def segmentation_eval(config, net,
mapping_assignment_dataloader,
mapping_test_dataloader,
sobel, using_IR=False, verbose=0, return_only=False):
torch.cuda.empty_cache() if not config.nocuda else None
net.eval()
stats_dict = cluster_subheads_eval(config, net,
mapping_assignment_dataloader=mapping_assignment_dataloader,
mapping_test_dataloader=mapping_test_dataloader,
sobel=sobel,
using_IR=using_IR,
get_data_fn=_segmentation_get_data,
verbose=verbose)
net.train()
acc = stats_dict["best"]
is_best = (len(config.epoch_acc) > 0) and (acc > max(config.epoch_acc))
torch.cuda.empty_cache() if not config.nocuda else None
if not return_only:
config.epoch_stats.append(stats_dict)
config.epoch_acc.append(acc)
config.epoch_avg_subhead_acc.append(stats_dict["avg"])
return is_best
else:
return stats_dict
# ------------------------------------------------------
net = archs.__dict__[config.arch](config)
model_path = os.path.join(config.out_dir, "best_net.pytorch")
net.load_state_dict(
torch.load(model_path, map_location=lambda storage, loc: storage))
net.cuda()
if given_config.use_eval:
print("doing eval mode")
net.eval()
net = torch.nn.DataParallel(net)
acc, nmi, ari, _ = cluster_subheads_eval(
config,
net,
mapping_assignment_dataloader=mapping_assignment_dataloader,
mapping_test_dataloader=mapping_test_dataloader,
sobel=True)
config.assign_set_szs_pc_acc[str(pc)] = (num_imgs, acc)
print("for model %d assign set sz pc %f, got %f, compared to best stored "
"acc %f" % (config.model_ind, pc, acc, max(config.epoch_acc)))
print(config.assign_set_szs_pc_acc)
sysout.flush()
if not given_config.dont_save:
print("writing to new config")
# store to config
with open(os.path.join(config.out_dir, "config.pickle"), "wb") as outfile:
pickle.dump(config, outfile)