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

ultralytics 8.0.53 DDP AMP and Edge TPU fixes (#1362)

Browse Source 
Co-authored-by: Richard Aljaste <richardaljasteabramson@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Vuong Kha Sieu <75152429+hotfur@users.noreply.github.com>
This commit is contained in:
Glenn Jocher
2023-03-12 02:08:13 +01:00
committed by GitHub
parent 177a68b39f
commit f921e1ac21
46 changed files with 1045 additions and 384 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ultralytics/yolo/cfg/__init__.py
View File

@ -299,7 +299,7 @@ def entrypoint(debug=''):
task = model.task
# Mode
if mode in {'predict', 'track'} and 'source' not in overrides:
if mode in ('predict', 'track') and 'source' not in overrides:
overrides['source'] = DEFAULT_CFG.source or ROOT / 'assets' if (ROOT / 'assets').exists() \
else 'https://ultralytics.com/images/bus.jpg'
LOGGER.warning(f"WARNING ⚠️ 'source' is missing. Using default 'source={overrides['source']}'.")

14
ultralytics/yolo/data/augment.py
View File

@ -14,7 +14,7 @@ from ..utils.checks import check_version
from ..utils.instance import Instances
from ..utils.metrics import bbox_ioa
from ..utils.ops import segment2box
from .utils import IMAGENET_MEAN, IMAGENET_STD, polygons2masks, polygons2masks_overlap
from .utils import polygons2masks, polygons2masks_overlap
# TODO: we might need a BaseTransform to make all these augments be compatible with both classification and semantic
@ -682,12 +682,14 @@ def v8_transforms(dataset, imgsz, hyp):
# Classification augmentations -----------------------------------------------------------------------------------------
def classify_transforms(size=224):
def classify_transforms(size=224, mean=(0.0, 0.0, 0.0), std=(1.0, 1.0, 1.0)): # IMAGENET_MEAN, IMAGENET_STD
# Transforms to apply if albumentations not installed
if not isinstance(size, int):
raise TypeError(f'classify_transforms() size {size} must be integer, not (list, tuple)')
# T.Compose([T.ToTensor(), T.Resize(size), T.CenterCrop(size), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])
return T.Compose([CenterCrop(size), ToTensor(), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])
if any(mean) or any(std):
return T.Compose([CenterCrop(size), ToTensor(), T.Normalize(mean, std, inplace=True)])
else:
return T.Compose([CenterCrop(size), ToTensor()])
def classify_albumentations(
@ -697,8 +699,8 @@ def classify_albumentations(
hflip=0.5,
vflip=0.0,
jitter=0.4,
mean=IMAGENET_MEAN,
std=IMAGENET_STD,
mean=(0.0, 0.0, 0.0), # IMAGENET_MEAN
std=(1.0, 1.0, 1.0), # IMAGENET_STD
auto_aug=False,
):
# YOLOv8 classification Albumentations (optional, only used if package is installed)

2
ultralytics/yolo/data/dataloaders/v5loader.py
View File

@ -496,7 +496,7 @@ class LoadImagesAndLabels(Dataset):
# Display cache
nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupt, total
if exists and LOCAL_RANK in {-1, 0}:
if exists and LOCAL_RANK in (-1, 0):
d = f'Scanning {cache_path}... {nf} images, {nm + ne} backgrounds, {nc} corrupt'
tqdm(None, desc=prefix + d, total=n, initial=n, bar_format=TQDM_BAR_FORMAT) # display cache results
if cache['msgs']:

2
ultralytics/yolo/data/dataset.py
View File

@ -133,7 +133,7 @@ class YOLODataset(BaseDataset):
# Display cache
nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupt, total
if exists and LOCAL_RANK in {-1, 0}:
if exists and LOCAL_RANK in (-1, 0):
d = f'Scanning {cache_path}... {nf} images, {nm + ne} backgrounds, {nc} corrupt'
tqdm(None, desc=self.prefix + d, total=n, initial=n, bar_format=TQDM_BAR_FORMAT) # display cache results
if cache['msgs']:

