Evaluation gives 0 score - pytorch

Question

I am working on a instance segmentation problem on mask rcnn with pytorch. Training part is working with below code but evaluation gives 0 score at every mAP. What's the problem in the code?
More info:
I use Albumentations for transforms and some files from pytorch vision for training.
Some problems I've been through:
When I use coco for bbox format instead of pascal voc it gives following error.
RuntimeError: Boolean value of Tensor with more than one value is ambiguous
When put labels out of the convert_seg_boolMask function, it gives following error.
RuntimeError: Boolean value of Tensor with more than one value is ambiguous
def get_transforms(train=False):
if train:
transform = A.Compose([
ToTensorV2()
],bbox_params=A.BboxParams("pascal_voc",label_fields=["labels","iscrowd"]))
else:
transform = A.Compose([
ToTensorV2()
],bbox_params=A.BboxParams("pascal_voc",label_fields=["labels","iscrowd"]))
return transform
class Dataset(datasets.VisionDataset):
def __init__(self, coco_, data_dir, transform=None, target_transform=None, transforms=None):
super().__init__(data_dir, transforms, transform, target_transform)
self.coco_info = coco_
self.data_dir = data_dir
self.transforms = transforms
if isinstance(self.coco_info,dict):
self.ids = [x["id"] for x in self.coco_info["images"] if len(self._load_target(x["id"]))>0]
def _load_image(self, id: int):
name = loadImgs(self.coco_info["images"],id)[0]['file_name']
image = cv2.imread(os.path.join(self.data_dir, name))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)/255
return image
def _load_target(self, id):
return loadAnns(self.coco_info["annotations"],id)
def n_classes(self):
category_names = list(set(sorted([a["name"] for a in self.coco_info["categories"]])))
self.classes = ["__background__"]+[i for i in category_names]
return self.classes
def __getitem__(self,idx):
id = self.ids[idx]
image = self._load_image(id)
target = copy.deepcopy(self._load_target(id))
image_shape = (image.shape[0],image.shape[1])
img_info = {
"img_shape":image_shape,
"image_id":id,
"labels":[t["category_id"]for t in target],
"segmentation":[t["segmentation"][0] for t in target],
"id": [t["id"] for t in target]
}
mask, labels = self.convert_seg_to_boolMask(img_info)
obj_ids = np.unique(mask)
obj_ids = obj_ids[1:]
masks = torch.tensor(mask == obj_ids[:, None, None])
boxes = []
bbox = np.array([t["bbox"] for t in target])
for xmin,ymin,width,height in bbox:
xmax = xmin+width
ymax = ymin+height
boxes.append([xmin, ymin, xmax, ymax])
boxes = torch.tensor(boxes)
labels = torch.tensor(labels)
image_id = torch.tensor([id])
iscrowd = torch.tensor([t["iscrowd"] for t in target])
transformed = self.transforms(image=image, masks=masks, bboxes=boxes, labels=labels, iscrowd=iscrowd)
image = transformed['image']
masks = torch.tensor(transformed["masks"])
boxes = torch.tensor(transformed['bboxes'])
labels = torch.tensor(transformed["labels"])
iscrowd = torch.tensor(transformed["iscrowd"])
area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])
image_id = torch.tensor(image_id)
target = {}
target["boxes"] = boxes
target["labels"] = labels
target["masks"] = masks
target["image_id"] = image_id
target["area"] = area
target["iscrowd"] = iscrowd
return image, target
def __len__(self):
return len(self.ids)
def convert_seg_to_boolMask(self,img_info):
mask = np.zeros(img_info["img_shape"], dtype=np.uint8)
mask = Image.fromarray(mask)
draw = ImageDraw.Draw(mask)
for seg, i in zip(img_info["segmentation"],img_info["id"]):
points = [tuple([k,l]) for k,l in zip(seg[0::2],seg[1::2])]
draw.polygon(xy=points,
outline=tuple([i]),
fill=tuple([i]))
mask = np.array(mask)
labels = img_info["labels"]
return mask, labels