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

New YOLOv8 Results() class for prediction outputs (#314)

Browse Source 
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
Co-authored-by: Laughing-q <1185102784@qq.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Laughing <61612323+Laughing-q@users.noreply.github.com>
Co-authored-by: Viet Nhat Thai <60825385+vietnhatthai@users.noreply.github.com>
Co-authored-by: Paula Derrenger <107626595+pderrenger@users.noreply.github.com>
This commit is contained in:
Ayush Chaurasia
2023-01-17 19:02:34 +05:30
committed by GitHub
parent 0cb87f7dd3
commit c6985da9de
32 changed files with 813 additions and 259 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
ultralytics/yolo/engine/model.py
View File

@ -54,8 +54,8 @@ class YOLO:
# Load or create new YOLO model
{'.pt': self._load, '.yaml': self._new}[Path(model).suffix](model)
def __call__(self, source, **kwargs):
return self.predict(source, **kwargs)
def __call__(self, source=None, stream=False, verbose=False, **kwargs):
return self.predict(source, stream, verbose, **kwargs)
def _new(self, cfg: str, verbose=True):
"""
@ -111,13 +111,20 @@ class YOLO:
self.model.fuse()
@smart_inference_mode()
def predict(self, source, return_outputs=False, **kwargs):
def predict(self, source=None, stream=False, verbose=False, **kwargs):
"""
Visualize prediction.
Perform prediction using the YOLO model.
Args:
source (str): Accepts all source types accepted by yolo
**kwargs : Any other args accepted by the predictors. To see all args check 'configuration' section in docs
source (str | int | PIL | np.ndarray): The source of the image to make predictions on.
Accepts all source types accepted by the YOLO model.
stream (bool): Whether to stream the predictions or not. Defaults to False.
verbose (bool): Whether to print verbose information or not. Defaults to False.
**kwargs : Additional keyword arguments passed to the predictor.
Check the 'configuration' section in the documentation for all available options.
Returns:
(dict): The prediction results.
"""
overrides = self.overrides.copy()
overrides["conf"] = 0.25
@ -127,8 +134,8 @@ class YOLO:
predictor = self.PredictorClass(overrides=overrides)
predictor.args.imgsz = check_imgsz(predictor.args.imgsz, min_dim=2) # check image size
predictor.setup(model=self.model, source=source, return_outputs=return_outputs)
return predictor() if return_outputs else predictor.predict_cli()
predictor.setup(model=self.model, source=source)
return predictor(stream=stream, verbose=verbose)
@smart_inference_mode()
def val(self, data=None, **kwargs):

