Fi֢ddlmZddlZddlZddlZddlZddlZddlZddlm Z ddl m Z ddl m Z ddl mZddlmZddlZddlZddlmZddlmZddlmcmZddlmZdd lmZm Z m!Z!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'dd l(m)Z)dd l*m+Z+dd l,m-Z-e)e$d Z.e)e$dZ/e)e$dZ0e)e$dZ1e)e$dZ2e)e$dZ3e)e$dZ4e)e$dZ5e)e$dZ6e)e$dZ7e)e$dZ8e)e%dZ9e)e%dZ:e)e%dZ;e)e%dZdZ?d@dAd Z@ejd!ZBejd"ZCdBd#ZDd$ZEd%ZFd&ZGdCd'ZHd(ZId)ZJd*ZKdDd+ZLdDd,ZMd-ZNdEd.ZOdFd/ZPdGd0ZQd1ZRdHd2ZSdId3ZTdJd4ZUd5ZVGd6d7ZWdKdLd8ZXd9ZYe dMd:ZZdNd;Z[Gd<d=Z\ dO dPd>Z]y)Q) annotationsN)contextmanager)deepcopy)datetime)Path)Any) __version__) DEFAULT_CFG_DICTDEFAULT_CFG_KEYSLOGGER NUM_THREADSPYTHON_VERSION TORCH_VERSIONTORCHVISION_VERSIONWINDOWScolorstr) check_version)CPUInfo) torch_loadz1.9.0z1.10.0z1.11.0z1.13.0z2.0.0z2.1.0z2.3.0z2.4.0z2.8.0z2.9.0z2.10.0z0.10.0z0.11.0z0.13.0z0.18.0z==2.4.0zKnown issue with torch==2.4.0 on Windows with CPU, recommend upgrading to torch>=2.4.1 to resolve https://github.com/ultralytics/ultralytics/issues/15049c#pKtjxrtj}|xrtjdk(}|r1|dvr-|rtj|gntjd|r4|dk(r.|rtj|gntjyyyw)ziEnsure all processes in distributed training wait for the local master (rank 0) to complete a task first.nccl>r) device_idsNr)dist is_availableis_initialized get_backendbarrier) local_rank initializeduse_idss _/var/www/pod-logistic/pod-ai/venv/lib/python3.12/site-packages/ultralytics/utils/torch_utils.pytorch_distributed_zero_firstr#>s##%?$*=*=*?K:d..0F:Gz018  -dlln zQ18  -dlln'{sB4B6c d}|S)zXApply torch.inference_mode() decorator if torch>=1.10.0, else torch.no_grad() decorator.ctrtjr|Strtj|Stj |S)zLApply appropriate torch decorator for inference mode based on torch version.) TORCH_1_9torchis_inference_mode_enabled TORCH_1_10inference_modeno_grad)fns r"decoratez&smart_inference_mode..decorateNs> 88:ILJE((LRP PEMMLRP P)r-s r"smart_inference_moder0KsQ Or.ctr!tjj||Stjjj|S)aHGet the appropriate autocast context manager based on PyTorch version and AMP setting. This function returns a context manager for automatic mixed precision (AMP) training that is compatible with both older and newer versions of PyTorch. It handles the differences in the autocast API between PyTorch versions. Args: enabled (bool): Whether to enable automatic mixed precision. device (str, optional): The device to use for autocast. Returns: (torch.amp.autocast): The appropriate autocast context manager. Examples: >>> with autocast(enabled=True): ... # Your mixed precision operations here ... pass Notes: - For PyTorch versions 1.13 and newer, it uses `torch.amp.autocast`. - For older versions, it uses `torch.cuda.amp.autocast`. )enabled) TORCH_1_13r'ampautocastcuda)r2devices r"r5r5Xs:,yy!!&'!::zz~~&&w//r.cddlm}d|vr tj|d<|j ddS#t$rYwxYw)z=Return a string with system CPU information, i.e. 