j/dZddlmZddlZddlmZddlZddlZddl m Z m Z m Z ddl mZmZ dddZdde jd dfddZdS)zlFunctions for estimating the best YOLO batch size to use a fraction of the available CUDA memory in PyTorch.) annotationsN)deepcopy) DEFAULT_CFGLOGGERcolorstr)autocast profile_opsTmodeltorch.nn.Moduleimgszintampboolbatch int | float max_num_obj dataset_sizereturnct|5tt||d|cxkrdkrnn|nd||cdddS#1swxYwYdS)a)Compute optimal YOLO training batch size using the autobatch() function. Args: model (torch.nn.Module): YOLO model to check batch size for. imgsz (int, optional): Image size used for training. amp (bool, optional): Use automatic mixed precision if True. batch (int | float, optional): Fraction of GPU memory to use. If -1, use default. max_num_obj (int, optional): The maximum number of objects from dataset. dataset_size (int, optional): Total number of training images. If > 0, batch size will not exceed this value. Returns: (int): Optimal batch size computed using the autobatch() function. Notes: If 0.0 < batch < 1.0, it's used as the fraction of GPU memory to use. Otherwise, a default fraction of 0.6 is used. )enabledgg?333333?)fractionrrN)r autobatchrtrain)r rrrrrs ^/var/www/html/Carbon-Document/venv/lib/python3.11/site-packages/ultralytics/utils/autobatch.pycheck_train_batch_sizers2 #      UOO ! ! # # !E////C/////UUS#%                      sAA##A'*A'rrfloat batch_sizectd}tj|dd|dzdt|j}|jdvrtj|d||Stj j j rtj|d||Sd }d tj d d d } tj|} | j|z tj||z } tj||z } | | zz } tj|| d| jddd| dd| dd| dd dkrgdngd}d krfd|D}|jdd fd|D}t/||d||fdt1t3|D}|r t3|nggf\}}t5j||d}t9t;| |z|dz |d z }d vr:d }|||kr|t?|dz d }|dks|d!krtj|d"|d#|d$|}d krtA|}t5j!||| z| zz }tj|d%|d&| d'|zdd(dd)|dzd*d+ |tj"S#tF$rG}tj|d,|d-|d$|cYd }~tj"Sd }~wwxYw#tj"wxYw).aAutomatically estimate the best YOLO batch size to use a fraction of the available CUDA memory. Args: model (torch.nn.Module): YOLO model to compute batch size for. imgsz (int, optional): The image size used as input for the YOLO model. fraction (float, optional): The fraction of available CUDA memory to use. batch_size (int, optional): The default batch size to use if an error is detected. max_num_obj (int, optional): The maximum number of objects from dataset. dataset_size (int, optional): Total number of training images. If > 0, batch size will not exceed this value. Returns: (int): The optimal batch size. z AutoBatch: z'Computing optimal batch size for imgsz=z at dz% CUDA memory utilization.>cpumpsz4intended for CUDA devices, using default batch-size zHRequires torch.backends.cudnn.benchmark=False, using default batch-size i@zCUDA:CUDA_VISIBLE_DEVICES0rz (z) z.2fz G total, z G reserved, z G allocated, zG free)r r()r r)r*r+r( @c g|] }|k| Sr/).0brs r zautobatch..`s#CCCQl1B1Bq1B1B1Bchannelsc>g|]}tj|Sr/)torchempty)r0r1chrs rr2zautobatch..cs)EEEAu{1b%//EEEr3r )ndevicercg|]x\}\}}|rnt|dttfrLd|dcxkrkr9nn6|dks&|dz r|d|dz dkn||dgyS)r)rr ) isinstancerr )r0ixyresultsts rr2zautobatch..gs   6Aq 1Q4#u..  AaD 1 awq1u~1Aq8I1I1I !I 2J1I1Ir3)degNizbatch=z. outside safe range, using default batch-size .zUsing batch-size z for  zG/zG (z.0fu%) ✅zerror detected: z, using default batch-size )$rrinfonext parametersr;typewarningr7backendscudnn benchmarkosgetenvstripcudaget_device_properties total_memorymemory_reservedmemory_allocatednameyamlgetr enumeratezipnppolyfitrroundindexmaxminpolyval empty_cache Exception)r rrr!rrprefixr;gbd propertiesraf batch_sizesimgxyfit_xfit_ypr1r>er9rArBs ` ` @@@rrr3s,m $ $F K6ww%wwXX[^wwwxxx %""$$ % % ,F {n$$&bbV`bbccc ~%&vvjtvvwww BC 0#66<<>>qACCA11&99J"$A ""6**R/A ##F++b0A QU A K6w1ww ww1www1wwwVWwwwjkwwwwxxx'("ff""""2J2J2JKaCCCC+CCC  A & &B!!EEEEEEEEc5Af+VVV     &s;'@'@AA   $&3sBxxB8 u Jue + + + q8|$$qt+qt3 4 4 7?? d##AKN""AE1 . q55AHH NfjjAjj]gjjj k k kA !  A|$$AJq!$$q(1,1 vttttttALtttPQtttZbehZhttttuuu      &^^!^^Q[^^^___           s1E4M N*#N%N*N-%N**N-- O )r Tr r r)r rrrrrrrrrrrrr)r rrrrr r!rrrrrrr)__doc__ __future__rrNcopyrnumpyr[r7ultralytics.utilsrrrultralytics.utils.torch_utilsrr rrrr/r3rrxsrr""""""  ;;;;;;;;;;????????       J!' P!P!P!P!P!P!P!r3