dict=self.ts_properties,
csv_filename=csv_path_and_filename,
epoch_key="Training set",
append=False)
# begin writing `convergence.log` file
util.write_model_status(append=False)
# create `generation/` subdirectory to write generation output to
os.makedirs(self.constants.job_dir + "generation/", exist_ok=True)
