FPh_&ddlZddlZddlZddlZddlZddlmZddlmZddl m Z ddl m Z ddl ZddlZddlmZddlmZddlmcmZddlmZmZmZmZddlmZ ddlZeej8dZ!eej8d Z"ed e#fd Z$d Z%d Z&d+dZ'dZ(dZ)dZ*d,dZ+dZ,dZ-dZ.d-dZ/d-dZ0dZ1d.dZ2dZ3d/dZ4dZ5d0dZ6dZ7dZ8d1d Z9d2d!Z:Gd"d#Z;d3d$e eGd)d*Z?y#e $rdZYwxYw)5N)contextmanager)deepcopy)Path)Union)DEFAULT_CFG_DICTDEFAULT_CFG_KEYSLOGGER __version__) check_versionz1.9.0z2.0.0 local_rankc# Ktjjxrtjj}|r|dvrt j |gd|r|dk(rt j dgyyyw)zcDecorator to make all processes in distributed training wait for each local_master to do something.)r) device_idsNr)torch distributed is_availableis_initializeddistbarrier)r initializeds hC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\ultralytics/utils/torch_utils.pytorch_distributed_zero_firstrsi##002Yu7H7H7W7W7YKz0  - zQ $'{sBBc d}|S)zXApplies torch.inference_mode() decorator if torch>=1.9.0 else torch.no_grad() decorator.ctrtjr|Strtj|Stj|S)zNApplies appropriate torch decorator for inference mode based on torch version.) TORCH_1_9ris_inference_mode_enabledinference_modeno_grad)fns rdecoratez&smart_inference_mode..decorate-s> 88:IKIE((KBO O5==KBO O)r s rsmart_inference_moder#*sP Or!cddl}d}|j}|j|d|vr|dn|d|vr|dn|dd}|jddjd djd dS) z=Return a string with system CPU information, i.e. 'Apple M2'.rN) brand_raw hardware_rawarch_string_rawunknownz(R)zCPU z@ )cpuinfo get_cpu_infogetreplace)r,kinfostrings rr-r-7sx6A    !D XXaddlad! !!A$PY ZF >>% $ , ,VR 8 @ @r JJr!c t|tjr|Sdtdt j dtjd}t |j}dD]}|j|d}|dk(}|dv}|s|rd tjd <n!|r|d k(rd }tjjd d }|tjd <tjjr:tjjt|jddk\st!j"|tjjdk(rdnd} t%d|dtjjdtjjd|d| |s|stjjr|r|j'dnd } t| } | dkDr7|dkDr2|| zdk7r*t%d|d| d|| z| zd|| z| z| zd| d dt|dzz} t)| D]T\} }tjj+| }|| dk(rdn| d|d|j,d|j.d z d!d"z }Vd#}nW|rBt0r>> select_device('cuda:0') device(type='cuda', index=0) >>> select_device('cpu') device(type='cpu') Note: Sets the 'CUDA_VISIBLE_DEVICES' environment variable for specifying which GPUs to use. zUltralytics YOLOvu 🚀 Python-z torch- )zcuda:none()[]'r4r+cpu)mpszmps:0z-1CUDA_VISIBLE_DEVICEScuda0N,rz}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']:  r(z'batch=z"' must be a multiple of GPU count z . Try 'batch=z ' or 'batch=z/', the nearest batch sizes evenly divisible by .zCUDA:z (, iz.0fzMiB) zcuda:0zMPS (z) r<zCPU () isinstancerdevicer platformpython_versionstrlowerr/osenvironr.r>r device_countlenr r1 ValueErrorsplit enumerateget_device_propertiesname total_memory TORCH_2_0backendsr<r-rstrip)rEbatchnewlineverbosesremover;r<visibleinstalldevicesnspaceidpargs r select_devicereAs&D&%,,' K= h6M6M6O5PPWX]XiXiWjjklA [   F?