'Apple M2'.rPERSISTENT_CACHEcpu_infounknown)ultralytics.utilsr:rname Exceptiongetr9s r" get_cpu_inforAtsO3)) +2<<> Z (    I 66   s5 AActjj|}|jd|jdz ddS)zGReturn a string with system GPU information, i.e. 'Tesla T4, 15102MiB'., iz.0fMiB)r'r6get_device_propertiesr> total_memory)index propertiess r" get_gpu_inforIs>11%8Joo b!8!8G!DS I MMr.c Z t|tjst|j dr|Sdt dt dtd}t|j}dD]}|j|d}|j dr: d d l }ttdrtjjstd |d |dd }|dk(rd }nB|j dr"|dd j!rt#|dd }ntd |dtjj%}||k\rtd |d|dtjj'||r;t)j*|d|dtjj-|dtjd|Sd|vrd dlm} |j3d} | j5| j7dd} t9t;| D],} | | dk(s | rt| j=d nd| | <.dj?d| D}|dk(} |dv}| s|rdt@jBd <nU|rR|d!k(rd"}d|vr1dj?|j3dDcgc]}|s| c}}t@jBjEd d }|t@jBd <tjFjr9tjFj%t;|j3dk\st)j*|tjFj%d k(rd#nd}td$|d%tjFjd&tjFj%d'|d(| | sw|sutjFjrW|r|j3dnd"}dt;|z}tI|D]#\} }|| d k(rdn|d)|dtK| dz }%d*}nW|rBtLr>> select_device("cuda:0") device(type='cuda', index=0) >>> select_device("cpu") device(type='cpu') Notes: Sets the 'CUDA_VISIBLE_DEVICES' environment variable for specifying which GPUs to use. )tpuintelvulkanz Ultralytics u 🚀 Python-z torch- )zcuda:none()[]'rNnpurNzInvalid NPU 'device=z;'. Install 'torch_npu' at https://github.com/Ascend/pytorchz)' requested. Ascend NPU is not available.:z' format. Use 'npu' or 'npu:0'.z' requested. Only z NPU(s) available.zNPU:z (z) znpu:z-1)GPUInfo,g?)countmin_memory_fractionc3&K|] }|s| ywNr/).0ps r" z select_device..s0a!0scpu>mps:0mpsCUDA_VISIBLE_DEVICESr60z}See https://pytorch.org/get-started/locally/ for up-to-date torch install instructions if no CUDA devices are seen by torch. zInvalid CUDA 'device=z' requested. Use 'device=cpu' or pass valid CUDA device(s) if available, i.e. 'device=0' or 'device=0,1,2,3' for Multi-GPU. torch.cuda.is_available(): z torch.cuda.device_count(): z% os.environ['CUDA_VISIBLE_DEVICES']:  zCUDA:zcuda:0zMPS (rezCPU (>rcre)- isinstancer'r7str startswithr rrlowerreplace torch_npu ImportError ValueErrorhasattrrVrisdigitint device_count set_devicer infoget_device_nameultralytics.utils.autodevicerZsplitselect_idle_gpur\rangelenpopjoinosenvironr@r6 enumeraterI TORCH_2_0backendsrerAset_num_threadsr rstrip)r7newlineverbosesremovernsuffixidxnrZpartsselectedircrexvisibleinstalldevicesspacedargs r" select_devicers6&%,,'3v;+A+AB\+] {m=0@ VWXA [   F?,+, y ue$EII,B,B,D3F8;def f R<C   s #qr (:(:(<fQRj/C3F8;Z[\ \ II " " $ !83F8;MaSPbcd d S!  