imgs_t.save(os.path.join(args.gen_img_dir, filename))
if i_iter % num_calmIoU == 0:
enc_shared.eval()
print('evaluating models ...')
for i_val, (images_val, labels_val) in tqdm(enumerate(val_loader)):
images_val = Variable(images_val.cuda(), volatile=True)
labels_val = Variable(labels_val, volatile=True)
_, _, pred, _ = enc_shared(images_val)
pred = upsample_512(pred)
pred = pred.data.max(1)[1].cpu().numpy()
gt = labels_val.data.cpu().numpy()
cty_running_metrics.update(gt, pred)
cty_score, cty_class_iou = cty_running_metrics.get_scores()
for k, v in cty_score.items():
print(k, v)
cty_running_metrics.reset()
City_tmp.append(cty_score['Mean IoU : \t'])
epoch_tmp.append(i_iter)
if i_iter % 10000 == 0 and i_iter != 0:
save_models(model_dict, './weight_' + str(i_iter))
if cty_score['Mean IoU : \t'] > best_iou:
best_iter = i_iter
best_iou = cty_score['Mean IoU : \t']
save_models(model_dict, './weight/')
enc_shared.eval()
print('evaluating models ...')
with torch.no_grad():
for i_val, (images_val, labels_val) in tqdm(enumerate(val_loader)):
images_val = Variable(images_val.cuda())
labels_val = Variable(labels_val)
_, aug_pred, pred, _ = enc_shared(images_val)
#pred = upsample_512(pred)
pred = 0.5 * aug_pred + pred
pred = upsample_540(pred)
pred = pred.data.max(1)[1].cpu().numpy()
gt = labels_val.data.cpu().numpy()
cty_running_metrics.update(gt, pred)
cty_score, cty_class_iou = cty_running_metrics.get_scores()
for k, v in cty_score.items():
print(k, v)
cty_running_metrics.reset()
City_tmp.append(cty_score['Mean IoU : \t'])
epoch_tmp.append(i_iter)
if i_iter % 3000 == 0 and i_iter != 0:
save_models(model_dict, save_model_path + 'weight_' + str(i_iter))
if cty_score['Mean IoU : \t'] > best_iou:
best_iter = i_iter
best_iou = cty_score['Mean IoU : \t']
save_models(model_dict, save_model_path + 'weight_best/')