+@ E/C $ $C c-1 )*  V F**..!7>-3 )* '')ejj.E.E.G3v~~^aceOfKg.g KKN=BZZ=T=T=VZ[=[:ac 4VH==>CZZ=T=T=Vwug-OPQsR_`eij`jmn`n_op'',zA~'9&::ijkillmop ps1vz"g&DAq 003A !q&Be,E!BqvvhbSZA[\_@``fg gA' u~~11>>@ u\^$C (( u\^$C (( Aahhj1 << r!ctjjrtjjt jS)zPyTorch-accurate time.)rr>r synchronizetimer"r!r time_syncris. zz   99;r!c tj|j|j|j|j |j |j|jdjdj|jj}|jjj|jd}tj |jj#tj$|j&|j(z}|jj+tj,||j|jj.|j0Dtj2|jj5d|jjn |j0}|j0|jj7|j8j#tj$|j(|j&zz }|j0j+tj,||j;ddj;d|z|S)z\Fuse Conv2d() and BatchNorm2d() layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/.T) kernel_sizestridepaddingdilationgroupsbiasFrrrEr()nnConv2d in_channels out_channelsrkrlrmrnrorequires_grad_toweightrEcloneviewrdiagdivsqrteps running_varcopy_mmshaperpzerossizemul running_meanreshape)convbn fusedconvw_convw_bnb_convb_bns rfuse_conv_and_bnrs $**++&*&6&6!%"&,,#'==!%#%&4^E%:22dkk>P>P;Q[[   % %d&7&7 >TVTZTZCZ8[\ \D NN$r1(=>FFrJTQR r!c 2tj|j|j|j|j |j |j|j|jd jdj|jj}|jjj|jd}t!j"|jj%t!j&|j(|j*z}|jj-t!j.||j|jj0|j2Dt!j4|jj7d|jjn |j2}|j2|jj9|j:j%t!j&|j*|j(zz }|j2j-t!j.||j=ddj=d|z|S)z0Fuse ConvTranspose2d() and BatchNorm2d() layers.T)rkrlrmoutput_paddingrnrorpFrr(rq)rrConvTranspose2drtrurkrlrmrrnrorvrwrxrEryrzrr{r|r}r~rrrrrprrrrr)deconvr fuseddconvw_deconvrrrs rfuse_deconv_and_bnrs##F$6$6$*$7$7060B0B+1==,2NN393H3H-3__+1==)-/0>~e/DRR H\H\E]}}""$))&*=*=rBH ::biimmEJJrvv/F$GH IDEHHT8499*:K:K:Q:QRSQWP[P[PcU[[++A.v}}7K7K LioititF 77RYY]]2??377 2>>TVTZTZCZ8[\ \DOO%((4A)>?GGKdRS r!c|syt|}t|}tt|j }|rt j dddddddddd d dd d dd dddd t|jD]\}\}} |jdd}t j d||| j| jt| j| j| j| jfzt!||} t#|ddrdnd} | rd| ddnd} t#|ddxst#|dij%dd} t'| j(jddxsd}t j |d| d|d |d!|d"| |||| fS)#z_ Model information. imgsz may be int or list, i.e. imgsz=640 or imgsz=[640, 320]. Nlayerz>5r4rRz>40gradientz>9 parametersz>12rz>20muz>10sigmaz module_list.r+z)%5g %40s %9s %12g %20s %10.3g %10.3g %10sis_fusedcy)NFr"r"r!rzmodel_info..sUr!z (fused)rC.1fz GFLOPs yaml_fileyamlyoloYOLOModelz summaryz: z layers, z parameters, z gradients)get_num_paramsget_num_gradientsrMlistmodulesr r1rPnamed_parametersr/ requires_gradnumelrmeanstddtype get_flopsgetattrr.rstem)modeldetailedrYimgszn_pn_gn_lrarRrcflopsfusedfsr model_names r model_infors   C E "C d5==?# $C rl!F3?LAya<<3D KKCD!//1779d177mQVVXWXW\W\W^`a`g`ghi j@ eU #EC'%]CEJ2E$)2eC[ rB{B/b75&"3M3Q3QR]_a3bIi%%--ff=HJ KK:,hugRuIcU-PSuT^_a^bcd S% r!cBtd|jDS)z6Return the total number of parameters in a YOLO model.c3<K|]}|jywNr.0xs r z!get_num_params..s5"4Qqwwy"4sumrrs rrrs 5%"2"2"45 55r!cBtd|jDS)zEReturn the total number of parameters with gradients in a YOLO model.c3VK|]!}|js|j#ywr)rrrs rrz$get_num_gradients..sH"4Qqwwy"4s))rrs rrrs H%"2"2"4H HHr!cx|jjrHddlm}||jg|j jd}|j dn5t|jtt|jdd}t|jjdd|d<|S) a Return model info dict with useful model information. Example: YOLOv8n info for loggers ```python 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) ProfileModelsrqz model/name)zmodel/parametersz model/GFLOPs inferencezmodel/speed_PyTorch(ms)) argsprofileultralytics.utils.benchmarksrlastrEpoprrroundr validatorspeed)trainerrresultss rmodel_info_for_loggersrs||>w~~FNNPQRS L!!