مقایسه‌ی تشخیص و بخش‌بندی (Detection و Segmentation)

def nms(boxes, scores, iou_threshold=0.45):
    order = np.argsort(-scores)              # ابتدا بهترین امتیاز
    keep = []
    while len(order) > 0:
        i = order[0]
        keep.append(i)
        if len(order) == 1:
            break
        rest = order[1:]
        ious = box_iou(boxes[[i]], boxes[rest])[0]
        order = rest[ious <= iou_threshold]  # حذف همپوشانی‌ها
    return np.array(keep, dtype=np.int64)

def dice_loss(logits, targets, num_classes, eps=1e-6):
    probs = F.softmax(logits, dim=1)
    targets_one_hot = F.one_hot(targets, num_classes).permute(0, 3, 1, 2).float()
    dims = (0, 2, 3)                                 # دسته، H، W
    intersection = (probs * targets_one_hot).sum(dim=dims)
    denom = probs.sum(dim=dims) + targets_one_hot.sum(dim=dims)
    dice = (2 * intersection + eps) / (denom + eps)  # مقدار eps برای کلاس‌های غایب
    return 1 - dice.mean()                           # معیار macro Dice
def combined_loss(logits, targets, num_classes, lam=1.0):
    ce = F.cross_entropy(logits, targets)            # ابعاد (N, C, H, W) و (N, H, W)
    dc = dice_loss(logits, targets, num_classes)
    return ce + lam * dc, {"ce": ce.item(), "dice": dc.item()}

model = maskrcnn_resnet50_fpn_v2(weights=MaskRCNN_ResNet50_FPN_V2_Weights.DEFAULT)
in_features = model.roi_heads.box_predictor.cls_score.in_features
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)  # پس‌زمینه را حساب می‌کند
in_features_mask = model.roi_heads.mask_predictor.conv5_mask.in_channels
model.roi_heads.mask_predictor = MaskRCNNPredictor(in_features_mask, 256, num_classes)
for p in model.backbone.parameters():
    p.requires_grad = False                                                  # FPN را نیز منجمد می‌کند
loss_dict = model(images, targets)                                           # حالت train: دیکشنری توابع زیان
losses = sum(loss for loss in loss_dict.values())
binary_masks = (p['masks'] > 0.5).squeeze(1)                                 # حالت eval: ابعاد (N, H, W)