122
ultralytics/yolo/engine/predictor.py
View File

@ -27,13 +27,14 @@ Usage - formats:
"""
import platform
from collections import defaultdict
from itertools import chain
from pathlib import Path
import cv2
from ultralytics.nn.autobackend import AutoBackend
from ultralytics.yolo.configs import get_config
from ultralytics.yolo.data.dataloaders.stream_loaders import LoadImages, LoadScreenshots, LoadStreams
from ultralytics.yolo.data.dataloaders.stream_loaders import LoadImages, LoadPilAndNumpy, LoadScreenshots, LoadStreams
from ultralytics.yolo.data.utils import IMG_FORMATS, VID_FORMATS
from ultralytics.yolo.utils import DEFAULT_CONFIG, LOGGER, SETTINGS, callbacks, colorstr, ops
from ultralytics.yolo.utils.checks import check_file, check_imgsz, check_imshow
@ -89,7 +90,6 @@ class BasePredictor:
self.vid_path, self.vid_writer = None, None
self.annotator = None
self.data_path = None
self.output = {}
self.callbacks = defaultdict(list, {k: [v] for k, v in callbacks.default_callbacks.items()}) # add callbacks
callbacks.add_integration_callbacks(self)
@ -99,29 +99,18 @@ class BasePredictor:
def get_annotator(self, img):
raise NotImplementedError("get_annotator function needs to be implemented")
def write_results(self, pred, batch, print_string):
def write_results(self, results, batch, print_string):
raise NotImplementedError("print_results function needs to be implemented")
def postprocess(self, preds, img, orig_img):
return preds
def setup(self, source=None, model=None, return_outputs=False):
def setup(self, source=None, model=None):
# source
source = str(source if source is not None else self.args.source)
is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
webcam = source.isnumeric() or source.endswith('.streams') or (is_url and not is_file)
screenshot = source.lower().startswith('screen')
if is_url and is_file:
source = check_file(source) # download
source, webcam, screenshot, from_img = self.check_source(source)
# model
device = select_device(self.args.device)
model = model or self.args.model
self.args.half &= device.type != 'cpu' # half precision only supported on CUDA
model = AutoBackend(model, device=device, dnn=self.args.dnn, fp16=self.args.half)
stride, pt = model.stride, model.pt
imgsz = check_imgsz(self.args.imgsz, stride=stride) # check image size
stride, pt = self.setup_model(model)
imgsz = check_imgsz(self.args.imgsz, stride=stride, min_dim=2) # check image size
# Dataloader
bs = 1 # batch_size
@ -131,7 +120,7 @@ class BasePredictor:
imgsz=imgsz,
stride=stride,
auto=pt,
transforms=getattr(model.model, 'transforms', None),
transforms=getattr(self.model.model, 'transforms', None),
vid_stride=self.args.vid_stride)
bs = len(self.dataset)
elif screenshot:
@ -139,32 +128,47 @@ class BasePredictor:
imgsz=imgsz,
stride=stride,
auto=pt,
transforms=getattr(model.model, 'transforms', None))
transforms=getattr(self.model.model, 'transforms', None))
elif from_img:
self.dataset = LoadPilAndNumpy(source,
imgsz=imgsz,
stride=stride,
auto=pt,
transforms=getattr(self.model.model, 'transforms', None))
else:
self.dataset = LoadImages(source,
imgsz=imgsz,
stride=stride,
auto=pt,
transforms=getattr(model.model, 'transforms', None),
transforms=getattr(self.model.model, 'transforms', None),
vid_stride=self.args.vid_stride)
self.vid_path, self.vid_writer = [None] * bs, [None] * bs
model.warmup(imgsz=(1 if pt or model.triton else bs, 3, *imgsz)) # warmup
self.model.warmup(imgsz=(1 if pt or self.model.triton else bs, 3, *imgsz)) # warmup
self.model = model
self.webcam = webcam
self.screenshot = screenshot
self.from_img = from_img
self.imgsz = imgsz
self.done_setup = True
self.device = device
self.return_outputs = return_outputs
return model
@smart_inference_mode()
def __call__(self, source=None, model=None, return_outputs=False):
def __call__(self, source=None, model=None, verbose=False, stream=False):
if stream:
return self.stream_inference(source, model, verbose)
else:
return list(chain(*list(self.stream_inference(source, model, verbose)))) # merge list of Result into one
def predict_cli(self):
# Method used for cli prediction. It uses always generator as outputs as not required by cli mode
gen = self.stream_inference(verbose=True)
for _ in gen: # running CLI inference without accumulating any outputs (do not modify)
pass
def stream_inference(self, source=None, model=None, verbose=False):
self.run_callbacks("on_predict_start")
model = self.model if self.done_setup else self.setup(source, model, return_outputs)
model.eval()
if not self.done_setup:
self.setup(source, model)
self.seen, self.windows, self.dt = 0, [], (ops.Profile(), ops.Profile(), ops.Profile())
for batch in self.dataset:
self.run_callbacks("on_predict_batch_start")
@ -177,17 +181,17 @@ class BasePredictor:
# Inference
with self.dt[1]:
preds = model(im, augment=self.args.augment, visualize=visualize)
preds = self.model(im, augment=self.args.augment, visualize=visualize)
# postprocess
with self.dt[2]:
preds = self.postprocess(preds, im, im0s)
results = self.postprocess(preds, im, im0s)
for i in range(len(im)):
if self.webcam:
path, im0s = path[i], im0s[i]
p = Path(path)
s += self.write_results(i, preds, (p, im, im0s))
p, im0 = (path[i], im0s[i]) if self.webcam or self.from_img else (path, im0s)
p = Path(p)
if verbose or self.args.save or self.args.save_txt:
s += self.write_results(i, results, (p, im, im0))
if self.args.show:
self.show(p)
@ -195,30 +199,50 @@ class BasePredictor:
if self.args.save:
self.save_preds(vid_cap, i, str(self.save_dir / p.name))
if self.return_outputs:
yield self.output
self.output.clear()
yield results
# Print time (inference-only)
LOGGER.info(f"{s}{'' if len(preds) else '(no detections), '}{self.dt[1].dt * 1E3:.1f}ms")
if verbose:
LOGGER.info(f"{s}{'' if len(preds) else '(no detections), '}{self.dt[1].dt * 1E3:.1f}ms")
self.run_callbacks("on_predict_batch_end")
# Print results
t = tuple(x.t / self.seen * 1E3 for x in self.dt) # speeds per image
LOGGER.info(
f'Speed: %.1fms pre-process, %.1fms inference, %.1fms postprocess per image at shape {(1, 3, *self.imgsz)}'
% t)
if verbose:
t = tuple(x.t / self.seen * 1E3 for x in self.dt) # speeds per image
LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms postprocess per image at shape '
f'{(1, 3, *self.imgsz)}' % t)
if self.args.save_txt or self.args.save:
s = f"\n{len(list(self.save_dir.glob('labels/*.txt')))} labels saved to {self.save_dir / 'labels'}" if self.args.save_txt else ''
s = f"\n{len(list(self.save_dir.glob('labels/*.txt')))} labels saved to {self.save_dir / 'labels'}" \
if self.args.save_txt else ''
LOGGER.info(f"Results saved to {colorstr('bold', self.save_dir)}{s}")
self.run_callbacks("on_predict_end")
def predict_cli(self, source=None, model=None, return_outputs=False):
# as __call__ is a generator now so have to treat it like a generator
for _ in (self.__call__(source, model, return_outputs)):
pass
def setup_model(self, model):
device = select_device(self.args.device)
model = model or self.args.model
self.args.half &= device.type != 'cpu' # half precision only supported on CUDA
model = AutoBackend(model, device=device, dnn=self.args.dnn, fp16=self.args.half)
self.model = model
self.device = device
self.model.eval()
return model.stride, model.pt
def check_source(self, source):
source = source if source is not None else self.args.source
webcam, screenshot, from_img = False, False, False
if isinstance(source, (str, int, Path)): # int for local usb carame
source = str(source)
is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
webcam = source.isnumeric() or source.endswith('.streams') or (is_url and not is_file)
screenshot = source.lower().startswith('screen')
if is_url and is_file:
source = check_file(source) # download
else:
from_img = True
return source, webcam, screenshot, from_img
def show(self, p):
im0 = self.annotator.result()