/w}} =!)GMM":A>@*/w/@/@/F/F{/SUV)WG %& Nr!c t|}t|j}t|dr-t t |j j dnd}tjd|jd||f|j}tr+tjt||gdddz d znd}t|tr|n||g}||dz|z |dz|z S#t $rYywxYw) zReturn a YOLO model's FLOPs.rl r(rqFinputsrYreAr)) de_parallelnextrhasattrmaxintrlremptyrrEthoprrrDr Exception)rrrcrlimrs rrrs E" !!# $5.&s9%8AGG%8srr)rTrrrrrrrlrrrrEprofilerrr key_averagesrDr)rrrcrlrprofrs rget_flops_with_torch_profilerrsE" !!# $6=eX6N#c%,,**,-r2TVZ[[ [[!QWWQZ8 J ^^ # #t # 4 "I59T%6%6%899C?#E40uena 6)E!H4v= 5 4s < DD&cZ|jD]}t|}|tjur!|tjurd|_d|_B|tjtjtjtjtjfvsd|_ y)z*Initialize model weights to random values.gMbP?gQ?TN) rtyperrrs BatchNorm2dr~momentum Hardswish LeakyReLUReLUReLU6SiLUinplace)rmts rinitialize_weightsr-sq ]]_ G  >  ".. AEAJ 2<<rww"''J JAIr!c dk(r|S|jdd\}}t|zt|zf}tj||dd}|sfd||fD\}}tj|d||d z d||dz gd S) zScales and pads an image tensor of shape img(bs,3,y,x) based on given ratio and grid size gs, optionally retaining the original shape. ?r)NbilinearF)rmode align_cornersc3ZK|]"}tj|zz z$ywr)mathceil)rrgsratios rrzscale_img..Ds(?1 !e)b.)B.s(+rr(gS㥛?)value)rrF interpolatepad)imgr same_shaperhwrZs ` ` r scale_imgr:s | 99QR=DAq QYQY(A --!*E JC ?A?1 55q!ad(Aq1Q4x0 >>r!ct|tjrt|j }t j ||z |zS)z'Returns nearest x divisible by divisor.)rDrTensorrrrr)rdivisors rmake_divisiblerHs8'5<<(gkkm$ 99Q[ !G ++r!c|jjD]7\}}t|r||vs|jds||vr+t |||9y)zdCopies attributes from object 'b' to object 'a', with options to include/exclude certain attributes._N)__dict__itemsrM startswithsetattr)abincludeexcluder0vs r copy_attrrOsK   "1 LQg-!,,s2CqG|  Aq!  #r!cZtdttjDdz S)zYReturn second-most (for maturity) recently supported ONNX opset by this version of torch.c3BK|]}d|vst|ddyw)symbolic_opsetN)r)rr0s rrz#get_latest_opset..Zs%N$4q8HA8Ms1RS6{$4s r()rvarsronnxr"r!rget_latest_opsetr&Xs" ND$4N NQR RRr!c|jDcic]?\}|vs tfd|Ds |j|jk(s=|Ac}}Scc}}w)zjReturns a dictionary of intersecting keys with matching shapes, excluding 'exclude' keys, using da values.c3&K|]}|v ywrr")rrr0s rrz"intersect_dicts.._s:Ww!1A:ws)rallr)dadbrr0rs ` rintersect_dictsr,]sWXXZ sZTQ17s:Ww:W7W\]\c\cgijkglgrgr\rAqDZ ss ss AAAAc~t|tjjtjjfS)z+Returns True if model is of type DP or DDP.)rDrrparallel DataParallelDistributedDataParallelrs r is_parallelr1bs( ebkk66 8[8[\ ]]r!c4t|r |jS|S)zODe-parallelize a model: returns single-GPU model if model is of type DP or DDP.)r1modulers rrrgs&u-5<<858r!cfdS)zaReturns a lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf.ctdtj|tjzz z dz z zzS)Nr(r))rcospi)rstepsy1y2s