KK1#T#b)B)B3)G(HL M||d3%L)) v~8 S!9,,5;;t3DZ],^s5z" DAQx43;3x||A/a D0U00 E/C $ $C c-/ )*  V F &=XX&,,s*;AQqqABF**..!7>-3 )* '')ejj.E.E.G3v||\_O`Ka.a KKN::**,14  'x0016 0G0G0I/J/ 0G0G0I/J8 )   suzz668'-&,,s#3c!f g& LDAq !q&Be,E!B|A6GsK KA L u~~11>>@ u\^$C (( u\^$C (( n k* Aahhj1 << g y3F8;vwx x yXBsT 8T(T( T%ctjjrtjjt jS)zReturn PyTorch-accurate time.)r'r6r synchronizetimer/r.r" time_syncrs. zz   99;r.c|jj|jd}tj|jj tj |j|jz}tj||j|jj|j_ |j5tj|j|jjn |j}|j|jj|j j tj |j|jzz }tj||j#ddj#d|z}|j&|j%dt'j(|n||j_ |j+dS)aFuse Conv2d and BatchNorm2d layers for inference optimization. Args: conv (nn.Conv2d): Convolutional layer to fuse. bn (nn.BatchNorm2d): Batch normalization layer to fuse. Returns: (nn.Conv2d): The fused convolutional layer with gradients disabled. Examples: >>> conv = nn.Conv2d(3, 16, 3) >>> bn = nn.BatchNorm2d(16) >>> fused_conv = fuse_conv_and_bn(conv, bn) rr7rYbiasFweightview out_channelsr'diagdivsqrteps running_varmmshapedatarzerosr7mul running_meanreshaperegister_parameternn Parameterrequires_grad_)convbnw_convw_bnb_convb_bn fused_biass r"fuse_conv_and_bnr s` [[  d// 4F ::biimmEJJrvv/F$GH IDxxf-224;;3D3DEDKKKO))J[U[[**4;;3E3E FaeajajF 77RYY]]2??377 2>>TVTZTZCZ8[\ \D$r1 56>>rBTIJ yy  Z(@A#    u %%r.c|jj|jd}tj|jj tj |j|jz}tj||j|jj|j_ |j5tj|j|jjn |j}|j|jj|j j tj |j|jzz }tj||j#ddj#d|z}|j&|j%dt'j(|n||j_ |j+dS)aFuse ConvTranspose2d and BatchNorm2d layers for inference optimization. Args: deconv (nn.ConvTranspose2d): Transposed convolutional layer to fuse. bn (nn.BatchNorm2d): Batch normalization layer to fuse. Returns: (nn.ConvTranspose2d): The fused transposed convolutional layer with gradients disabled. Examples: >>> deconv = nn.ConvTranspose2d(16, 3, 3) >>> bn = nn.BatchNorm2d(3) >>> fused_deconv = fuse_deconv_and_bn(deconv, bn) rrrYrFr)deconvrw_deconvrrrrs r"fuse_deconv_and_bnr-sa }}!!&"5"5r:H ::biimmEJJrvv/F$GH ID$166v}}7J7JKFMMOUkkNaU[[,,V]]5I5I JgmgrgrF 77RYY]]2??377 2>>TVTZTZCZ8[\ \D$r1 56>>rBTIJ {{!!&",,z*BC%    ''r.c|syt|}t|}tdjd|j D}t |}|rzdddddd d d d d dd dd dd }t j|t|jD].\} \} } | jdd} | jj} t | jr| jD]\} }t j| d| d| d| d |jd |j!dt#|j$d |j'd|j)dt+|j,jddd t j| d| d| d dd ddgd dd dd dd 1t/||}t1|ddrdnd}|rd|d d!nd}t1|d"dxst1|d#ij3d"d}t5|j6jd$d%xsd&}t j|d'|d(|d)d*|d)d+|d)d,| ||||fS)-a$Print and return detailed model information layer by layer. Args: model (nn.Module): Model to analyze. detailed (bool, optional): Whether to print detailed layer information. verbose (bool, optional): Whether to print model information. imgsz (int | list, optional): Input image size. Returns: (tuple): Tuple containing: - n_l (int): Number of layers. - n_p (int): Number of parameters. - n_g (int): Number of gradients. - flops (float): GFLOPs. N collectionsc3\K|]$\}}t|jdk(s||f&yw)rN)r|_modules)r`rms r"rbzmodel_info..bs-2tda_bcdcmcm_nrs_sAq62ts , ,layerz>5r>z>40typez>20gradientz>10 parametersz>12rmusigmaz module_list.rUz>5g.z>12gz>10.3gztorch.z>15Fr-is_fusedcy)NFr/r/r.r"zmodel_info..ssr.z (fused)rC.1fz GFLOPs yaml_fileyamlyoloYOLOModelz summaryz: r[z layers, z parameters, z gradients)get_num_paramsget_num_gradients __import__ OrderedDict named_modulesr|r rvritemsrm __class____name__ _parametersnamed_parameters requires_gradnumellistrmeanstdrjdtype get_flopsgetattrr@rstem)modeldetailedrimgszn_pn_glayersn_lhrmnrmtpnraflopsfusedfsr model_names r" model_inforNs"   C E "C  & 2 22teFYFYF[2t tF f+Crl6#,vcl:c2B.