def evaluate_model(self):
""" Evaluates model by calculating the UC-JSD from generated structures.
Saves model scores in `validation.csv` and then saves model state.
"""
self.model.eval() # sets layers to eval mode (e.g. norm, dropout)
with torch.no_grad(): # deactivates autograd engine
# generate graphs required for model evaluation
# note that evaluation of the generated graphs happens in
# `generate_graphs()`, and molecules are saved as `self` attributes
self.generate_graphs(n_samples=self.C.n_samples, evaluation=True)
print("* Evaluating model.", flush=True)
anal.evaluate_model(valid_dataloader=self.valid_dataloader,
train_dataloader=self.train_dataloader,
nll_per_action=self.nll_per_action,
model=self.model)
self.nll_per_action = None # don't need anymore
print(f"* Saving model state at Epoch {self.current_epoch}.", flush=True)
# `pickle.HIGHEST_PROTOCOL` good for large objects
model_path_and_filename = (self.C.job_dir + f"model_restart_{self.current_epoch}.pth")
torch.save(obj=self.model,
f=model_path_and_filename,
pickle_protocol=pickle.HIGHEST_PROTOCOL)
def testing_phase(self):
""" Evaluates model using test set data.
"""
self.test_dataloader = self.get_dataloader(self.test_h5_path,
"test set")
self.get_ts_properties()
self.restart_epoch = util.get_restart_epoch()
print(
f"* Loading model from previous saved state (Epoch {self.restart_epoch}).",
flush=True)
self.model = torch.load(self.C.job_dir +
f"model_restart_{self.restart_epoch}.pth")
self.model.eval()
with torch.no_grad():
self.generate_graphs(n_samples=self.C.n_samples)
print("* Evaluating model.", flush=True)
anal.evaluate_model(valid_dataloader=self.test_dataloader,
train_dataloader=self.train_dataloader,
nll_per_action=self.nll_per_action,
model=self.model)
self.print_time_elapsed()