U zhY @sddlZddlmZmZmZmZmZmZmZm Z m Z m Z ddl Z ddl mZddl mZddlmZedddZe eeefee eeffe eeefee eeffeeeefdd d Zejed ee eeffeegefe e ed fe ed feeeeffd d dZeje e ed fe ed feeeeffdddZeje e ed fe ed feeeeffdddZdejeeeeeeeeejdd d!Zejd"d#d$Zejd"d%d&ZGd'd(d(ejZ Gd)d*d*ejZ!d?ejee e!e ed ffd+d,d-Z"d@ejee e e ed fe ed ffd+d.d/Z#e e ed fd fe ed fd0d1d2Z$eeje e e ed fe ed ffd3d4d5Z%dAe eje e de e&fe j'j(d7d8d9Z)dBe eje e de e&fe j'j(d7d:d;Z*dS)CN) AnyCallableDictIterableListNoReturnSequenceTupleTypeUnion)Tensor)NamedMemberAccessor)returncCs tddS)Na$make_functional(module): we don't yet support models that do parameter tying (also sometimes known as weight sharing). Please try to rewrite your model by replacing all instances of the tied parameter with another and/or comment your support in https://github.com/pytorch/functorch/issues/446) RuntimeErrorrrR/var/www/html/venv/lib/python3.8/site-packages/torch/_functorch/make_functional.pyraise_parameter_tying_error sr) named_paramstied_named_paramsrcCst|}t|}t|}t|}||s6ti}|D]\}}|gf||<qB|D]&\}}||kstt||d|q`t|S)a[ named_params is a dictionary of tensors: {'A': A, 'B': B} tied_named_params is another dictionary of tensors {'A': A, 'B': B, 'B_tied': B} with potentially tied (or 'duplicated') tensors This function creates a mapping from the names in named_params to the names in tied_named_params: {'A': ['A'], 'B': ['B', 'B_tied']}. )dictsetkeysissubsetAssertionErroritemsappendvalues)rrZtensors_dict_keysZtied_tensors_dict_keysZtensor_to_mappingkeyZtensorrrrcreate_names_map*s    r.)mod named_memberssubclassrc Cst|dd}t|dd}t||}i}t|}|D]:\}} | |kr\|tj| dd|| <|| } ||| q6t|dkrd\} } n t|\} } | | |fS)NF)Zremove_duplicateTmeta)devicer)rr)tuplerr torchZ empty_like set_tensorlenzip) r r!r"Zall_named_membersZunique_named_members names_mapmemoaccessornamep replacementnamesparamsrrr_extract_membersFs     r2)r rcCst||jtjS)aZ This function removes all the Parameters from the model and return them as a tuple as well as their original attribute names. The weights must be re-loaded with `load_weights` before the model can be used again. Note that this function modifies the model in place and after this call, mod.parameters() will be empty. )r2Znamed_parametersnn Parameterr rrrextract_weights_s r6cCst||jddS)NcSs|SNr)xrrrpz!extract_buffers..)r2Z named_buffersr5rrrextract_buffersmsr;F)r r0r1 as_paramsrcCs*t|}|rdd|D}|||dS)a  Reload a set of weights so that `mod` can be used again to perform a forward pass. Note that the `params` are regular Tensors (that can have history) and so are left as Tensors. This means that mod.parameters() will still be empty after this call. cSsg|]}t|qSr)r3r4).0r.rrr sz load_weights..Nr Z set_tensors)r r0r1r<r,rrr load_weightsss r@)r r*elemsrc Csfg}t|}t||D]F\\}}}t|D]0\}} |dkrR||| |q.|| |q.q|S)Nr)r r)r enumeraterZ swap_tensorr') r r*rAresultr,_ attr_nameselemi attr_namerrr _swap_statesrI)r r0buffersr<rcCst|}|||dSr7r?)r r0rJr<r,rrr load_bufferssrKr)modelweights weight_namesrJ buffer_namesrcCsPt|t|kstt|||t|dkrLt|t|ks@tt||||S)zload_state(model, weights, weight_names, buffers=(), buffer_names=()) -> model load_state takes `weights` and `buffers` and assigns them to the model. This is the inverse operation of `make_functional_deprecated_v1`. r)r(rr@rK)rLrMrNrJrOrrr load_states    rP)rLcsFt}t|dkr tdt\}}fdd}||fS)amake_functional_deprecated_v1(model) -> weights, func, weight_names Given an nn.Module, make_functional_deprecated_v1 extracts the state (weights) and returns a functional version of the model, `func`. This makes it so that it is possible use transforms over the parameters of `model`. `func` can be invoked as follows: ``` x = torch.randn(4, 3) model = nn.Linear(3, 3) weights, func, _ = make_functional_deprecated_v1(model) func(weights, (x,)) ``` And here is an example of applying the grad transform: ``` x = torch.randn(4, 3) model = nn.Linear(3, 3) weights, _, func = make_functional_deprecated_v1(model) grad_weights = grad(func)(weights, (x,)) ``` To put the state back into a model, use `load_state`. rzmake_functional_deprecated_v1(model): `model` has buffers. Please use