rrzone_cycle..ns3q488AK%$788A="r'JROr!r")r9r:r8s```r one_cycler;ls OOr!ctj|tjj|tj|tj j |tj j ||rtrftjdddtjj_ dtjd<t|tjd<ytj dytjddtjj_ y) zeInitialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html.T) warn_onlyz:4096:8CUBLAS_WORKSPACE_CONFIGPYTHONHASHSEEDuBWARNING ⚠️ Upgrade to torch>=2.0.0 for deterministic training.FN)randomseednpr manual_seedr>manual_seed_allrTuse_deterministic_algorithmsrUcudnn deterministicrJrKrHr warning)rArGs r init_seedsrIqs KKIINN4 d JJ4  JJt$   . .tt D15ENN .4=BJJ0 1+.t9BJJ' ( NN_ ` **51-2*r!c&eZdZdZddZdZddZy)ModelEMAa^Updated Exponential Moving Average (EMA) from https://github.com/rwightman/pytorch-image-models Keeps a moving average of everything in the model state_dict (parameters and buffers) For EMA details see https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage To disable EMA set the `enabled` attribute to `False`. ctt|j|_||_fd|_|jj D]}|jdd|_y)z Create EMA.cBdtj| z z zS)Nr()rexp)rdecaytaus rrz#ModelEMA.__init__..suDHHaR#X,>(>?r!FTN) rrevalemaupdatesrOrrvenabled)selfrrOrPrSrcs `` r__init__zModelEMA.__init__sWK./446 ? $$&A  U #' r!c|jr|xjdz c_|j|j}t|j }|j j j D]=\}}|jjs||z}|d|z ||jzz }?yy)zUpdate EMA parameters.r(N) rTrSrOr state_dictrRrris_floating_pointdetach)rUrrbmsdr0rs rupdatezModelEMA.updates << LLA L 4<<(Ae$//1C++-335177,,FA!a%3q6==?22A6 r!cN|jrt|j|||yy)zCUpdates attributes and saves stripped model with optimizer removed.N)rTrrR)rUrrrs r update_attrzModelEMA.update_attrs << dhhw 8 r!N)gH.?ir)r") process_groupreducer)__name__ __module__ __qualname____doc__rVr\r^r"r!rrKrKs  39r!rKfrZreturnc  tj|tjd}d|vrtjd|dyt |ddr t |dj|d_d|vr it|dnd}|jd r|d |d<d D]}d||< d |d <|dj|djD] }d |_ |jDcic]\}}|tvs||c}}|d<tj||xs|t j"j%|xs|dz }tjd|d|rd|dndd|ddycc}}w)a' Strip optimizer from 'f' to finalize training, optionally save as 's'. Args: f (str): file path to model to strip the optimizer from. Default is 'best.pt'. s (str): file path to save the model with stripped optimizer to. If not provided, 'f' will be overwritten. Returns: None Example: ```python from pathlib import Path from ultralytics.utils.torch_utils import strip_optimizer for f in Path('path/to/weights').rglob('*.pt'): strip_optimizer(f) ``` r;) map_locationrz Skipping z , not a valid Ultralytics model.Nr train_argsrR) optimizer best_fitnessrRrSrepochFg.AzOptimizer stripped from r@z saved as r+r4rMB)rloadrEr r1rdictrrr.halfrrrrsaverJpathgetsize)rerZrrr0rcrmbs rstrip_optimizerrusx(  15<<#67Aa is"BCDqz6"qz/' 6Ba6G 2 2!L/ 2TDuuU|uX' :!;AgJgJOO wZ " " $%(, N 1=M8Mq!t NAlO JJq!