params['trainer']['patience']):
print("Stopping at epoch {:2d}".format(epoch_num))
break
save_checkpoint(
model,
optimizer,
args.folder,
epoch_num,
val_metric_per_epoch,
is_best=int(
np.argmax(val_metric_per_epoch)) == (len(val_metric_per_epoch) -
1))
print("STOPPING. now running the best model on the validation set", flush=True)
# Load best
restore_best_checkpoint(model, args.folder)
model.eval()
val_probs = []
val_labels = []
for b, (time_per_batch, batch) in enumerate(time_batch(val_loader)):
with torch.no_grad():
batch = _to_gpu(batch)
output_dict = model(**batch)
val_probs.append(output_dict['label_probs'].detach().cpu().numpy())
val_labels.append(batch['label'].detach().cpu().numpy())
val_labels = np.concatenate(val_labels, 0)
val_probs = np.concatenate(val_probs, 0)
acc = float(np.mean(val_labels == val_probs.argmax(1)))
print("Final val accuracy is {:.3f}".format(acc))
np.save(os.path.join(args.folder, f'valpreds.npy'), val_probs)
is_best=int(
np.argmax(val_metric_per_epoch)) == (len(val_metric_per_epoch) -
1),
q_att1=q_att1,
a_att1=a_att1,
q_att2=q_att2,
a_att2=a_att2)
writer.add_scalar('val_loss', val_loss_avg, global_step=epoch_num)
writer.add_scalar('val_accuracy',
val_metric_per_epoch[-1],
global_step=epoch_num)
print("STOPPING. now running the best model on the validation set", flush=True)
logger.write("STOPPING. now running the best model on the validation set")
# Load best
restore_best_checkpoint(model, folder)
model.eval()
val_probs = []
val_labels = []
for b, (time_per_batch, batch) in enumerate(time_batch(val_loader)):
with torch.no_grad():
batch = _to_gpu(batch)
output_dict = model(**batch)
val_probs.append(output_dict['label_probs'].detach().cpu().numpy())
val_labels.append(batch['label'].detach().cpu().numpy())
val_labels = np.concatenate(val_labels, 0)
val_probs = np.concatenate(val_probs, 0)
acc = float(np.mean(val_labels == val_probs.argmax(1)))
print("Final val accuracy is {:.4f}".format(acc))
logger.write("Final val accuracy is {:.4f}".format(acc))
np.save(os.path.join(folder, f'valpreds.npy'), val_probs)