}s5Qqwwy5sumrrs r"rr{s 5%"2"2"45 55r.cBtd|jDS)zEReturn the total number of parameters with gradients in a YOLO model.c3VK|]!}|js|j#ywr_)rrrs r"rbz$get_num_gradients..sHQqwwyHs))rrs r"rrs H%"2"2"4H HHr.cx|jjrHddlm}||jg|j j d}|jdn5t|jtt|jdd}t|jjdd|d<|S) aVReturn model info dict with useful model information. Args: trainer (ultralytics.engine.trainer.BaseTrainer): The trainer object containing model and validation data. Returns: (dict): Dictionary containing model parameters, GFLOPs, and inference speeds. Examples: YOLOv8n info for loggers >>> results = { ... "model/parameters": 3151904, ... "model/GFLOPs": 8.746, ... "model/speed_ONNX(ms)": 41.244, ... "model/speed_TensorRT(ms)": 3.211, ... "model/speed_PyTorch(ms)": 18.755, ...} r) ProfileModelsrz model/namerW)zmodel/parametersz model/GFLOPs inferencezmodel/speed_PyTorch(ms))argsprofileultralytics.utils.benchmarksrlastr7runr}rrroundr validatorspeed)trainerrresultss r"model_info_for_loggersrs&||>w~~FJJLQO L!!/w}} =!)GMM":A> */w/@/@/F/F{/SUV)WG %& Nr.c ddl}|sy t|}t|j }t |t s||g} t|dr-tt|jjdnd}tjd|jd||f|j}|jt!||gd dd z d z}||dz|z |dz|z S#t$rd}YwxYw#t"$r]tjd|jdg||j}|jt!||gd dd z d zcYSwxYw#t"$rYywxYw) aCalculate FLOPs (floating point operations) for a model in GFLOPs. Attempts two calculation methods: first with a stride-based tensor for efficiency, then falls back to full image size if needed (e.g., for RTDETR models). Returns 0.0 if thop library is unavailable or calculation fails. Args: model (nn.Module): The model to calculate FLOPs for. imgsz (int | list, optional): Input image size. Returns: (float): The model's GFLOPs (billions of floating point operations). rNstride rYrFinputsreA)thopro unwrap_modelnextrrirrqmaxrsrr'emptyrr7rrr?)rrrrarimrs r"rrsv U# !!# $%&ENE Z9@9QSU\\--/0"5WYFaVV.9AGG9rrc34K|]}|jywr_r!rs r"rbz0get_flops_with_torch_profiler..r"r#)rrrrrirrqrrsrr'rrr7profilerrr key_averagesr?)rrrarrprofrs r"get_flops_with_torch_profilerr(s   E U    A eT " @6=eX6N#c%,,**,-r2TVZ[[ [[!QWWQZ8 J ^^ # #t # 4  "I 9T%6%6%899C?a 6)E!H4v= L   @ [[!QWWQZ0%0 B ^^ # #t # 4  "I   9T%6%6%899C? L @sDBD' D?D'D$ D''AG F GF *GGcZ|jD]}t|}|tjur!