7
ultralytics/yolo/engine/exporter.py
View File

@ -63,7 +63,6 @@ from ultralytics.nn.autobackend import check_class_names
from ultralytics.nn.modules import C2f, Detect, Segment
from ultralytics.nn.tasks import DetectionModel, SegmentationModel
from ultralytics.yolo.cfg import get_cfg
from ultralytics.yolo.data.utils import IMAGENET_MEAN, IMAGENET_STD
from ultralytics.yolo.utils import (DEFAULT_CFG, LINUX, LOGGER, MACOS, __version__, callbacks, colorstr,
get_default_args, yaml_save)
from ultralytics.yolo.utils.checks import check_imgsz, check_requirements, check_version
@ -148,7 +147,7 @@ class Exporter:
self.run_callbacks('on_export_start')
t = time.time()
format = self.args.format.lower() # to lowercase
if format in {'tensorrt', 'trt'}: # engine aliases
if format in ('tensorrt', 'trt'): # engine aliases
format = 'engine'
fmts = tuple(export_formats()['Argument'][1:]) # available export formats
flags = [x == format for x in fmts]
@ -408,8 +407,6 @@ class Exporter:
scale = 1 / 255
classifier_config = None
if self.model.task == 'classify':
bias = [-x for x in IMAGENET_MEAN]
scale = 1 / 255 / (sum(IMAGENET_STD) / 3)
classifier_config = ct.ClassifierConfig(list(self.model.names.values())) if self.args.nms else None
model = self.model
elif self.model.task == 'detect':
@ -531,7 +528,7 @@ class Exporter:
# Export to TF
int8 = '-oiqt -qt per-tensor' if self.args.int8 else ''
cmd = f'onnx2tf -i {f_onnx} -o {f} -nuo --non_verbose {int8}'
LOGGER.info(f"\n{prefix} running '{cmd}'")
LOGGER.info(f"\n{prefix} running '{cmd.strip()}'")
subprocess.run(cmd, shell=True)
yaml_save(f / 'metadata.yaml', self.metadata) # add metadata.yaml

2
ultralytics/yolo/engine/model.py
View File

@ -319,7 +319,7 @@ class YOLO:
self.trainer.hub_session = self.session # attach optional HUB session
self.trainer.train()
# update model and cfg after training
if RANK in {0, -1}:
if RANK in (-1, 0):
self.model, _ = attempt_load_one_weight(str(self.trainer.best))
self.overrides = self.model.args
self.metrics = getattr(self.trainer.validator, 'metrics', None) # TODO: no metrics returned by DDP

2
ultralytics/yolo/engine/results.py
View File

@ -185,7 +185,7 @@ class Boxes:
if boxes.ndim == 1:
boxes = boxes[None, :]
n = boxes.shape[-1]
assert n in {6, 7}, f'expected `n` in [6, 7], but got {n}' # xyxy, (track_id), conf, cls
assert n in (6, 7), f'expected `n` in [6, 7], but got {n}' # xyxy, (track_id), conf, cls
# TODO
self.is_track = n == 7
self.boxes = boxes

