^j$~dZddlmZddlZddlZddlmZddlmZm Z ddl Z ddl m Z m Z mZddlmZGdd e Zy) zAExponential Moving Average callback compatible with ``ModelEma``.) annotationsN)deepcopy)AnyOptional)CallbackLightningModuleTrainer) AveragedModelceZdZdZ d dfd Z ddZddZ d ddZddZ ddZ ddZ dd Z dd Z dd Z dd Zdd ZddZxZS)RFDETREMACallbacka/Exponential Moving Average with optional tau-based warm-up. Drop-in replacement for ``rfdetr.util.utils.ModelEma`` implemented as a plain Lightning callback around :class:`torch.optim.swa_utils.AveragedModel`. The ``_avg_fn`` reproduces the exact same formula as ``ModelEma`` (1-indexed ``updates`` counter, optional ``tau`` warm-up). Args: decay: Base EMA decay factor. Corresponds to ``TrainConfig.ema_decay``. tau: Warm-up time constant (in optimizer steps). When > 0 the effective decay ramps from 0 towards *decay* following ``decay * (1 - exp(-updates / tau))``. Corresponds to ``TrainConfig.ema_tau``. use_buffers: Whether buffers are averaged in addition to parameters. update_interval_steps: Update EMA every N optimizer steps. ct|||_||_||_t dt ||_d|_d|_ d|_ d|_ d|_ y)Nr) super__init___decay_tau _use_buffersmaxint_update_interval_steps_average_model_latest_update_step_latest_update_epoch_swapped_state_dict_pending_average_state_dict)selfdecaytau use_buffersupdate_interval_steps __class__s h/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/training/callbacks/ema.pyrzRFDETREMACallback.__init__$sc   '&)!S1F-G&H#7;#$ $&!FJ EI(c|dz}|jdkDr4|jdtj| |jz z z}n |j}||z|d|z zzS)aNCompute the EMA update for a single parameter tensor. Matches the ``ModelEma`` formula where ``updates`` is 1-indexed: PTL's ``num_averaged`` starts at 0 (incremented *after* calling ``avg_fn``), so ``updates = num_averaged + 1`` reproduces the same sequence of effective decay values. Args: averaged_param: Current EMA parameter value. model_param: Corresponding live model parameter value. num_averaged: Number of models averaged so far (0-indexed). Returns: Updated EMA parameter tensor. rrg?)rrmathexp)raveraged_param model_param num_averagedupdateseffective_decays r#_avg_fnzRFDETREMACallback._avg_fn7sc(" 99q="kkQ7(TYY:N1O-OPO"kkO/+AV2WWWr$c|dk7ryt||j|j|j|_|jj |j -|jj|j d|_yt|drt|d}t|tr|jjjj|d}|js |jrIt!j"dt%|jdt%|jd t&d t)|dyy) aInitialise the averaged model at fit start. Args: trainer: The Lightning Trainer instance. pl_module: The ``RFDETRModelModule`` being trained. stage: Current trainer stage (``"fit"``, ``"validate"``, ...). fitN)modeldevicer avg_fn_pending_legacy_ema_stateFstrictz?Legacy EMA checkpoint loaded with non-exact key match; missing=z unexpected=.) stacklevel)r r1rr-revalrload_state_dicthasattrgetattr isinstancedictmoduler0 missing_keysunexpected_keyswarningswarnlen UserWarningdelattr)rtrainer pl_modulestagelegacy_ema_state incompatibles r#setupzRFDETREMACallback.setupRs2 E> +##))<<     "  + + 7    / /0P0P Q/3D , Y ; <&y2MN *D1#2299??OOP`inOo ,, 0L0LMM##&|'@'@#A"BC&&),*F*F&G%HK$#$  I: ;=r$c|duxs|duS)aReturn ``True`` after every optimizer step and every epoch end. The base ``WeightAveraging`` only updates on steps. This override also triggers an update at epoch boundaries, matching RF-DETR's existing EMA behaviour. Args: step_idx: Index of the last optimizer step, or ``None``. epoch_idx: Index of the last epoch, or ``None``. Returns: Whether the averaged model should be updated. N)rstep_idx epoch_idxs r# should_updatezRFDETREMACallback.should_updateys"t#>2F#J $.NN3H"$M!+5>>:T+U(r$c@|j t|jjdsy|jjjj j Dcic]$\}}||j j&c}}Scc}}w)z;Expose EMA model weights for external checkpoint callbacks.Nr0)rr;r?r0rSitemsdetachclone)rkvs r#get_ema_model_state_dictz*RFDETREMACallback.get_ema_model_state_dictsw    &gd6I6I6P6PRY.Z262E2E2L2L2R2R2]2]2_2e2e2gh$!Q188:##%%hhhs-)B)g-?dTr) rfloatrrr boolr!rreturnNone)r( torch.Tensorr)r|r*rrzr|)rGr rHrrIstrrzr{)NN)rO Optional[int]rPr~rzry)rHrrzr{) rGr rHrrXrrYrrZrrzr{)rGr rHrrzr{)rzdict[str, Any])rSrrzr{)rzz!Optional[dict[str, torch.Tensor]])__name__ __module__ __qualname____doc__rr-rLrQrTr\r`rcrfrhrSr:rv __classcell__)r"s@r#r r s " %& JJJ J # J  J&X$X"X X  X6%%%( V ir$r )r __future__rr&rBcopyrtypingrrtorchpytorch_lightningrrr torch.optim.swa_utilsr r rNr$r#rs6 H"  @@/AiAir$