xdobro23/YOLOv8-16bit
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ultralytics 8.0.65 YOLOv8 Pose models (#1347)

Browse Source 
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Mert Can Demir <validatedev@gmail.com>
Co-authored-by: Laughing <61612323+Laughing-q@users.noreply.github.com>
Co-authored-by: Fabian Greavu <fabiangreavu@gmail.com>
Co-authored-by: Yonghye Kwon <developer.0hye@gmail.com>
Co-authored-by: Eric Pedley <ericpedley@gmail.com>
Co-authored-by: JustasBart <40023722+JustasBart@users.noreply.github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Aarni Koskela <akx@iki.fi>
Co-authored-by: Sergio Sanchez <sergio.ssm.97@gmail.com>
Co-authored-by: Bogdan Gheorghe <112427971+bogdan-galileo@users.noreply.github.com>
Co-authored-by: Jaap van de Loosdrecht <jaap@vdlmv.nl>
Co-authored-by: Noobtoss <96134731+Noobtoss@users.noreply.github.com>
Co-authored-by: nerdyespresso <106761627+nerdyespresso@users.noreply.github.com>
Co-authored-by: Farid Inawan <frdteknikelektro@gmail.com>
Co-authored-by: Laughing-q <1185102784@qq.com>
Co-authored-by: Alexander Duda <Alexander.Duda@me.com>
Co-authored-by: Mehran Ghandehari <mehran.maps@gmail.com>
Co-authored-by: Snyk bot <snyk-bot@snyk.io>
Co-authored-by: majid nasiri <majnasai@gmail.com>
This commit is contained in:
Ayush Chaurasia
2023-04-06 03:55:32 +05:30
committed by GitHub
parent 9af3e69b1a
commit 1cb92d7f42
57 changed files with 1578 additions and 489 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ultralytics/yolo/v8/pose/__init__.py Normal file
View File

@ -0,0 +1,7 @@
# Ultralytics YOLO 🚀, GPL-3.0 license
from .predict import PosePredictor, predict
from .train import PoseTrainer, train
from .val import PoseValidator, val
__all__ = 'PoseTrainer', 'train', 'PoseValidator', 'val', 'PosePredictor', 'predict'

103
ultralytics/yolo/v8/pose/predict.py Normal file
View File

@ -0,0 +1,103 @@
# Ultralytics YOLO 🚀, GPL-3.0 license
from ultralytics.yolo.engine.results import Results
from ultralytics.yolo.utils import DEFAULT_CFG, ROOT, ops
from ultralytics.yolo.utils.plotting import colors, save_one_box
from ultralytics.yolo.v8.detect.predict import DetectionPredictor
class PosePredictor(DetectionPredictor):
def postprocess(self, preds, img, orig_img):
preds = ops.non_max_suppression(preds,
self.args.conf,
self.args.iou,
agnostic=self.args.agnostic_nms,
max_det=self.args.max_det,
classes=self.args.classes,
nc=len(self.model.names))
results = []
for i, pred in enumerate(preds):
orig_img = orig_img[i] if isinstance(orig_img, list) else orig_img
shape = orig_img.shape
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], shape).round()
pred_kpts = pred[:, 6:].view(len(pred), *self.model.kpt_shape) if len(pred) else pred[:, 6:]
pred_kpts = ops.scale_coords(img.shape[2:], pred_kpts, shape)
path, _, _, _, _ = self.batch
img_path = path[i] if isinstance(path, list) else path
results.append(
Results(orig_img=orig_img,
path=img_path,
names=self.model.names,
boxes=pred[:, :6],
keypoints=pred_kpts))
return results
def write_results(self, idx, results, batch):
p, im, im0 = batch
log_string = ''
if len(im.shape) == 3:
im = im[None] # expand for batch dim
self.seen += 1
imc = im0.copy() if self.args.save_crop else im0
if self.source_type.webcam or self.source_type.from_img: # batch_size >= 1
log_string += f'{idx}: '
frame = self.dataset.count
else:
frame = getattr(self.dataset, 'frame', 0)
self.data_path = p
self.txt_path = str(self.save_dir / 'labels' / p.stem) + ('' if self.dataset.mode == 'image' else f'_{frame}')
log_string += '%gx%g ' % im.shape[2:] # print string
