Merge pull request #183 from wenh06/master

modify the evaluate function in train.py along with the modification of the model inference output

Author: ersheng-ai

tool/utils_iou.py

@@ -171,13 +171,17 @@ def bboxes_iou(bboxes_a, bboxes_b, fmt='voc', iou_type='iou'):
     if iou_type.lower() == 'diou':
         return diou
 
+    """ the legacy custom cosine similarity:
+
     # bb_a of shape `(N,2)`, bb_b of shape `(K,2)`
     v = torch.einsum('nm,km->nk', bb_a, bb_b)
     v = _true_divide(v, (torch.norm(bb_a, p='fro', dim=1)[:,np.newaxis] * torch.norm(bb_b, p='fro', dim=1)))
     # avoid nan for torch.acos near \pm 1
     # https://github.com/pytorch/pytorch/issues/8069
     eps = 1e-7
     v = torch.clamp(v, -1+eps, 1-eps)
+    """
+    v = F.cosine_similarity(bb_a[:,np.newaxis,:], bb_b, dim=-1)
     v = (_true_divide(2*torch.acos(v), np.pi)).pow(2)
     with torch.no_grad():
         alpha = (_true_divide(v, 1-iou+v)) * ((iou>=0.5).type(iou.type()))

tool/utils_iou_test.py

@@ -138,11 +138,17 @@ def bboxes_iou_test(bboxes_a, bboxes_b, fmt='voc', iou_type='iou'):
     # if iou_type.lower() == 'diou':
     #     return diou
 
+    """ the legacy custom cosine similarity:
+
     # bb_a of shape `(N,2)`, bb_b of shape `(K,2)`
     v = torch.einsum('nm,km->nk', bb_a, bb_b)
     v = _true_divide(v, (torch.norm(bb_a, p='fro', dim=1)[:,np.newaxis] * torch.norm(bb_b, p='fro', dim=1)))
+    # avoid nan for torch.acos near \pm 1
+    # https://github.com/pytorch/pytorch/issues/8069
     eps = 1e-7
     v = torch.clamp(v, -1+eps, 1-eps)
+    """
+    v = F.cosine_similarity(bb_a[:,np.newaxis,:], bb_b, dim=-1)
     v = (_true_divide(2*torch.acos(v), np.pi)).pow(2)
     alpha = (_true_divide(v, 1-iou+v))*((iou>=0.5).type(iou.type()))
 

train.py

@@ -487,18 +487,21 @@ def evaluate(model, data_loader, cfg, device, logger=None, **kwargs):
         # outputs = [{k: v.to(cpu_device) for k, v in t.items()} for t in outputs]
         model_time = time.time() - model_time
 
-        outputs = outputs.cpu().detach().numpy()
+        # outputs = outputs.cpu().detach().numpy()
         res = {}
-        for img, target, output in zip(images, targets, outputs):
+        # for img, target, output in zip(images, targets, outputs):
+        for img, target, (boxes, confs) in zip(images, targets, outputs):
             img_height, img_width = img.shape[:2]
-            boxes = output[...,:4].copy()  # output boxes in yolo format
+            # boxes = output[...,:4].copy()  # output boxes in yolo format
+            boxes = boxes.squeeze(2).cpu().detach().numpy()
             boxes[...,:2] = boxes[...,:2] - boxes[...,2:]/2  # to coco format
             boxes[...,0] = boxes[...,0]*img_width
             boxes[...,1] = boxes[...,1]*img_height
             boxes[...,2] = boxes[...,2]*img_width
             boxes[...,3] = boxes[...,3]*img_height
             boxes = torch.as_tensor(boxes, dtype=torch.float32)
-            confs = output[...,4:].copy()
+            # confs = output[...,4:].copy()
+            confs = confs.cpu().detach().numpy()
             labels = np.argmax(confs, axis=1).flatten()
             labels = torch.as_tensor(labels, dtype=torch.int64)
             scores = np.max(confs, axis=1).flatten()