NgF)dZddlZddlmZddlZddlmZddlmZm Z m Z m Z m Z ddl mZdZefd Zd Zd Zd ZGd dZgdfdZgddfdZgdfdZgdfdZddejjdejjfdZdS)zM Model / state_dict utils Hacked together by / Copyright 2020 Ross Wightman Ndeepcopy)FrozenBatchNorm2d)BatchNormAct2dSyncBatchNormActFrozenBatchNormAct2dfreeze_batch_norm_2dunfreeze_batch_norm_2d)ModelEmact|trt|jSt |drt|jSt |drt|jS|S)Nmodule _orig_mod) isinstancer unwrap_modelemahasattrrr)models L/var/www/html/ai-engine/env/lib/python3.11/site-packages/timm/utils/model.pyrrsm%""EI&&& 5( # #  -- - UK ( ( 00 0Lc<||SN) state_dict)r unwrap_fns rget_state_dictrs 9U   & & ( ((rctj|gddzS)zA calculate average channel square mean of output activations raxisr)torchmeanitemrinputoutputs ravg_sq_ch_meanr( s7 :fkkyyyk11Q6 7 7 < < > >>rcztj|gdSz> calculate average channel variance of output activations rr r"r#varr$r%s r avg_ch_varr-&2 :fjjiiij00 1 1 6 6 8 88rcztj|gdSr*r+r%s ravg_ch_var_residualr0,r.rc$eZdZdZdZdZdZdS)ActivationStatsHookaIterates through each of `model`'s modules and matches modules using unix pattern matching based on `hook_fn_locs` and registers `hook_fn` to the module if there is a match. Arguments: model (nn.Module): model from which we will extract the activation stats hook_fn_locs (List[str]): List of `hook_fn` locations based on Unix type string matching with the name of model's modules. hook_fns (List[Callable]): List of hook functions to be registered at every module in `layer_names`. Inspiration from https://docs.fast.ai/callback.hook.html. Refer to https://gist.github.com/amaarora/6e56942fcb46e67ba203f3009b30d950 for an example on how to plot Signal Propogation Plots using `ActivationStatsHook`. c"||_||_||_t|t|krt dt d|D|_t||D]\}}|||dS)Nz_Please provide `hook_fns` for each `hook_fn_locs`, their lengths are different.c3(K|] }|jgfVdSr)__name__).0hook_fns r z/ActivationStatsHook.__init__..Ks*IIW7+R0IIIIIIr) r hook_fn_locshook_fnslen ValueErrordictstatszip register_hook)selfrr9r: hook_fn_locr7s r__init__zActivationStatsHook.__init__Ds (  |  H - -.// /IIIIIII $' h$?$? 5 5 K   {G 4 4 4 4 5 5rcfd}|S)Ncl|||}jj|dSr)r>r5append)rr&r'outr7rAs rappend_activation_statszAActivationStatsHook._create_hook..append_activation_statsPs:'&%00C Jw' ( / / 4 4 4 4 4r)rAr7rHs`` r _create_hookz ActivationStatsHook._create_hookOs* 5 5 5 5 5 5'&rc|jD]C\}}tj||s|||DdSr)r named_modulesfnmatchregister_forward_hookrJ)rArBr7namers rr@z!ActivationStatsHook.register_hookVsm J4466 E ELD&?455   ( ():):7)C)C D D D D E ErN)r5 __module__ __qualname____doc__rCrJr@rIrrr2r22sP" 5 5 5'''EEEEErr2)rrTcvtjdd|}t|||}||}|jS)a Extract average square channel mean and variance of activations during forward pass to plot Signal Propogation Plots (SPP). Paper: https://arxiv.org/abs/2101.08692 Example Usage: https://gist.github.com/amaarora/6e56942fcb46e67ba203f3009b30d950 gg?)r9r:)r"normalr2r>)rr9r: input_shapexhook_s rextract_spp_statsr[]s@  R[))A u<( S S SD aA :rTfreezec~|dvs Jdttjjjjtjjjjttfrtdt|tr|g}|}fd|D}t|s+tt\}}t||D]\}}|D]}|dkrdnd|_|rd}|dkrmt#|} t|tjjjjtjjjjttfr ||| t%|} t|t&t(fr ||| d S) a4 Freeze or unfreeze parameters of the specified modules and those of all their hierarchical descendants. This is done in place. Args: root_module (nn.Module, optional): Root module relative to which the `submodules` are referenced. submodules (list[str]): List of modules for which the parameters will be (un)frozen. They are to be provided as named modules relative to the root module (accessible via `root_module.named_modules()`). An empty list means that the whole root module will be (un)frozen. Defaults to [] include_bn_running_stats (bool): Whether to also (un)freeze the running statistics of batch norm 2d layers. Defaults to `True`. mode (bool): Whether