self.annotator = self.get_annotator(im0)
det = results[idx].boxes # TODO: make boxes inherit from tensors
if len(det) == 0:
return f'{log_string}(no detections), '
for c in det.cls.unique():
n = (det.cls == c).sum() # detections per class
log_string += f"{n} {self.model.names[int(c)]}{'s' * (n > 1)}, "
kpts = reversed(results[idx].keypoints)
for k in kpts:
self.annotator.kpts(k, shape=results[idx].orig_shape)
# write
for j, d in enumerate(reversed(det)):
c, conf, id = int(d.cls), float(d.conf), None if d.id is None else int(d.id.item())
if self.args.save_txt: # Write to file
kpt = (kpts[j][:, :2] / d.orig_shape[[1, 0]]).reshape(-1).tolist()
box = d.xywhn.view(-1).tolist()
line = (c, *box, *kpt) + (conf, ) * self.args.save_conf + (() if id is None else (id, ))
with open(f'{self.txt_path}.txt', 'a') as f:
f.write(('%g ' * len(line)).rstrip() % line + '\n')
if self.args.save or self.args.show: # Add bbox to image
name = ('' if id is None else f'id:{id} ') + self.model.names[c]
label = (f'{name} {conf:.2f}' if self.args.show_conf else name) if self.args.show_labels else None
if self.args.boxes:
self.annotator.box_label(d.xyxy.squeeze(), label, color=colors(c, True))
if self.args.save_crop:
save_one_box(d.xyxy,
imc,
file=self.save_dir / 'crops' / self.model.model.names[c] / f'{self.data_path.stem}.jpg',
BGR=True)
return log_string
def predict(cfg=DEFAULT_CFG, use_python=False):
model = cfg.model or 'yolov8n-pose.pt'
source = cfg.source if cfg.source is not None else ROOT / 'assets' if (ROOT / 'assets').exists() \
else 'https://ultralytics.com/images/bus.jpg'
args = dict(model=model, source=source)
if use_python:
from ultralytics import YOLO
YOLO(model)(**args)
else:
predictor = PosePredictor(overrides=args)
predictor.predict_cli()
if __name__ == '__main__':
predict()

170
ultralytics/yolo/v8/pose/train.py Normal file
View File

@ -0,0 +1,170 @@
# Ultralytics YOLO 🚀, GPL-3.0 license
from copy import copy
import torch
import torch.nn as nn
from ultralytics.nn.tasks import PoseModel
from ultralytics.yolo import v8
from ultralytics.yolo.utils import DEFAULT_CFG
from ultralytics.yolo.utils.loss import KeypointLoss
from ultralytics.yolo.utils.metrics import OKS_SIGMA
from ultralytics.yolo.utils.ops import xyxy2xywh
from ultralytics.yolo.utils.plotting import plot_images, plot_results
from ultralytics.yolo.utils.tal import make_anchors
from ultralytics.yolo.utils.torch_utils import de_parallel
from ultralytics.yolo.v8.detect.train import Loss
# BaseTrainer python usage
class PoseTrainer(v8.detect.DetectionTrainer):
def __init__(self, cfg=DEFAULT_CFG, overrides=None):
if overrides is None:
overrides = {}
overrides['task'] = 'pose'
super().__init__(cfg, overrides)
def get_model(self, cfg=None, weights=None, verbose=True):
model = PoseModel(cfg, ch=3, nc=self.data['nc'], data_kpt_shape=self.data['kpt_shape'], verbose=verbose)
if weights:
model.load(weights)
return model
def set_model_attributes(self):
super().set_model_attributes()
self.model.kpt_shape = self.data['kpt_shape']
def get_validator(self):
self.loss_names = 'box_loss', 'pose_loss', 'kobj_loss', 'cls_loss', 'dfl_loss'
return v8.pose.PoseValidator(self.test_loader, save_dir=self.save_dir, args=copy(self.args))
def criterion(self, preds, batch):
if not hasattr(self, 'compute_loss'):
self.compute_loss = PoseLoss(de_parallel(self.model))
return self.compute_loss(preds, batch)
def plot_training_samples(self, batch, ni):
images = batch['img']
kpts = batch['keypoints']
cls = batch['cls'].squeeze(-1)
bboxes = batch['bboxes']
paths = batch['im_file']
batch_idx = batch['batch_idx']
plot_images(images,
