def step_epoch(self):
'''
return map_mean, map_50, map_75
note: this function will set self.net.train() at last
'''
with torch.no_grad():
self.net.eval()
pred_bboxes, pred_labels, pred_scores, gt_bboxes, gt_labels = [], [], [], [], []
for i, (imgs, boxes, labels, locs,
scales) in enumerate(self.loader):
if i == 0:
batch_size = int(imgs.shape[0])
temp = self.net(imgs, locs)
pred_cls_i, pred_cls_p, pred_reg = get_pred(
temp, self.net.module.nms_th,
self.net.module.nms_iou) # DataParallel
for idx in range(len(pred_cls_i)):
pred_cls_i[idx] = pred_cls_i[idx].cpu().detach().numpy()
pred_cls_p[idx] = pred_cls_p[idx].cpu().detach().numpy()
pred_reg[idx] = pred_reg[idx].cpu().detach().numpy()
_boxes, _label = [], []
for idx in range(boxes.shape[0]):
mask = labels[idx] > 0
_boxes.append(boxes[idx][mask].detach().numpy())
_label.append(labels[idx][mask].detach().numpy())
pred_bboxes += pred_reg
pred_labels += pred_cls_i
pred_scores += pred_cls_p
gt_bboxes += _boxes
gt_labels += _label
print(' Eval: {}/{}'.format(i * batch_size,
len(self.dataset)),
end='\r')
ap_iou = [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95]
ap_res = []
for iou_th in ap_iou:
res = eval_detection(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
iou_th=iou_th)
ap_res.append(res)
ap_sum = 0.0
for i in range(len(ap_res)):
ap_sum += float(ap_res[i]['map'])
map_mean = ap_sum / float(len(ap_res))
map_50 = float(ap_res[0]['map'])
map_75 = float(ap_res[5]['map'])
print('map_mean:', map_mean, 'map_50:', map_50, 'map_75:', map_75)
self.net.train()
return map_mean, map_50, map_75
_boxes.append(bbox[idx][mask].detach().numpy())
_label.append(label[idx][mask].detach().numpy())
pred_bboxes += reg_preds
pred_labels += cls_i_preds
pred_scores += cls_p_preds
gt_bboxes += _boxes
gt_labels += _label
print(' Eval: {}/{}'.format(i * cfg['nbatch_eval'],
len(dataset_eval)),
end='\r')
ap_iou = [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95]
ap_res = []
for iou_th in ap_iou:
res = eval_detection(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
iou_th=iou_th)
ap_res.append(res)
ap_sum = 0.0
for i in range(len(ap_res)):
ap_sum += float(ap_res[i]['map'])
map_mean = ap_sum / float(len(ap_res))
map_50 = float(ap_res[0]['map'])
map_75 = float(ap_res[5]['map'])
print('map_mean')
print(map_mean)
print('map_50')
print(map_50)