284
ultralytics/yolo/engine/results.py Normal file
View File

@ -0,0 +1,284 @@
from functools import lru_cache
import numpy as np
import torch
from ultralytics.yolo.utils import LOGGER, ops
class Results:
"""
A class for storing and manipulating inference results.
Args:
boxes (Boxes, optional): A Boxes object containing the detection bounding boxes.
masks (Masks, optional): A Masks object containing the detection masks.
probs (torch.Tensor, optional): A tensor containing the detection class probabilities.
orig_shape (tuple, optional): Original image size.
Attributes:
boxes (Boxes, optional): A Boxes object containing the detection bounding boxes.
masks (Masks, optional): A Masks object containing the detection masks.
probs (torch.Tensor, optional): A tensor containing the detection class probabilities.
orig_shape (tuple, optional): Original image size.
"""
def __init__(self, boxes=None, masks=None, probs=None, orig_shape=None) -> None:
self.boxes = Boxes(boxes, orig_shape) if boxes is not None else None # native size boxes
self.masks = Masks(masks, orig_shape) if masks is not None else None # native size or imgsz masks
self.probs = probs.softmax(0) if probs is not None else None
self.orig_shape = orig_shape
self.comp = ["boxes", "masks", "probs"]
def pandas(self):
pass
# TODO masks.pandas + boxes.pandas + cls.pandas
def __getitem__(self, idx):
r = Results(orig_shape=self.orig_shape)
for item in self.comp:
if getattr(self, item) is None:
continue
setattr(r, item, getattr(self, item)[idx])
return r
def cpu(self):
r = Results(orig_shape=self.orig_shape)
for item in self.comp:
if getattr(self, item) is None:
continue
setattr(r, item, getattr(self, item).cpu())
return r
def numpy(self):
r = Results(orig_shape=self.orig_shape)
for item in self.comp:
if getattr(self, item) is None:
continue
setattr(r, item, getattr(self, item).numpy())
return r
def cuda(self):
r = Results(orig_shape=self.orig_shape)
for item in self.comp:
if getattr(self, item) is None:
continue
setattr(r, item, getattr(self, item).cuda())
return r
def to(self, *args, **kwargs):
r = Results(orig_shape=self.orig_shape)
for item in self.comp:
if getattr(self, item) is None:
continue
setattr(r, item, getattr(self, item).to(*args, **kwargs))
return r
def __len__(self):
for item in self.comp:
if getattr(self, item) is None:
continue
return len(getattr(self, item))
def __str__(self):
return self.__repr__()
def __repr__(self):
s = f'Ultralytics YOLO {self.__class__} instance\n' # string
if self.boxes:
s = s + self.boxes.__repr__() + '\n'
if self.masks:
s = s + self.masks.__repr__() + '\n'
if self.probs:
s = s + self.probs.__repr__()
s += f'original size: {self.orig_shape}\n'
return s
class Boxes:
"""
A class for storing and manipulating detection boxes.
Args:
boxes (torch.Tensor) or (numpy.ndarray): A tensor or numpy array containing the detection boxes,
with shape (num_boxes, 6). The last two columns should contain confidence and class values.
orig_shape (tuple): Original image size, in the format (height, width).
Attributes:
boxes (torch.Tensor) or (numpy.ndarray): A tensor or numpy array containing the detection boxes,
with shape (num_boxes, 6).
orig_shape (torch.Tensor) or (numpy.ndarray): Original image size, in the format (height, width).
Properties:
xyxy (torch.Tensor) or (numpy.ndarray): The boxes in xyxy format.
conf (torch.Tensor) or (numpy.ndarray): The confidence values of the boxes.
cls (torch.Tensor) or (numpy.ndarray): The class values of the boxes.
xywh (torch.Tensor) or (numpy.ndarray): The boxes in xywh format.
xyxyn (torch.Tensor) or (numpy.ndarray): The boxes in xyxy format normalized by original image size.
xywhn (torch.Tensor) or (numpy.ndarray): The boxes in xywh format normalized by original image size.
"""
def __init__(self, boxes, orig_shape) -> None:
if boxes.ndim == 1:
boxes = boxes[None, :]
assert boxes.shape[-1] == 6 # xyxy, conf, cls
self.boxes = boxes