batch_idx,
cls,
bboxes,
kpts=kpts,
paths=paths,
fname=self.save_dir / f'train_batch{ni}.jpg')
def plot_metrics(self):
plot_results(file=self.csv, pose=True) # save results.png
# Criterion class for computing training losses
class PoseLoss(Loss):
def __init__(self, model): # model must be de-paralleled
super().__init__(model)
self.kpt_shape = model.model[-1].kpt_shape
self.bce_pose = nn.BCEWithLogitsLoss()
is_pose = self.kpt_shape == [17, 3]
nkpt = self.kpt_shape[0] # number of keypoints
sigmas = torch.from_numpy(OKS_SIGMA).to(self.device) if is_pose else torch.ones(nkpt, device=self.device) / nkpt
self.keypoint_loss = KeypointLoss(sigmas=sigmas)
def __call__(self, preds, batch):
loss = torch.zeros(5, device=self.device) # box, cls, dfl, kpt_location, kpt_visibility
feats, pred_kpts = preds if isinstance(preds[0], list) else preds[1]
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_kpts = pred_kpts.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_size = pred_scores.shape[0]
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)
# pboxes
pred_bboxes = self.bbox_decode(anchor_points, pred_distri) # xyxy, (b, h*w, 4)
pred_kpts = self.kpts_decode(anchor_points, pred_kpts.view(batch_size, -1, *self.kpt_shape)) # (b, h*w, 17, 3)
_, 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 = max(target_scores.sum(), 1)
# cls loss
# loss[1] = self.varifocal_loss(pred_scores, target_scores, target_labels) / target_scores_sum # VFL way
loss[3] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum # BCE
# bbox loss
if fg_mask.sum():
target_bboxes /= stride_tensor
loss[0], loss[4] = self.bbox_loss(pred_distri, pred_bboxes, anchor_points, target_bboxes, target_scores,
target_scores_sum, fg_mask)
keypoints = batch['keypoints'].to(self.device).float().clone()
keypoints[..., 0] *= imgsz[1]
keypoints[..., 1] *= imgsz[0]
for i in range(batch_size):
if fg_mask[i].sum():
idx = target_gt_idx[i][fg_mask[i]]
gt_kpt = keypoints[batch_idx.view(-1) == i][idx] # (n, 51)
gt_kpt[..., 0] /= stride_tensor[fg_mask[i]]
gt_kpt[..., 1] /= stride_tensor[fg_mask[i]]
area = xyxy2xywh(target_bboxes[i][fg_mask[i]])[:, 2:].prod(1, keepdim=True)
pred_kpt = pred_kpts[i][fg_mask[i]]
kpt_mask = gt_kpt[..., 2] != 0
loss[1] += self.keypoint_loss(pred_kpt, gt_kpt, kpt_mask, area) # pose loss
# kpt_score loss
if pred_kpt.shape[-1] == 3:
loss[2] += self.bce_pose(pred_kpt[..., 2], kpt_mask.float()) # keypoint obj loss
loss[0] *= self.hyp.box # box gain
loss[1] *= self.hyp.pose / batch_size # pose gain
loss[2] *= self.hyp.kobj / batch_size # kobj gain
loss[3] *= self.hyp.cls # cls gain
loss[4] *= self.hyp.dfl # dfl gain
return loss.sum() * batch_size, loss.detach() # loss(box, cls, dfl)
def kpts_decode(self, anchor_points, pred_kpts):
y = pred_kpts.clone()
y[..., :2] *= 2.0
y[..., 0] += anchor_points[:, [0]] - 0.5
y[..., 1] += anchor_points[:, [1]] - 0.5
return y
def train(cfg=DEFAULT_CFG, use_python=False):
model = cfg.model or 'yolov8n-pose.yaml'
data = cfg.data or 'coco8-pose.yaml'
device = cfg.device if cfg.device is not None else ''
args = dict(model=model, data=data, device=device)
if use_python:
from ultralytics import YOLO
YOLO(model).train(**args)
else:
trainer = PoseTrainer(overrides=args)
trainer.train()
if __name__ == '__main__':
train()

213
ultralytics/yolo/v8/pose/val.py Normal file
View File

@ -0,0 +1,213 @@
# Ultralytics YOLO 🚀, GPL-3.0 license
from pathlib import Path
import numpy as np
import torch
from ultralytics.yolo.utils import DEFAULT_CFG, LOGGER, ops
from ultralytics.yolo.utils.checks import check_requirements
