import hydra import torch import torch.nn as nn import torch.nn.functional as F from ultralytics.nn.tasks import SegmentationModel from ultralytics.yolo import v8 from ultralytics.yolo.utils import DEFAULT_CONFIG from ultralytics.yolo.utils.loss import BboxLoss from ultralytics.yolo.utils.ops import crop_mask, xywh2xyxy, xyxy2xywh from ultralytics.yolo.utils.plotting import plot_images, plot_results from ultralytics.yolo.utils.tal import TaskAlignedAssigner, dist2bbox, make_anchors from ultralytics.yolo.utils.torch_utils import de_parallel # BaseTrainer python usage class SegmentationTrainer(v8.detect.DetectionTrainer): def load_model(self, model_cfg=None, weights=None, verbose=True): model = SegmentationModel(model_cfg or getattr(weights, 'yaml', None) or weights['model'].yaml, ch=3, nc=self.data["nc"], verbose=verbose) if weights: model.load(weights['model'] if isinstance(weights, dict) else weights, verbose) return model def get_validator(self): self.loss_names = 'box_loss', 'seg_loss', 'cls_loss', 'dfl_loss' return v8.segment.SegmentationValidator(self.test_loader, save_dir=self.save_dir, logger=self.console, args=self.args) def criterion(self, preds, batch): if not hasattr(self, 'compute_loss'): self.compute_loss = SegLoss(de_parallel(self.model), overlap=self.args.overlap_mask) return self.compute_loss(preds, batch) def plot_training_samples(self, batch, ni): images = batch["img"] masks = batch["masks"] cls = batch["cls"].squeeze(-1) bboxes = batch["bboxes"] paths = batch["im_file"] batch_idx = batch["batch_idx"] plot_images(images, batch_idx, cls, bboxes, masks, paths=paths, fname=self.save_dir / f"train_batch{ni}.jpg") def plot_metrics(self): plot_results(file=self.csv, segment=True) # save results.png # Criterion class for computing training losses class SegLoss: def __init__(self, model, overlap=True): # model must be de-paralleled device = next(model.parameters()).device # get model device h = model.args # hyperparameters m = model.model[-1] # Detect() module self.bce = nn.BCEWithLogitsLoss(reduction='none') self.hyp = h self.stride = m.stride # model strides self.nc = m.nc # number of classes self.no = m.no self.nm = m.nm # number of masks self.reg_max = m.reg_max self.overlap = overlap self.device = device self.use_dfl = m.reg_max > 1 self.assigner = TaskAlignedAssigner(topk=10, num_classes=self.nc, alpha=0.5, beta=6.0) self.bbox_loss = BboxLoss(m.reg_max - 1, use_dfl=self.use_dfl).to(device) self.proj = torch.arange(m.reg_max, dtype=torch.float, device=device) def preprocess(self, targets, batch_size, scale_tensor): if targets.shape[0] == 0: out = torch.zeros(batch_size, 0, 5, device=self.device) else: i = targets[:, 0] # image index _, counts = i.unique(return_counts=True) out = torch.zeros(batch_size, counts.max(), 5, device=self.device) for j in range(batch_size): matches = i == j n = matches.sum() if n: out[j, :n] = targets[matches, 1:] out[..., 1:5] = xywh2xyxy(out[..., 1:5].mul_(scale_tensor)) return out def bbox_decode(self, anchor_points, pred_dist): if self.use_dfl: b, a, c = pred_dist.shape # batch, anchors, channels pred_dist = pred_dist.view(b, a, 4, c // 4).softmax(3).matmul(self.proj.type(pred_dist.dtype)) # pred_dist = pred_dist.view(b, a, c // 4, 4).transpose(2,3).softmax(3).matmul(self.proj.type(pred_dist.dtype)) # pred_dist = (pred_dist.view(b, a, c // 4, 4).softmax(2) * self.proj.type(pred_dist.dtype).view(1, 1, -1, 1)).sum(2) return dist2bbox(pred_dist, anchor_points, xywh=False) def __call__(self, preds, batch): loss = torch.zeros(4, device=self.device) # box, cls, dfl feats, pred_masks, proto = preds if