51
ultralytics/yolo/engine/trainer.py
View File

@ -95,9 +95,9 @@ class BaseTrainer:
self.save_dir = Path(self.args.save_dir)
else:
self.save_dir = Path(
increment_path(Path(project) / name, exist_ok=self.args.exist_ok if RANK in {-1, 0} else True))
increment_path(Path(project) / name, exist_ok=self.args.exist_ok if RANK in (-1, 0) else True))
self.wdir = self.save_dir / 'weights' # weights dir
if RANK in {-1, 0}:
if RANK in (-1, 0):
self.wdir.mkdir(parents=True, exist_ok=True) # make dir
self.args.save_dir = str(self.save_dir)
yaml_save(self.save_dir / 'args.yaml', vars(self.args)) # save run args
@ -144,7 +144,7 @@ class BaseTrainer:
# Callbacks
self.callbacks = defaultdict(list, callbacks.default_callbacks) # add callbacks
if RANK in {0, -1}:
if RANK in (-1, 0):
callbacks.add_integration_callbacks(self)
def add_callback(self, event: str, callback):
@ -203,9 +203,14 @@ class BaseTrainer:
self.model = self.model.to(self.device)
self.set_model_attributes()
# Check AMP
callbacks_backup = callbacks.default_callbacks.copy() # backup callbacks as they are reset by check_amp()
self.amp = check_amp(self.model)
callbacks.default_callbacks = callbacks_backup # restore callbacks
self.amp = torch.tensor(True).to(self.device)
if RANK in (-1, 0): # Single-GPU and DDP
callbacks_backup = callbacks.default_callbacks.copy() # backup callbacks as check_amp() resets them
self.amp = torch.tensor(check_amp(self.model), device=self.device)
callbacks.default_callbacks = callbacks_backup # restore callbacks
if RANK > -1: # DDP
dist.broadcast(self.amp, src=0) # broadcast the tensor from rank 0 to all other ranks (returns None)
self.amp = bool(self.amp) # as boolean
self.scaler = amp.GradScaler(enabled=self.amp)
if world_size > 1:
self.model = DDP(self.model, device_ids=[rank])
@ -239,7 +244,7 @@ class BaseTrainer:
# dataloaders
batch_size = self.batch_size // world_size if world_size > 1 else self.batch_size
self.train_loader = self.get_dataloader(self.trainset, batch_size=batch_size, rank=rank, mode='train')
if rank in {0, -1}:
if rank in (-1, 0):
self.test_loader = self.get_dataloader(self.testset, batch_size=batch_size * 2, rank=-1, mode='val')
self.validator = self.get_validator()
metric_keys = self.validator.metrics.keys + self.label_loss_items(prefix='val')
@ -286,7 +291,7 @@ class BaseTrainer:
if hasattr(self.train_loader.dataset, 'close_mosaic'):
self.train_loader.dataset.close_mosaic(hyp=self.args)
if rank in {-1, 0}:
if rank in (-1, 0):
LOGGER.info(self.progress_string())
pbar = tqdm(enumerate(self.train_loader), total=nb, bar_format=TQDM_BAR_FORMAT)
self.tloss = None
@ -327,7 +332,7 @@ class BaseTrainer:
mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB)
loss_len = self.tloss.shape[0] if len(self.tloss.size()) else 1
losses = self.tloss if loss_len > 1 else torch.unsqueeze(self.tloss, 0)
if rank in {-1, 0}:
if rank in (-1, 0):
pbar.set_description(
('%11s' * 2 + '%11.4g' * (2 + loss_len)) %
(f'{epoch + 1}/{self.epochs}', mem, *losses, batch['cls'].shape[0], batch['img'].shape[-1]))
@ -342,7 +347,7 @@ class BaseTrainer:
self.scheduler.step()
self.run_callbacks('on_train_epoch_end')
if rank in {-1, 0}:
if rank in (-1, 0):
# Validation
self.ema.update_attr(self.model, include=['yaml', 'nc', 'args', 'names', 'stride', 'class_weights'])
@ -372,7 +377,7 @@ class BaseTrainer:
if self.stop:
break # must break all DDP ranks
if rank in {-1, 0}:
if rank in (-1, 0):
# Do final val with best.pt
LOGGER.info(f'\n{epoch - self.start_epoch + 1} epochs completed in '
f'{(time.time() - self.train_time_start) / 3600:.3f} hours.')
@ -603,7 +608,20 @@ class BaseTrainer:
def check_amp(model):
# Check PyTorch Automatic Mixed Precision (AMP) functionality. Return True on correct operation
"""
This function checks the PyTorch Automatic Mixed Precision (AMP) functionality of a YOLOv8 model.
If the checks fail, it means there are anomalies with AMP on the system that may cause NaN losses or zero-mAP
results, so AMP will be disabled during training.
Args:
model (nn.Module): A YOLOv8 model instance.
Returns:
bool: Returns True if the AMP functionality works correctly with YOLOv8 model, else False.
Raises:
AssertionError: If the AMP checks fail, indicating anomalies with the AMP functionality on the system.
"""
device = next(model.parameters()).device # get model device
if device.type in ('cpu', 'mps'):
return False # AMP only used on CUDA devices
@ -613,18 +631,21 @@ def check_amp(model):
a = m(im, device=device, verbose=False)[0].boxes.boxes # FP32 inference
with torch.cuda.amp.autocast(True):
b = m(im, device=device, verbose=False)[0].boxes.boxes # AMP inference
return a.shape == b.shape and torch.allclose(a, b.float(), rtol=0.1) # close to 10% absolute tolerance
del m
return a.shape == b.shape and torch.allclose(a, b.float(), atol=0.5) # close to 0.5 absolute tolerance
f = ROOT / 'assets/bus.jpg' # image to check
im = f if f.exists() else 'https://ultralytics.com/images/bus.jpg' if ONLINE else np.ones((640, 640, 3))
prefix = colorstr('AMP: ')
LOGGER.info(f'{prefix}running Automatic Mixed Precision (AMP) checks with YOLOv8n...')
try:
from ultralytics import YOLO
LOGGER.info(f'{prefix}running Automatic Mixed Precision (AMP) checks with YOLOv8n...')
assert amp_allclose(YOLO('yolov8n.pt'), im)
LOGGER.info(f'{prefix}checks passed ✅')
return True
except ConnectionError:
LOGGER.warning(f"{prefix}checks skipped ⚠️, offline and unable to download YOLOv8n. Setting 'amp=True'.")
except AssertionError:
LOGGER.warning(f'{prefix}checks failed ❌. Anomalies were detected with AMP on your system that may lead to '
f'NaN losses or zero-mAP results, so AMP will be disabled during training.')
return False
return True