to freeze ("freeze") or unfreeze ("unfreeze"). Defaults to `"freeze"`. )r\unfreezez,`mode` must be one of "freeze" or "unfreeze"zYou have provided a batch norm layer as the `root module`. Please use `timm.utils.model.freeze_batch_norm_2d` or `timm.utils.model.unfreeze_batch_norm_2d` instead.c:g|]}|SrI) get_submodule)r6m root_modules r z$_freeze_unfreeze..s'CCC1+++A..CCCrr\FTc|dd}t|dkr7||d|d|dS|||dS)N.r r)rsplitr;r` add_module)rrO submodulesplits r_add_submodulez(_freeze_unfreeze.._add_submodulesr C++u::>>((q22==eAh RRRRR%%dI66666rN)rr"nnmodules batchnorm BatchNorm2d SyncBatchNormrrAssertionErrorstrr;listr?named_children parameters requires_gradr r rr) rb submodulesinclude_bn_running_statsmoderLnraprjress ` r_freeze_unfreezer|os ) ) ) )+Y ) ) )+ H  & 2 H  & 4     m lmm m*c""" \ MCCCC CCCJ z??M$(k.H.H.J.J)K$L$L! zM:..881 B BA'+x'7'7eeTAOO # 8 7 7 7x*1--a(2>(2@&( "8 #N;3777-Q//a"35I!JKK8"N;3777?88rc,t|||ddS)a Freeze parameters of the specified modules and those of all their hierarchical descendants. This is done in place. Args: root_module (nn.Module): Root module relative to which `submodules` are referenced. submodules (list[str]): List of modules for which the parameters will be frozen. They are to be provided as named modules relative to the root module (accessible via `root_module.named_modules()`). An empty list means that the whole root module will be frozen. Defaults to `[]`. include_bn_running_stats (bool): Whether to also freeze the running statistics of `BatchNorm2d` and `SyncBatchNorm` layers. These will be converted to `FrozenBatchNorm2d` in place. Hint: During fine tuning, it's good practice to freeze batch norm stats. And note that these are different to the affine parameters which are just normal PyTorch parameters. Defaults to `True`. Hint: If you want to freeze batch norm ONLY, use `timm.utils.model.freeze_batch_norm_2d`. Examples:: >>> model = timm.create_model('resnet18') >>> # Freeze up to and including layer2 >>> submodules = [n for n, _ in model.named_children()] >>> print(submodules) ['conv1', 'bn1', 'act1', 'maxpool', 'layer1', 'layer2', 'layer3', 'layer4', 'global_pool', 'fc'] >>> freeze(model, submodules[:submodules.index('layer2') + 1]) >>> # Check for yourself that it works as expected >>> print(model.layer2[0].conv1.weight.requires_grad) False >>> print(model.layer3[0].conv1.weight.requires_grad) True >>> # Unfreeze >>> unfreeze(model) r\rwrxNr|rbrvrws rr\r\s#>[*G_fnoooooorc,t|||ddS)a Unfreeze parameters of the specified modules and those of all their hierarchical descendants. This is done in place. Args: root_module (nn.Module): Root module relative to which `submodules` are referenced. submodules (list[str]): List of submodules for which the parameters will be (un)frozen. They are to be provided as named modules relative to the root module (accessible via `root_module.named_modules()`). An empty list means that the whole root module will be unfrozen. Defaults to `[]`. include_bn_running_stats (bool): Whether to also unfreeze the running statistics of `FrozenBatchNorm2d` layers. These will be converted to `BatchNorm2d` in place. Defaults to `True`. See example in docstring for `freeze`. r^r~Nrrs rr^r^s#[*G_fpqqqqqqrFrreturncJ|st|}fd||S)NcL|D]\}}t|dr$t|||nIt|dr|n$t|dr||dS)Nfusereparameterizeswitch_to_deploy)rsrsetattrrrr)ra child_namechild_fuses rrz#reparameterize_model.._fuses!"!1!1!3!3   Juf%% ):uzz||4444 011 )$$&&&& 233 )&&((( E%LLLL  rr)rinplacers @rreparameterize_modelrsC   E%LLL Lr)F)rRrMcopyrr"torchvision.ops.miscr timm.layersrrrr r model_emar rrr(r-r0r2r[r|r\r^rkModulerrIrrrs 22222211111111111111   %1))))??? 999 999 (E(E(E(E(E(E(E(E^%$$ $.0$U]B8B8B8B8J$&ppppD&($ r r r r 58?r