|tjurd|_d|_B|tjtjtjtjtjhvsd|_ y)zHInitialize model weights, biases, and module settings to default values.gMbP?gQ?TN) modulesrrConv2d BatchNorm2drmomentum Hardswish LeakyReLUReLUReLU6SiLUinplace)rrts r"initialize_weightsr5sv ]]_ G  >  ".. AEAJ 2<<rww"''J JAIr.c dk(r|S|jdd\}}t|zt|zf}tj||dd}|sfd||fD\}}tj|d||d z d||dz gd S) aScale and pad an image tensor, optionally maintaining aspect ratio and padding to gs multiple. Args: img (torch.Tensor): Input image tensor. ratio (float, optional): Scaling ratio. same_shape (bool, optional): Whether to maintain the same shape. gs (int, optional): Grid size for padding. Returns: (torch.Tensor): Scaled and padded image tensor. ?rNbilinearF)sizemode align_cornersc3ZK|]"}tj|zz z$ywr_)mathceil)r`rgsratios r"rbzscale_img..s'?1 !e)b.)B.?s(+rrYgS㥛?)value)rrsF interpolatepad)imgr@ same_shaper?rwrs ` ` r" scale_imgrHs | 99QR=DAq QYQY(A --!*E JC ?A?1 55q!ad(Aq1Q4x0 >>r.c|jjD]7\}}t|r||vs|jds||vr+t |||9y)aCopy attributes from object 'b' to object 'a', with options to include/exclude certain attributes. Args: a (Any): Destination object to copy attributes to. b (Any): Source object to copy attributes from. include (tuple, optional): Attributes to include. If empty, all attributes are included. exclude (tuple, optional): Attributes to exclude. _N)__dict__rr|rksetattr)abincludeexcludekvs r" copy_attrrSsR   "1 LQg-!,,s2CqG|  Aq!  r.c|jDcic]?\}|vs tfd|Ds |j|jk(s=|Ac}}Scc}}w)aLReturn a dictionary of intersecting keys with matching shapes, excluding 'exclude' keys, using da values. Args: da (dict): First dictionary. db (dict): Second dictionary. exclude (tuple, optional): Keys to exclude. Returns: (dict): Dictionary of intersecting keys with matching shapes. c3&K|]}|v ywr_r/)r`rrQs r"rbz"intersect_dicts..-s:W!1A::Ws)rallr)dadbrPrQrRs ` r"intersect_dictsrY"sX XXZ s sTQ17s:Ww:W7W\]\c\cgijkglgrgr\rAqD ss ss AAAAc~t|tjjtjjfS)zReturn True if model is of type DP or DDP. Args: model (nn.Module): Model to check. Returns: (bool): True if model is DataParallel or DistributedDataParallel. )rirparallel DataParallelDistributedDataParallelrs r" is_parallelr^0s* ebkk66 8[8[\ ]]r.c t|dr1t|jtjr |j}n?t|dr1t|j tjr |j }n|S})aVUnwrap compiled and parallel models to get the base model. Args: m (nn.Module): A model that may be wrapped by torch.compile (._orig_mod) or parallel wrappers such as DataParallel/DistributedDataParallel (.module). Returns: (nn.Module): The unwrapped base model without compile or parallel wrappers. _orig_modmodule)rqrir`rModulera)rs r"rr<sY  1k "z!++ryy'I A Q !j299&EAH r.cfdS)aMReturn a lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf. Args: y1 (float, optional): Initial value. y2 (float, optional): Final value. steps (int, optional): Number of steps. Returns: (function): Lambda function for computing the sinusoidal ramp. ctdtj|tjzz z dz dz zzS)NrYrr)rr=cospi)rstepsy1y2s r"rzone_cycle..Zs=S!dhhq477{U':;;q@!DRPSUUr.r/)rhrirgs```r" one_cyclerjOs  VUr.c.tj|tjj|tj|tj j |tj j ||rtrftjdddtjj_ dtjd<t|tjd<ytj