def ding_test(cfg:BaseConfigByEpoch, net=None, val_dataloader=None, show_variables=False, convbuilder=None,
init_hdf5=None, extra_msg=None, weights_dict=None):
with Engine(local_rank=0, for_val_only=True) as engine:
engine.setup_log(
name='test', log_dir='./', file_name=DETAIL_LOG_FILE)
if convbuilder is None:
convbuilder = ConvBuilder(base_config=cfg)
if net is None:
net_fn = get_model_fn(cfg.dataset_name, cfg.network_type)
model = net_fn(cfg, convbuilder).cuda()
else:
model = net.cuda()
if val_dataloader is None:
val_data = create_dataset(cfg.dataset_name, cfg.dataset_subset,
global_batch_size=cfg.global_batch_size, distributed=False)
num_examples = num_val_examples(cfg.dataset_name)
assert num_examples % cfg.global_batch_size == 0
val_iters = num_val_examples(cfg.dataset_name) // cfg.global_batch_size
print('batchsize={}, {} iters'.format(cfg.global_batch_size, val_iters))
criterion = get_criterion(cfg).cuda()
engine.register_state(
scheduler=None, model=model, optimizer=None)
if show_variables:
engine.show_variables()
assert not engine.distributed
if weights_dict is not None:
engine.load_from_weights_dict(weights_dict)
else:
if cfg.init_weights:
engine.load_checkpoint(cfg.init_weights)
if init_hdf5:
engine.load_hdf5(init_hdf5)
# engine.save_by_order('smi2_by_order.hdf5')
# engine.load_by_order('smi2_by_order.hdf5')
# engine.save_hdf5('model_files/stami2_lrs4Z.hdf5')
model.eval()
eval_dict, total_net_time = run_eval(val_data, val_iters, model, criterion, 'TEST', dataset_name=cfg.dataset_name)
val_top1_value = eval_dict['top1'].item()
val_top5_value = eval_dict['top5'].item()
val_loss_value = eval_dict['loss'].item()
msg = '{},{},{},top1={:.5f},top5={:.5f},loss={:.7f},total_net_time={}'.format(cfg.network_type, init_hdf5 or cfg.init_weights, cfg.dataset_subset,
val_top1_value, val_top5_value, val_loss_value, total_net_time)
if extra_msg is not None:
msg += ', ' + extra_msg
log_important(msg, OVERALL_LOG_FILE)
return eval_dict
def val_during_train(epoch, iteration, tb_tags,
engine, model, val_data, criterion, descrip_str,
dataset_name, test_batch_size, tb_writer):
model.eval()
num_examples = num_val_examples(dataset_name)
assert num_examples % test_batch_size == 0
val_iters = num_examples // test_batch_size
eval_dict, total_net_time = run_eval(val_data, val_iters, model, criterion, descrip_str,
dataset_name=dataset_name)
val_top1_value = eval_dict['top1'].item()
val_top5_value = eval_dict['top5'].item()
val_loss_value = eval_dict['loss'].item()
for tag, value in zip(tb_tags, [val_top1_value, val_top5_value, val_loss_value]):
tb_writer.add_scalars(tag, {'Val': value}, iteration)
engine.log(
'val at epoch {}, top1={:.5f}, top5={:.5f}, loss={:.6f}'.format(epoch, val_top1_value,
val_top5_value,
val_loss_value))
model.train()