len(preds) == 3 else preds[1] batch_size, _, mask_h, mask_w = proto.shape # batch size, number of masks, mask height, mask width pred_distri, pred_scores = torch.cat([xi.view(feats[0].shape[0], self.no, -1) for xi in feats], 2).split( (self.reg_max * 4, self.nc), 1) # b, grids, .. pred_scores = pred_scores.permute(0, 2, 1).contiguous() pred_distri = pred_distri.permute(0, 2, 1).contiguous() pred_masks = pred_masks.permute(0, 2, 1).contiguous() dtype = pred_scores.dtype imgsz = torch.tensor(feats[0].shape[2:], device=self.device, dtype=dtype) * self.stride[0] # image size (h,w) anchor_points, stride_tensor = make_anchors(feats, self.stride, 0.5) # targets batch_idx = batch["batch_idx"].view(-1, 1) targets = torch.cat((batch_idx, batch["cls"].view(-1, 1), batch["bboxes"]), 1) targets = self.preprocess(targets.to(self.device), batch_size, scale_tensor=imgsz[[1, 0, 1, 0]]) gt_labels, gt_bboxes = targets.split((1, 4), 2) # cls, xyxy mask_gt = gt_bboxes.sum(2, keepdim=True).gt_(0) masks = batch["masks"].to(self.device).float() if tuple(masks.shape[-2:]) != (mask_h, mask_w): # downsample masks = F.interpolate(masks[None], (mask_h, mask_w), mode="nearest")[0] # pboxes pred_bboxes = self.bbox_decode(anchor_points, pred_distri) # xyxy, (b, h*w, 4) _, target_bboxes, target_scores, fg_mask, target_gt_idx = self.assigner( pred_scores.detach().sigmoid(), (pred_bboxes.detach() * stride_tensor).type(gt_bboxes.dtype), anchor_points * stride_tensor, gt_labels, gt_bboxes, mask_gt) target_scores_sum = target_scores.sum() # cls loss # loss[1] = self.varifocal_loss(pred_scores, target_scores, target_labels) / target_scores_sum # VFL way loss[2] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum # BCE # bbox loss if fg_mask.sum(): loss[0], loss[3] = self.bbox_loss(pred_distri, pred_bboxes, anchor_points, target_bboxes / stride_tensor, target_scores, target_scores_sum, fg_mask) for i in range(batch_size): if fg_mask[i].sum(): mask_idx = target_gt_idx[i][fg_mask[i]] + 1 if self.overlap: gt_mask = torch.where(masks[[i]] == mask_idx.view(-1, 1, 1), 1.0, 0.0) else: gt_mask = masks[batch_idx == i][mask_idx] xyxyn = target_bboxes[i][fg_mask[i]] / imgsz[[1, 0, 1, 0]] marea = xyxy2xywh(xyxyn)[:, 2:].prod(1) mxyxy = xyxyn * torch.tensor([mask_w, mask_h, mask_w, mask_h], device=self.device) loss[1] += self.single_mask_loss(gt_mask, pred_masks[i][fg_mask[i]], proto[i], mxyxy, marea) # seg loss # WARNING: Uncomment lines below in case of Multi-GPU DDP unused gradient errors # else: # loss[1] += proto.sum() * 0 # else: # loss[1] += proto.sum() * 0 loss[0] *= 7.5 # box gain loss[1] *= 7.5 / batch_size # seg gain loss[2] *= 0.5 # cls gain loss[3] *= 1.5 # dfl gain return loss.sum() * batch_size, loss.detach() # loss(box, cls, dfl) def single_mask_loss(self, gt_mask, pred, proto, xyxy, area): # Mask loss for one image pred_mask = (pred @ proto.view(self.nm, -1)).view(-1, *proto.shape[1:]) # (n, 32) @ (32,80,80) -> (n,80,80) loss = F.binary_cross_entropy_with_logits(pred_mask, gt_mask, reduction="none") return (crop_mask(loss, xyxy).mean(dim=(1, 2)) / area).mean() @hydra.main(version_base=None, config_path=str(DEFAULT_CONFIG.parent), config_name=DEFAULT_CONFIG.name) def train(cfg): cfg.model = cfg.model or "yolov8n-seg.yaml" cfg.data = cfg.data or "coco128-seg.yaml" # or yolo.ClassificationDataset("mnist") trainer = SegmentationTrainer(cfg) trainer.train() if __name__ == "__main__": """ CLI usage: python ultralytics/yolo/v8/segment/train.py model=yolov8n-seg.yaml data=coco128-segments epochs=100 imgsz=640 TODO: Direct cli support, i.e, yolov8 classify_train args.epochs 10 """ train()