batch = dataset[0]
imgId = batch['img_id']
img_path = img_map[imgId]
for key in batch:
if isinstance(batch[key], torch.Tensor):
batch[key] = batch[key].unsqueeze(0).numpy()
else:
try:
batch[key] = batch[key][np.newaxis, :]
except:
continue
feat = {}
feat['hm'] = torch.from_numpy(batch['hm'])
feat['wh'] = torch.from_numpy(
gen_oracle_map(batch['wh'], batch['ind'], 128, 128))
feat['reg'] = torch.from_numpy(
gen_oracle_map(batch['reg'], batch['ind'], 128, 128))
feat['kps'] = torch.from_numpy(
gen_oracle_map(batch['kps'], batch['ind'], 128, 128))
dets = centerface_decode(feat['hm'],
feat['wh'],
feat['kps'],
reg=feat['reg'],
K=100)
dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2])
dets_out = centerface_post_process(dets.copy(), batch['c'], batch['s'],
feat['hm'].shape[2],
feat['hm'].shape[3],
feat['hm'].shape[1])
def _evaluate(self, eval=False):
logger.info("Epoch {} evaluation start".format(self.epoch))
title = "EVALUATE" if eval else f"TEST[{self.epoch}/{self.cfg.SOLVER.MAX_EPOCHS}]"
results = {}
self._eval_epoch_start()
out_size = self.out_sizes[0]
with torch.no_grad():
self._eval_epoch_start()
for batch in tqdm(self.vdata, desc=title):
if self.use_gpu:
for key in batch:
if isinstance(batch[key], dict):
for sub_key in batch[key]:
if self.cfg.DISTRIBUTED:
batch[key][sub_key] = batch[key][
sub_key].to(self.device,
non_blocking=True)
else:
batch[key][sub_key] = batch[key][
sub_key].cuda()
elif not isinstance(batch[key], torch.Tensor):
continue
else:
if self.cfg.DISTRIBUTED:
batch[key] = batch[key].to(self.device,
non_blocking=True)
else:
batch[key] = batch[key].cuda()
if self.cfg.ORACLE:
feat = {}
feat['hm'] = batch[out_size]['hm']
feat['wh'] = torch.from_numpy(
gen_oracle_map(
batch[out_size]['wh'].detach().cpu().numpy(),
batch[out_size]['ind'].detach().cpu().numpy(),
batch['inp'].shape[3] // self.cfg.MODEL.STRIDES[0],
batch['inp'].shape[2] //
self.cfg.MODEL.STRIDES[0])).cuda()
feat['reg'] = torch.from_numpy(
gen_oracle_map(
batch[out_size]['reg'].detach().cpu().numpy(),
batch[out_size]['ind'].detach().cpu().numpy(),
batch['inp'].shape[3] // self.cfg.MODEL.STRIDES[0],
batch['inp'].shape[2] //
self.cfg.MODEL.STRIDES[0])).cuda()
else:
feat = self.graph.run(batch['inp'])[out_size]
feat['hm'].sigmoid_()
if self.cfg.DB.TARGET_FORMAT == 'centerface_bbox':
feat['wh'].exp_()
dets = centernet_det_decode(feat['hm'],
feat['wh'],
reg=feat['reg'],
K=100)
dets = dets.detach().cpu().numpy().reshape(
1, -1, dets.shape[1])
dets_out = centernet_det_post_process(dets.copy(),
batch['c'].cpu().numpy(),
batch['s'].cpu().numpy(),
feat['hm'].shape[2],
feat['hm'].shape[3],
feat['hm'].shape[1])
results[batch['img_id'][0]] = dets_out[0]
cce = coco_eval(self.vdata.dataset.coco[0], results,
self.cfg.OUTPUT_DIR)
logger.info(
'Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[0]))
logger.info(
'Average Precision (AP) @[ IoU=0.50 | area= all | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[1]))
logger.info(
'Average Precision (AP) @[ IoU=0.75 | area= all | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[2]))
logger.info(
'Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[3]))
logger.info(
'Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[4]))
logger.info(
'Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[5]))
logger.info(
'Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[0], cce.stats[6]))
logger.info(
'Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[1], cce.stats[7]))
logger.info(
'Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[8]))
logger.info(
'Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[9]))
logger.info(
'Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[10]))
logger.info(
'Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets={:>3d} ] = {:.3f}'
.format(cce.params.maxDets[2], cce.stats[11]))
if not eval:
self.accu = cce.stats[0]
self._eval_epoch_end()