Untitled

tiles = 2 # 3??


# v for loopu nekako takole:
thr = xy???
if (bboxes[:,:,2]-bboxes[:,:,0]).mean() > thr or (bboxes[:,:,3]-bboxes[:,:,1]).mean() > thr:
    tile = img.shape[2] // tiles
    imgs = []
    for i in range(tiles):
        for j in range(tiles):
            imgs.append(img[:, :, i * tile:(i + 1) * tile, j * tile:(j + 1) * tile])

    # change bboxes:
    # if a bbox is in the image, it has to have all coordinates > 0 and x2 and y2 < image size
    bboxes_ = []
    for i in range(tiles):
        for j in range(tiles):
            bboxes_.append(bboxes - torch.tensor([j * tile, i * tile, j * tile, i * tile]).to(device))

    # check on which tile a bbox is
    for i in range(tiles*tiles):
        bboxes_[i] = bboxes_[i][torch.logical_not(
            torch.logical_or((bboxes_[i] < 0).any(dim=2), (bboxes_[i] > imgs[i].shape[3]).any(dim=2)))]
            
            
     # # pad bboxes to have the same number of bboxes
    max_num_bboxes = max([bboxes_[i].shape[0] for i in range(tiles*tiles)])
    for i in range(tiles*tiles):
        bboxes_[i] = torch.cat([bboxes_[i], torch.zeros((max_num_bboxes - bboxes_[i].shape[0], 4)).to(device)])

    bboxes_batch = torch.stack([bb for bb in bboxes_])*tiles
    img_batch = torch.stack([im[0] for im in imgs])

    # Upscale image tensor to 1024,1024
    img_batch = torch.nn.functional.interpolate(img_batch, size=1024, mode='bilinear', align_corners=False)
    



#### Kjer imaš roi-pooling popravi, da se concatenirajo exemplarji in replicirajo v batch size (tj #tiles)