from ultralytics.yolo.utils.metrics import OKS_SIGMA, PoseMetrics, box_iou, kpt_iou
from ultralytics.yolo.utils.plotting import output_to_target, plot_images
from ultralytics.yolo.v8.detect import DetectionValidator
class PoseValidator(DetectionValidator):
def __init__(self, dataloader=None, save_dir=None, pbar=None, args=None):
super().__init__(dataloader, save_dir, pbar, args)
self.args.task = 'pose'
self.metrics = PoseMetrics(save_dir=self.save_dir)
def preprocess(self, batch):
batch = super().preprocess(batch)
batch['keypoints'] = batch['keypoints'].to(self.device).float()
return batch
def get_desc(self):
return ('%22s' + '%11s' * 10) % ('Class', 'Images', 'Instances', 'Box(P', 'R', 'mAP50', 'mAP50-95)', 'Pose(P',
'R', 'mAP50', 'mAP50-95)')
def postprocess(self, preds):
preds = ops.non_max_suppression(preds,
self.args.conf,
self.args.iou,
labels=self.lb,
multi_label=True,
agnostic=self.args.single_cls,
max_det=self.args.max_det,
nc=self.nc)
return preds
def init_metrics(self, model):
super().init_metrics(model)
self.kpt_shape = self.data['kpt_shape']
is_pose = self.kpt_shape == [17, 3]
nkpt = self.kpt_shape[0]
self.sigma = OKS_SIGMA if is_pose else np.ones(nkpt) / nkpt
def update_metrics(self, preds, batch):
# Metrics
for si, pred in enumerate(preds):
idx = batch['batch_idx'] == si
cls = batch['cls'][idx]
bbox = batch['bboxes'][idx]
kpts = batch['keypoints'][idx]
nl, npr = cls.shape[0], pred.shape[0] # number of labels, predictions
nk = kpts.shape[1] # number of keypoints
shape = batch['ori_shape'][si]
correct_kpts = torch.zeros(npr, self.niou, dtype=torch.bool, device=self.device) # init
correct_bboxes = torch.zeros(npr, self.niou, dtype=torch.bool, device=self.device) # init
self.seen += 1
if npr == 0:
if nl:
self.stats.append((correct_bboxes, correct_kpts, *torch.zeros(
(2, 0), device=self.device), cls.squeeze(-1)))
if self.args.plots:
self.confusion_matrix.process_batch(detections=None, labels=cls.squeeze(-1))
continue
# Predictions
if self.args.single_cls:
pred[:, 5] = 0
predn = pred.clone()
ops.scale_boxes(batch['img'][si].shape[1:], predn[:, :4], shape,
ratio_pad=batch['ratio_pad'][si]) # native-space pred
pred_kpts = predn[:, 6:].view(npr, nk, -1)
ops.scale_coords(batch['img'][si].shape[1:], pred_kpts, shape, ratio_pad=batch['ratio_pad'][si])
# Evaluate
if nl:
height, width = batch['img'].shape[2:]
tbox = ops.xywh2xyxy(bbox) * torch.tensor(
(width, height, width, height), device=self.device) # target boxes
ops.scale_boxes(batch['img'][si].shape[1:], tbox, shape,
ratio_pad=batch['ratio_pad'][si]) # native-space labels
tkpts = kpts.clone()
tkpts[..., 0] *= width
tkpts[..., 1] *= height
tkpts = ops.scale_coords(batch['img'][si].shape[1:], tkpts, shape, ratio_pad=batch['ratio_pad'][si])
labelsn = torch.cat((cls, tbox), 1) # native-space labels
correct_bboxes = self._process_batch(predn[:, :6], labelsn)
correct_kpts = self._process_batch(predn[:, :6], labelsn, pred_kpts, tkpts)
if self.args.plots:
self.confusion_matrix.process_batch(predn, labelsn)
# Append correct_masks, correct_boxes, pconf, pcls, tcls
self.stats.append((correct_bboxes, correct_kpts, pred[:, 4], pred[:, 5], cls.squeeze(-1)))
# Save
if self.args.save_json:
self.pred_to_json(predn, batch['im_file'][si])
# if self.args.save_txt:
# save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
def _process_batch(self, detections, labels, pred_kpts=None, gt_kpts=None):
"""
Return correct prediction matrix
Arguments:
detections (array[N, 6]), x1, y1, x2, y2, conf, class
labels (array[M, 5]), class, x1, y1, x2, y2
pred_kpts (array[N, 51]), 51 = 17 * 3
gt_kpts (array[N, 51])
Returns:
