def _accumulate_predictions_from_multiple_gpus(predictions_per_gpu, logger):
if is_main_process():
logger.info("Accumulating ...")
all_predictions = all_gather(predictions_per_gpu)
if not is_main_process():
return
predictions = list()
for p in all_predictions:
predictions.extend(p)
return predictions
def test(cfg, test_engine, loader, datasets, all_hooks):
total_timer = Timer()
total_timer.tic()
all_results = [[] for _ in range(4)]
eval = Evaluation(cfg)
with torch.no_grad():
loader = iter(loader)
for i in range(len(loader)):
all_hooks.iter_tic()
all_hooks.data_tic()
inputs, targets, idx = next(loader)
all_hooks.data_toc()
all_hooks.infer_tic()
result = test_engine(inputs, targets)
all_hooks.infer_toc()
all_hooks.post_tic()
eval_results = eval.post_processing(result, targets, idx, datasets)
all_results = [
results + eva
for results, eva in zip(all_results, eval_results)
]
all_hooks.post_toc()
all_hooks.iter_toc()
if is_main_process():
all_hooks.log_stats(i, 0, len(loader), len(datasets))
all_results = list(zip(*all_gather(all_results)))
all_results = [[item for sublist in results for item in sublist]
for results in all_results]
if is_main_process():
total_timer.toc(average=False)
logging_rank('Total inference time: {:.3f}s'.format(
total_timer.average_time))
eval.evaluation(datasets, all_results)