self.orig_shape = torch.as_tensor(orig_shape, device=boxes.device) if isinstance(boxes, torch.Tensor) \
else np.asarray(orig_shape)
@property
def xyxy(self):
return self.boxes[:, :4]
@property
def conf(self):
return self.boxes[:, -2]
@property
def cls(self):
return self.boxes[:, -1]
@property
@lru_cache(maxsize=2) # maxsize 1 should suffice
def xywh(self):
return ops.xyxy2xywh(self.xyxy)
@property
@lru_cache(maxsize=2)
def xyxyn(self):
return self.xyxy / self.orig_shape[[1, 0, 1, 0]]
@property
@lru_cache(maxsize=2)
def xywhn(self):
return self.xywh / self.orig_shape[[1, 0, 1, 0]]
def cpu(self):
boxes = self.boxes.cpu()
return Boxes(boxes, self.orig_shape)
def numpy(self):
boxes = self.boxes.numpy()
return Boxes(boxes, self.orig_shape)
def cuda(self):
boxes = self.boxes.cuda()
return Boxes(boxes, self.orig_shape)
def to(self, *args, **kwargs):
boxes = self.boxes.to(*args, **kwargs)
return Boxes(boxes, self.orig_shape)
def pandas(self):
LOGGER.info('results.pandas() method not yet implemented')
'''
new = copy(self) # return copy
ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns
cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns
for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update
setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
return new
'''
@property
def shape(self):
return self.boxes.shape
def __len__(self): # override len(results)
return len(self.boxes)
def __str__(self):
return self.__repr__()
def __repr__(self):
return (f"Ultralytics YOLO {self.__class__} masks\n" + f"type: {type(self.boxes)}\n" +
f"shape: {self.boxes.shape}\n" + f"dtype: {self.boxes.dtype}")
def __getitem__(self, idx):
boxes = self.boxes[idx]
return Boxes(boxes, self.orig_shape)
class Masks:
"""
A class for storing and manipulating detection masks.
Args:
masks (torch.Tensor): A tensor containing the detection masks, with shape (num_masks, height, width).
orig_shape (tuple): Original image size, in the format (height, width).
Attributes:
masks (torch.Tensor): A tensor containing the detection masks, with shape (num_masks, height, width).
orig_shape (tuple): Original image size, in the format (height, width).
Properties:
segments (list): A list of segments which includes x,y,w,h,label,confidence, and mask of each detection masks.
"""
def __init__(self, masks, orig_shape) -> None:
self.masks = masks # N, h, w
self.orig_shape = orig_shape
@property
@lru_cache(maxsize=1)
def segments(self):
return [
ops.scale_segments(self.masks.shape[1:], x, self.orig_shape, normalize=True)
for x in reversed(ops.masks2segments(self.masks))]
@property
def shape(self):
return self.masks.shape
def cpu(self):
masks = self.masks.cpu()
return Masks(masks, self.orig_shape)
def numpy(self):
masks = self.masks.numpy()
return Masks(masks, self.orig_shape)
def cuda(self):
masks = self.masks.cuda()
return Masks(masks, self.orig_shape)
def to(self, *args, **kwargs):
masks = self.masks.to(*args, **kwargs)
return Masks(masks, self.orig_shape)
def __len__(self): # override len(results)
return len(self.masks)
def __str__(self):
return self.__repr__()
def __repr__(self):
return (f"Ultralytics YOLO {self.__class__} masks\n" + f"type: {type(self.masks)}\n" +
f"shape: {self.masks.shape}\n" + f"dtype: {self.masks.dtype}")
def __getitem__(self, idx):
masks = self.masks[idx]
return Masks(masks, self.im_shape, self.orig_shape)
if __name__ == "__main__":
# test examples
results = Results(boxes=torch.randn((2, 6)), masks=torch.randn((2, 160, 160)), orig_shape=[640, 640])
results = results.cuda()
print("--cuda--pass--")
results = results.cpu()
print("--cpu--pass--")
results = results.to("cuda:0")
print("--to-cuda--pass--")
results = results.to("cpu")
print("--to-cpu--pass--")
results = results.numpy()
print("--numpy--pass--")
# box = Boxes(boxes=torch.randn((2, 6)), orig_shape=[5, 5])
# box = box.cuda()
# box = box.cpu()
# box = box.numpy()
# for b in box:
# print(b)