dyt#y)zInitialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html. Args: seed (int, optional): Random seed. deterministic (bool, optional): Whether to set deterministic algorithms. T) warn_onlyz:4096:8CUBLAS_WORKSPACE_CONFIGPYTHONHASHSEEDz3Upgrade to torch>=2.0.0 for deterministic training.N)randomseednpr' manual_seedr6manual_seed_allruse_deterministic_algorithmsrcudnn deterministicrrrjr warningunset_deterministic)rprvs r" init_seedsry]s KKIINN4 d JJ4  JJt$   . .tt D15ENN .4=BJJ0 1+.t9BJJ' ( NNP Qr.ctjddtjj_t j jddt j jddy)z@Unset all the configurations applied for deterministic training.FrmNrn)r'rtrrurvrrr}r/r.r"rxrxvsH &&u-).ENN&JJNN,d3JJNN#T*r.c&eZdZdZddZdZddZy)ModelEMAaUpdated Exponential Moving Average (EMA) implementation. Keeps a moving average of everything in the model state_dict (parameters and buffers). For EMA details see References. To disable EMA set the `enabled` attribute to `False`. Attributes: ema (nn.Module): Copy of the model in evaluation mode. updates (int): Number of EMA updates. decay (function): Decay function that determines the EMA weight. enabled (bool): Whether EMA is enabled. References: - https://github.com/rwightman/pytorch-image-models - https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage ctt|j|_||_fd|_|jj D]}|jdd|_y)a7Initialize EMA for 'model' with given arguments. Args: model (nn.Module): Model to create EMA for. decay (float, optional): Maximum EMA decay rate. tau (int, optional): EMA decay time constant. updates (int, optional): Initial number of updates. cBdtj| z z zS)NrY)r=exp)rdecaytaus r"rz#ModelEMA.__init__..s uDHHaR#X,>(>?r.FTN) rrevalemaupdatesrrrr2)selfrrrrras `` r"__init__zModelEMA.__init__s^L/0557 ? $$& $A  U # $ r.c|jr|xjdz c_|j|j}t|j }|j j j D]=\}}|jjs||z}|d|z ||jzz }?yy)zgUpdate EMA parameters. Args: model (nn.Module): Model to update EMA from. rYN) r2rrr state_dictrrris_floating_pointdetach)rrrmsdrQrRs r"updatezModelEMA.updates << LLA L 4<<(Au%002C++-335 3177,,FA!a%3q6==?22A 3 r.cN|jrt|j|||yy)a(Copy attributes from model to EMA, with options to include/exclude certain attributes. Args: model (nn.Module): Model to copy attributes from. include (tuple, optional): Attributes to include. exclude (tuple, optional): Attributes to exclude. N)r2rSr)rrrOrPs r" update_attrzModelEMA.update_attrs" << dhhw 8 r.N)gH.?ir)r/) process_groupreducer)r __module__ __qualname____doc__rrrr/r.r"r|r|~s$ 3" 9r.r|c  t|tjd}t|tsJdd|vsJd tjjtd d d }|jd r|d |d<t|dd r t |dj|d_t|ddr d|d_|dj!