| import torch |
| import torch.nn.functional as F |
| |
| import numpy as np |
|
|
|
|
| def validater(args, val_data, logger, epoch_num, sam, |
| loss_validation): |
| patch_size = args.rand_crop_size[0] |
| device = args.device |
| with torch.no_grad(): |
| loss_summary = [] |
| |
| |
| for idx, (image, label, _) in enumerate(val_data): |
| |
| |
| |
| |
|
|
| image, label = image.to(device), label.to(device) |
|
|
| image_embedding = sam.image_encoder(image) |
| prev_masks = interaction(args, sam, image_embedding, label, num_clicks=11) |
|
|
| masks = prev_masks |
| loss = loss_validation(masks, label) |
| loss_summary.append(loss.detach().cpu().numpy()) |
| logger.info( |
| 'epoch: {}/{}, iter: {}/{}'.format(epoch_num, args.max_epoch, idx, len(val_data)) + ": loss:" + str( |
| loss_summary[-1].flatten()[0])) |
| logger.info("- Val metrics: " + str(np.mean(loss_summary))) |
| return loss_summary |
|
|
|
|
| def get_next_click3D_torch_2(prev_seg, gt_semantic_seg): |
|
|
| mask_threshold = 0.5 |
|
|
| batch_points = [] |
| batch_labels = [] |
| |
|
|
| pred_masks = (prev_seg > mask_threshold) |
| true_masks = (gt_semantic_seg > 0) |
| fn_masks = torch.logical_and(true_masks, torch.logical_not(pred_masks)) |
| fp_masks = torch.logical_and(torch.logical_not(true_masks), pred_masks) |
|
|
| to_point_mask = torch.logical_or(fn_masks, fp_masks) |
|
|
| for i in range(gt_semantic_seg.shape[0]): |
|
|
| points = torch.argwhere(to_point_mask[i]) |
| point = points[np.random.randint(len(points))] |
| |
| if fn_masks[i, 0, point[1], point[2], point[3]]: |
| is_positive = True |
| else: |
| is_positive = False |
|
|
| bp = point[1:].clone().detach().reshape(1, 1, 3) |
| bl = torch.tensor([int(is_positive), ]).reshape(1, 1) |
| batch_points.append(bp) |
| batch_labels.append(bl) |
|
|
| return batch_points, batch_labels |
| def get_points(args, prev_masks, gt3D, click_points, click_labels): |
| batch_points, batch_labels = get_next_click3D_torch_2(prev_masks, gt3D) |
|
|
| points_co = torch.cat(batch_points, dim=0).to(args.device) |
| points_la = torch.cat(batch_labels, dim=0).to(args.device) |
|
|
| click_points.append(points_co) |
| click_labels.append(points_la) |
|
|
| points_multi = torch.cat(click_points, dim=1).to(args.device) |
| labels_multi = torch.cat(click_labels, dim=1).to(args.device) |
|
|
| |
| |
| |
| |
| points_input = points_co |
| labels_input = points_la |
| return points_input, labels_input, click_points, click_labels |
|
|
| def batch_forward(args, sam_model, image_embedding, gt3D, low_res_masks, points=None): |
|
|
| sparse_embeddings, dense_embeddings = sam_model.prompt_encoder( |
| points=points, |
| boxes=None, |
| masks=low_res_masks, |
| ) |
| low_res_masks, iou_predictions = sam_model.mask_decoder( |
| image_embeddings=image_embedding.to(args.device), |
| image_pe=sam_model.prompt_encoder.get_dense_pe(), |
| sparse_prompt_embeddings=sparse_embeddings, |
| dense_prompt_embeddings=dense_embeddings, |
| multimask_output=False, |
| ) |
| prev_masks = F.interpolate(low_res_masks, size=gt3D.shape[-3:], mode='trilinear', align_corners=False) |
| return low_res_masks, prev_masks |
|
|
| def interaction(args, sam_model, image_embedding, gt3D, num_clicks): |
| |
| prev_masks = torch.zeros_like(gt3D).to(gt3D.device) |
| random_insert = np.random.randint(2, 9) |
|
|
| click_points, click_labels = [], [] |
| for num_click in range(num_clicks): |
| points_input, labels_input, click_points, click_labels = get_points(args, prev_masks, gt3D, click_points, click_labels) |
|
|
| if num_click == random_insert or num_click == num_clicks - 1: |
| prev_masks = batch_forward(args, sam_model, image_embedding, gt3D, prev_masks, points=None) |
| else: |
| prev_masks = batch_forward(args, sam_model, image_embedding, gt3D, prev_masks, points=[points_input, labels_input]) |
| |
| |
| |
| return prev_masks |