correct (array[N, 10]), for 10 IoU levels
"""
if pred_kpts is not None and gt_kpts is not None:
# `0.53` is from https://github.com/jin-s13/xtcocoapi/blob/master/xtcocotools/cocoeval.py#L384
area = ops.xyxy2xywh(labels[:, 1:])[:, 2:].prod(1) * 0.53
iou = kpt_iou(gt_kpts, pred_kpts, sigma=self.sigma, area=area)
else: # boxes
iou = box_iou(labels[:, 1:], detections[:, :4])
correct = np.zeros((detections.shape[0], self.iouv.shape[0])).astype(bool)
correct_class = labels[:, 0:1] == detections[:, 5]
for i in range(len(self.iouv)):
x = torch.where((iou >= self.iouv[i]) & correct_class) # IoU > threshold and classes match
if x[0].shape[0]:
matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]),
1).cpu().numpy() # [label, detect, iou]
if x[0].shape[0] > 1:
matches = matches[matches[:, 2].argsort()[::-1]]
matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
# matches = matches[matches[:, 2].argsort()[::-1]]
matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
correct[matches[:, 1].astype(int), i] = True
return torch.tensor(correct, dtype=torch.bool, device=detections.device)
def plot_val_samples(self, batch, ni):
plot_images(batch['img'],
batch['batch_idx'],
batch['cls'].squeeze(-1),
batch['bboxes'],
kpts=batch['keypoints'],
paths=batch['im_file'],
fname=self.save_dir / f'val_batch{ni}_labels.jpg',
names=self.names)
def plot_predictions(self, batch, preds, ni):
pred_kpts = torch.cat([p[:, 6:].view(-1, *self.kpt_shape)[:15] for p in preds], 0)
plot_images(batch['img'],
*output_to_target(preds, max_det=15),
kpts=pred_kpts,
paths=batch['im_file'],
fname=self.save_dir / f'val_batch{ni}_pred.jpg',
names=self.names) # pred
def pred_to_json(self, predn, filename):
stem = Path(filename).stem
image_id = int(stem) if stem.isnumeric() else stem
box = ops.xyxy2xywh(predn[:, :4]) # xywh
box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner
for p, b in zip(predn.tolist(), box.tolist()):
self.jdict.append({
'image_id': image_id,
'category_id': self.class_map[int(p[5])],
'bbox': [round(x, 3) for x in b],
'keypoints': p[6:],
'score': round(p[4], 5)})
def eval_json(self, stats):
if self.args.save_json and self.is_coco and len(self.jdict):
anno_json = self.data['path'] / 'annotations/person_keypoints_val2017.json' # annotations
pred_json = self.save_dir / 'predictions.json' # predictions
LOGGER.info(f'\nEvaluating pycocotools mAP using {pred_json} and {anno_json}...')
try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
check_requirements('pycocotools>=2.0.6')
from pycocotools.coco import COCO # noqa
from pycocotools.cocoeval import COCOeval # noqa
for x in anno_json, pred_json:
assert x.is_file(), f'{x} file not found'
anno = COCO(str(anno_json)) # init annotations api
pred = anno.loadRes(str(pred_json)) # init predictions api (must pass string, not Path)
for i, eval in enumerate([COCOeval(anno, pred, 'bbox'), COCOeval(anno, pred, 'keypoints')]):
if self.is_coco:
eval.params.imgIds = [int(Path(x).stem) for x in self.dataloader.dataset.im_files] # im to eval
eval.evaluate()
eval.accumulate()
eval.summarize()
idx = i * 4 + 2
stats[self.metrics.keys[idx + 1]], stats[
self.metrics.keys[idx]] = eval.stats[:2] # update mAP50-95 and mAP50
except Exception as e:
LOGGER.warning(f'pycocotools unable to run: {e}')
return stats
def val(cfg=DEFAULT_CFG, use_python=False):
model = cfg.model or 'yolov8n-pose.pt'
data = cfg.data or 'coco128-pose.yaml'
args = dict(model=model, data=data)
if use_python:
from ultralytics import YOLO
YOLO(model).val(**args)
else:
validator = PoseValidator(args=args)
validator(model=args['model'])
if __name__ == '__main__':
val()