|dj#D] }d|_ it&|jdi}dD]}d||< d|d<|j)D cic]\}} |t*vs|| c} }|d<i|||xsi} tj,| |xs|t.j0j3|xs|dz } t j4d|d|rd|dndd| dd| S#t $r'}t jd|d|icYd}~Sd}~wwxYwcc} }w)aStrip optimizer from 'f' to finalize training, optionally save as 's'. Args: f (str | Path): File path to model to strip the optimizer from. s (str, optional): File path to save the model with stripped optimizer to. If not provided, 'f' will be overwritten. updates (dict, optional): A dictionary of updates to overlay onto the checkpoint before saving. Returns: (dict): The combined checkpoint dictionary. Examples: >>> from pathlib import Path >>> from ultralytics.utils.torch_utils import strip_optimizer >>> for f in Path("path/to/model/checkpoints").rglob("*.pt"): ... strip_optimizer(f) rc) map_locationz%checkpoint is not a Python dictionaryrz'model' missing from checkpointz Skipping z!, not a valid Ultralytics model: Nz2AGPL-3.0 License (https://ultralytics.com/license)zhttps://docs.ultralytics.com)dateversionlicensedocsrr criterionF train_args) optimizer best_fitnessrrscalerrepochg.AzOptimizer stripped from r[z saved as rUrNrMB)rr'r7ridictr?r rwrnow isoformatr r@rqrrhalfrrr rr saverpathgetsizerv) frrremetadatararrQrRcombinedmbs r"strip_optimizerrs$ qu||E': ;!T"K$KK"!|>>>|  ((*G. H uuU|uX' qz6"qz/' qz;'#' gJOO wZ " " $   ; :!%% b"9 :D D!AgJ(, N1=M8Mq!tNAlO4(3a3GMr3H JJxa a 3 &B KK*1#QAA3a/@2.NaPRSVxWYZ[ OG 1#%FqcJK 6Os*AGHH HG?9H?Hc|djD]j}|jD]U\}}|dvs t|tjs&|j tj usC|j||<Wl|S)aConvert the state_dict of a given optimizer to FP16, focusing on the 'state' key for tensor conversions. Args: state_dict (dict): Optimizer state dictionary. Returns: (dict): Converted optimizer state dictionary with FP16 tensors. state>step exp_avg_sq)valuesrrir'Tensorrfloat32r)rrrQrRs r"$convert_optimizer_state_dict_to_fp16rs|G$++-$KKM $DAq..:a3NSTSZSZ^c^k^kSk668a $$ r.c#BKtd}tjjrFtjj  |tjj ||d<y|y#tjj ||d<wxYww)aIMonitor and manage CUDA memory usage. This function checks if CUDA is available and, if so, empties the CUDA cache to free up unused memory. It then yields a dictionary containing memory usage information, which can be updated by the caller. Finally, it updates the dictionary with the amount of memory reserved by CUDA on the specified device. Args: device (torch.device, optional): The CUDA device to query memory usage for. Yields: (dict): A dictionary with a key 'memory' initialized to 0, which will be updated with the reserved memory. r)memoryrN)rr'r6r empty_cachememory_reserved)r7 cuda_infos r"cuda_memory_usager svAI zz    EO"'**"<">> from ultralytics.utils.torch_utils import profile_ops >>> input = torch.randn(16, 3, 640, 640) >>> m1 = lambda x: x * torch.sigmoid(x) >>> m2 = nn.SiLU() >>> profile_ops(input, [m1, m2], n=100) # profile over 100 iterations rNParamsz>12sGFLOPsz GPU_mem (GB)z>14sz forward (ms)z backward (ms)inputz>24soutputTtor)rrrFrrrrYc3<K|]}|jywr_)r)r`yis r"rbzprofile_ops..Zs 6b 6rnanric3hK|])}jd|z jd|z z+yw)rN)r)r`rrs r"rbzprofile_ops..gs/'iRSqQWWR[1_(M'is/2)r7rc3~K|]5}t|tjrt|jnd7yw)rN)rir'rtuplerrs r"rbzprofile_ops..ls,i]^Au||1LuQWW~RXXis;=c3<K|]}|jywr_rrs r"rbzprofile_ops..ms:a :r12z12.4gz>14.3fz14.4g)'rrorir'r7rr rvgccollectr6rrrrrqrrfloat16rrrr?r{rrrbackwardfloatrandnrrsrtolistrrrbrappend)ropsrr7 max_num_objrrrtftbr4rmemrJrys_ins_outrarrs @r" profile_opsr$s(G fell +v& KK D/(4(=nT=RSbcgRh 4.