6
ultralytics/yolo/engine/trainer.py
View File

@ -30,7 +30,7 @@ from ultralytics.yolo.data.utils import check_dataset, check_dataset_yaml
from ultralytics.yolo.utils import (DEFAULT_CONFIG, LOGGER, RANK, SETTINGS, TQDM_BAR_FORMAT, callbacks, colorstr,
yaml_save)
from ultralytics.yolo.utils.autobatch import check_train_batch_size
from ultralytics.yolo.utils.checks import check_file, print_args
from ultralytics.yolo.utils.checks import check_file, check_imgsz, print_args
from ultralytics.yolo.utils.dist import ddp_cleanup, generate_ddp_command
from ultralytics.yolo.utils.files import get_latest_run, increment_path
from ultralytics.yolo.utils.torch_utils import ModelEMA, de_parallel, init_seeds, one_cycle, strip_optimizer
@ -203,7 +203,9 @@ class BaseTrainer:
self.set_model_attributes()
if world_size > 1:
self.model = DDP(self.model, device_ids=[rank])
# Check imgsz
gs = max(int(self.model.stride.max() if hasattr(self.model, 'stride') else 32), 32) # grid size (max stride)
self.args.imgsz = check_imgsz(self.args.imgsz, stride=gs, floor=gs * 2)
# Batch size
if self.batch_size == -1:
if RANK == -1: # single-GPU only, estimate best batch size