2
ultralytics/yolo/engine/validator.py
View File

@ -79,7 +79,7 @@ class BaseValidator:
project = self.args.project or Path(SETTINGS['runs_dir']) / self.args.task
name = self.args.name or f'{self.args.mode}'
self.save_dir = save_dir or increment_path(Path(project) / name,
exist_ok=self.args.exist_ok if RANK in {-1, 0} else True)
exist_ok=self.args.exist_ok if RANK in (-1, 0) else True)
(self.save_dir / 'labels' if self.args.save_txt else self.save_dir).mkdir(parents=True, exist_ok=True)
if self.args.conf is None:

4
ultralytics/yolo/utils/__init__.py
View File

@ -126,7 +126,7 @@ class IterableSimpleNamespace(SimpleNamespace):
def set_logging(name=LOGGING_NAME, verbose=True):
# sets up logging for the given name
rank = int(os.getenv('RANK', -1)) # rank in world for Multi-GPU trainings
level = logging.INFO if verbose and rank in {-1, 0} else logging.ERROR
level = logging.INFO if verbose and rank in (-1, 0) else logging.ERROR
logging.config.dictConfig({
'version': 1,
'disable_existing_loggers': False,
@ -524,7 +524,7 @@ def set_sentry():
return event
if SETTINGS['sync'] and \
RANK in {-1, 0} and \
RANK in (-1, 0) and \
Path(sys.argv[0]).name == 'yolo' and \
not TESTS_RUNNING and \
ONLINE and \

4
ultralytics/yolo/utils/benchmarks.py
View File

@ -28,7 +28,7 @@ from pathlib import Path
from ultralytics import YOLO
from ultralytics.yolo.engine.exporter import export_formats
from ultralytics.yolo.utils import LINUX, LOGGER, ROOT, SETTINGS
from ultralytics.yolo.utils import LINUX, LOGGER, MACOS, ROOT, SETTINGS
from ultralytics.yolo.utils.checks import check_yolo
from ultralytics.yolo.utils.downloads import download
from ultralytics.yolo.utils.files import file_size
@ -51,6 +51,8 @@ def benchmark(model=Path(SETTINGS['weights_dir']) / 'yolov8n.pt', imgsz=160, hal
if model.task == 'classify':
assert i != 11, 'paddle cls exports coming soon'
assert i != 9 or LINUX, 'Edge TPU export only supported on Linux'
if i == 10:
assert MACOS or LINUX, 'TF.js export only supported on macOS and Linux'
if 'cpu' in device.type:
assert cpu, 'inference not supported on CPU'
if 'cuda' in device.type:

2
ultralytics/yolo/utils/downloads.py
View File

@ -118,7 +118,7 @@ def safe_download(url,
raise ConnectionError(f'❌ Download failure for {url}. Retry limit reached.') from e
LOGGER.warning(f'⚠️ Download failure, retrying {i + 1}/{retry} {url}...')
if unzip and f.exists() and f.suffix in {'.zip', '.tar', '.gz'}:
if unzip and f.exists() and f.suffix in ('.zip', '.tar', '.gz'):
unzip_dir = dir or f.parent # unzip to dir if provided else unzip in place
LOGGER.info(f'Unzipping {f} to {unzip_dir}...')
if f.suffix == '.zip':

2
ultralytics/yolo/utils/torch_utils.py
View File

@ -33,7 +33,7 @@ TORCH_1_12 = check_version(torch.__version__, '1.12.0')
def torch_distributed_zero_first(local_rank: int):
# Decorator to make all processes in distributed training wait for each local_master to do something
initialized = torch.distributed.is_available() and torch.distributed.is_initialized()
if initialized and local_rank not in {-1, 0}:
if initialized and local_rank not in (-1, 0):
dist.barrier(device_ids=[local_rank])
yield
if initialized and local_rank == 0:

2
ultralytics/yolo/v8/classify/val.py
View File

@ -43,6 +43,8 @@ class ClassificationValidator(BaseValidator):
return build_classification_dataloader(path=dataset_path,
imgsz=self.args.imgsz,
batch_size=batch_size,
augment=False,
shuffle=False,
workers=self.args.workers)
def print_results(self):

4
ultralytics/yolo/v8/detect/predict.py
View File

@ -30,8 +30,8 @@ class DetectionPredictor(BasePredictor):
results = []
for i, pred in enumerate(preds):
orig_img = orig_imgs[i] if isinstance(orig_imgs, list) else orig_imgs
shape = orig_img.shape
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], shape).round()
if not isinstance(orig_imgs, torch.Tensor):
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], orig_img.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))

9
ultralytics/yolo/v8/segment/predict.py
View File

@ -23,18 +23,19 @@ class SegmentationPredictor(DetectionPredictor):
proto = preds[1][-1] if len(preds[1]) == 3 else preds[1] # second output is len 3 if pt, but only 1 if exported
for i, pred in enumerate(p):
orig_img = orig_imgs[i] if isinstance(orig_imgs, list) else orig_imgs
shape = orig_img.shape
path, _, _, _, _ = self.batch
img_path = path[i] if isinstance(path, list) else path
if not len(pred): # save empty boxes
results.append(Results(orig_img=orig_img, path=img_path, names=self.model.names, boxes=pred[:, :6]))
continue
if self.args.retina_masks:
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], shape).round()
masks = ops.process_mask_native(proto[i], pred[:, 6:], pred[:, :4], shape[:2]) # HWC
if not isinstance(orig_imgs, torch.Tensor):
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], orig_img.shape)
masks = ops.process_mask_native(proto[i], pred[:, 6:], pred[:, :4], orig_img.shape[:2]) # HWC
else:
masks = ops.process_mask(proto[i], pred[:, 6:], pred[:, :4], img.shape[2:], upsample=True) # HWC
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], shape).round()
if not isinstance(orig_imgs, torch.Tensor):
pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], orig_img.shape)
results.append(
Results(orig_img=orig_img, path=img_path, names=self.model.names, boxes=pred[:, :6], masks=masks))
return results