$ )JJL JJ -UE7/) DDL"3-C5, )A '4 0V aA#Av.:a3NSTSZSZ^c^k^kSkqrA1iAB ]a Xa[!e LQORUUXYYgh# )q9A*62 0i({!aD({!0:DQ:MS 6A 66STYY[ddf#,;AaD 09X.44C1Q4!A$;$.22B1Q4!A$;$.22B".v6)!KK ! + #C'iWXW_W_WfWfWh'i$i j'-&+mm y2S88/90jcdfgbhi e>HBII>VC:1<<>::\] qfU5M#fbZ5zRVY]Q^_dgk^lmn5#r2tUCD   &&(Y, )/)` Ny *   )0#(>> device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") >>> # Try to compile and warm up a model with a 640x640 input >>> model = attempt_compile(model, device=device, imgsz=640, use_autocast=True, warmup=True) Notes: - If the current PyTorch build does not provide torch.compile, the function returns the input model immediately. - Warmup runs under torch.inference_mode and may use torch.autocast for CUDA/MPS to align compute precision. - CUDA devices are synchronized after warmup to account for asynchronous kernel execution. compileTdefaultzcompile:z starting torch.compile with 'z ' mode...z max-autotunez mode='zB' not recommended, using mode='max-autotune-no-cudagraphs' insteadzmax-autotune-no-cudagraphsinductor)r:backendz. torch.compile failed, continuing uncompiled: NrrYrWrr6>rer6z complete in rz s (compile z s + warmup zs)z compile complete in z s (no warmup))rqr'rr rvrwr perf_counterrr?rrrr*r5r6r)rr7r use_autocastwarmupr:rt0r t_compilet_warmdummyt1rJtotals r"attempt_compilers#H 5) $D  t| j !F KK6(8iHI ~&.pqr+   B e$ C!!#b(I F  Aq%v> FKK61JJLE      ! ! # !  >^^FKK0%e A%%%L  ! ;;& JJ " "6 *""$r)  E  vhmE#;k)CP[\bcf[ggijk L  vh3Ic?-PQ L7 &!OPQsST %% ! !sB8G00H.> H"H.0 H9HHH"H+ 'H..H7)rrs)r6)r2boolr7rj)rUFT)FT)r)r7Fr)r/r/)r/)r nn.Modulereturnr)rr7d)rF)zbest.ptrUN)rz str | Pathrrjrzdict[str, Any] | Nonerzdict[str, Any]r_) Nr)rFFr)rtorch.nn.Moduler7z torch.devicerrsrrrrr:z bool | strrr)^ __future__r functoolsrr=rror contextlibrcopyrrpathlibrtypingrnumpyrqr'torch.distributed distributedrtorch.nnrtorch.nn.functional functionalrB ultralyticsr r=r r r r rrrrrultralytics.utils.checksrultralytics.utils.cpurultralytics.utils.patchesrr&r) TORCH_1_11r3r TORCH_2_1 TORCH_2_3 TORCH_2_4 TORCH_2_8 TORCH_2_9 TORCH_2_10TORCHVISION_0_10TORCHVISION_0_11TORCHVISION_0_13TORCHVISION_0_18rwr#r0r5 lru_cacherArIrrrrrrrrrr(r5rHrSrYr^rrjryrxr|rrrrrrr/r.r"rs# % #   3)0 - 1 =( 3 =( 3 =( 3 - 1 - 1 - 1 - 1 - 1 - 1 =( 3  !4h? !4h? !4h? !4h? }]I6FNN B  M M 08  7 7 NN zz&B(B* Z6 I B%PB ?,  t ^& V2+=9=9@9x"0Rj00l L L L L L  L  LLr.