&q a 3 &B KK*1#QAA3a/@2.NaPRSVxWYZ[ Os :F Fc g}t|tjs t|}t j dddddddddddd d d t|t r|n|gD]}|j|}d |_t|t r|n|gD]O}t|d r|j|n|}t|d rFt|tjr,|jtjur|jn|}ddgd} }} tr"tj||gdddz dznd} t#|D]} t%| d<||} t%| d< t| t rt'd| Dn| j'j)t%| d<|| d| dz dz|z z }|| d| dz dz|z z }tj,j/r!tj,j1dz nd} d| fD\}}t|t2j4r t'd|j7Dnd}t j |d| d| d|d|dt9|d t9|d |j;|| | ||||gtj,j=R|S#t $rd} YwxYw#t $rt+d| d<Y[wxYw#t $r0}t j ||j;dYd}~d}~wwxYw)aU Ultralytics speed, memory and FLOPs profiler. Example: ```python from ultralytics.utils.torch_utils import profile input = torch.randn(16, 3, 640, 640) m1 = lambda x: x * torch.sigmoid(x) m2 = nn.SiLU() profile(input, [m1, m2], n=100) # profile over 100 iterations ``` Paramsz>12sGFLOPsz GPU_mem (GB)z>14sz forward (ms)z backward (ms)inputz>24soutputTrwrpr)rrrFrrr)r(c3<K|]}|jywr)r)ryis rrzprofile..s2"RVVXrnanic3~K|]5}t|tjrt|jnd7yw)rN)rDrrtuplerrs rrzprofile..s/ibh]^Au||1LuQWW~RXXbhs;=c3<K|]}|jywrrrs rrzprofile..s:>a >r12z12.4gz>14.3fz14.4gN)rDrrErer r1rrwrrrrfloat16rprrrrangerirbackwardfloatr>rmemory_reservedrrModulerrHappend empty_cache)ryopsr_rErrrtftbrrrymems_ins_outrces rrrs*G fell +v& KK8D/(40EnUYEZ[jkoZp4.$12!-UE7 : DDL"3-C58A '4 0V aA#Av.:a3NSTSZSZ^c^k^kSkqrA1iAB SW QsEB1EKaO]^ %qA$;AaD!A$;AaD,6@D6I222qUUW``b({!1Q4!A$;$.22B1Q4!A$;$.22B"=BJJHBII>VC:1<<>::\] qfU5M#fbZ5zRUVZR[\`Qabefkblmqarst5#r2tUCD JJ " " $=9;D N3  %,$U|!, % At$$ %sP*K00L 4A LD L 0 K?>K?L L L L  M)&MMceZdZdZddZdZy) EarlyStoppingzkEarly stopping class that stops training when a specified number of epochs have passed without improvement.cVd|_d|_|xs td|_d|_y)z Initialize early stopping object. Args: patience (int, optional): Number of epochs to wait after fitness stops improving before stopping. rinfFN)rk best_epochrpatience possible_stop)rUrs rrVzEarlyStopping.__init__s+  0E%L "r!c .|y||jk\r||_||_||jz }||jdz k\|_||jk\}|r=t j d|jd|jd|jd|S)a Check whether to stop training. Args: epoch (int): Current epoch of training fitness (float): Fitness value of current epoch Returns: (bool): True if training should stop, False otherwise Fr(z;Stopping training early as no improvement observed in last z( epochs. Best results observed at epoch z@, best model saved as best.pt. To update EarlyStopping(patience=z^) pass a new patience value, i.e. `patience=300` or use `patience=0` to disable EarlyStopping.)rkrrrr r1)rUrlfitnessdeltastops r__call__zEarlyStopping.__call__s ? d'' '#DO 'D '"t}}q'89 %  KKUVZVcVcUde::>//9JK<<@MM?K\] ^ r!N)2)rarbrcrdrVrr"r!rrr su #r!r)r+rFT)FT)r)rFr)r"r")r")rrd)rF)zbest.ptr+) N)@rrJrFr@rh contextlibrcopyrpathlibrtypingrnumpyrBrtorch.distributedrrtorch.nnrrtorch.nn.functional functionalrultralytics.utilsrrr r ultralytics.utils.checksr r ImportErrorrrTrrr#r-rerirrrrrrrrrrrrr&r,r1rr;rIrKrHrurrr"r!rrsV  % UU2 %++W 5 %++W 5 %S%% KSl02 86 I 4    ?,S t ^ 9 P 3* 9 9F(\uS$Y'(\(\T(\V6r((k DsDDD