make_functional_with_buffers_deprecated_v1(model) instead.cst}t||||Sr7)copydeepcopyr@)rMdata mutable_modelZ descriptorsrLrrfuns  z*make_functional_deprecated_v1..fun)listrJr(rr6)rLrJrMrDrVrrUrmake_functional_deprecated_v1s  rXcs:t\}}t\}}fdd}|||fS)a`make_functional_with_buffers_deprecated_v1(model) -> weights, buffers, func, weight_names, buffer_names Given an nn.Module, make_functional_with_buffers_deprecated_v1 extracts the state (weights and buffers) and returns a functional version of the model, `func`. `func` can be invoked as follows: ``` x = torch.randn(4, 3) model = nn.Linear(3, 3) weights, buffers, func, _, _ = make_functional_with_buffers_deprecated_v1(model) func(weights, buffers, (x,)) ``` And here is an example of applying the grad transform: ``` x = torch.randn(4, 3) model = nn.Linear(3, 3) weights, buffers, func, _, _ = make_functional_with_buffers_deprecated_v1(model) func(weights, buffers, (x,)) grad_weights = grad(func)(weights, buffers, (x,)) ``` To put the state back into a model, use `load_state`. cs*t}t||t||||Sr7)rQrRr@rK)rMrJrSrTZbuf_descriptorsrLZweight_descriptorsrrrVs   z7make_functional_with_buffers_deprecated_v1..fun)r6r;)rLrMrDrJrVrrYr*make_functional_with_buffers_deprecated_v1srZc seZdZdZejeedfeedfeee efeee efddfdd Z e deje edee dfee dffdd d Zee ee ed d d ZZS)FunctionalModuleWithBufferszW This is the callable object returned by :func:`make_functional_with_buffers`. .N)stateless_model param_namesrOparam_names_mapbuffer_names_maprcs6t||_||_||_t||_|j|dSr7)super__init__r\r]rOr all_names_mapupdate)selfr\r]rOr^r_ __class__rrras   z$FunctionalModuleWithBuffers.__init__FrLdisable_autograd_trackingrc CsTt|}t|\}}}t|\}}}|r>|D]} | dq.t|||||||fSNF)rQrRr6r;requires_grad_r[) rLrh model_copyr1r]r^rJrOr_paramrrr _create_froms   z(FunctionalModuleWithBuffers._create_from)r1rJrc OsDt|j|jt|t|}z|j||WSt|j|j|XdSr7)rIr\rbr%)rdr1rJargskwargs old_staterrrforward'sz#FunctionalModuleWithBuffers.forward)F__name__ __module__ __qualname____doc__r3Moduler strrrra staticmethodboolr rmrrrq __classcell__rrrerr[s&  r[c seZdZdZejeedfeee efddfdd Z e deje edee dffdd d Zee ed d d ZZS)FunctionalModulezJ This is the callable object returned by :func:`make_functional`. .N)r\r]r*rcs t||_||_||_dSr7)r`rar\r]r*)rdr\r]r*rerrra<s zFunctionalModule.__init__FrgcCs@t|}t|\}}}|r0|D]}|dq t||||fSri)rQrRr6rjr|)rLrhrkr1r]r*rlrrrrmGs   zFunctionalModule._create_from)r1rc Os8t|j|j|}z|j||WSt|j|j|XdSr7)rIr\r*)rdr1rnrorprrrrqSszFunctionalModule.forward)Frrrrrerr|7s   r|rgcCs.t|}t|dkr tdtj||dS)amake_functional(model, disable_autograd_tracking=False) -> func, params Given a ``torch.nn.Module``, :func:`make_functional` extracts the state (params) and returns a functional version of the model, ``func``. This makes it so that it is possible use transforms over the parameters of ``model``. ``func`` can be invoked as follows: .. code-block:: python import torch import torch.nn as nn from functorch import make_functional x = torch.randn(4, 3) model = nn.Linear(3, 3) func, params = make_functional(model) func(params, x) And here is an example of applying the grad transform over the parameters of a model. .. code-block:: python import torch import torch.nn as nn from functorch import make_functional, grad x = torch.randn(4, 3) t = torch.randn(4, 3) model = nn.Linear(3, 3) func, params = make_functional(model) def compute_loss(params, x, t): y = func(params, x) return nn.functional.mse_loss(y, t) grad_weights = grad(compute_loss)(params, x, t) If the model has any buffers, please use :func:`make_functional_with_buffers` instead. Args: model (torch.nn.Module): Input model. disable_autograd_tracking (bool): Flag to disable gradients tracking for output parameters. The returned params are unrelated to the set of params from the original model. If False (default), the params will have ``requires_grad=True`` on them (aka they will be trackable with regular PyTorch autograd), matching the requires_grad-ness of the params from the original model. Otherwise, the returned params will have ``requires_grad=False``. Default, False. If you plan on using regular PyTorch autograd (e.g., if you want to call ``.backward()`` or ``torch.autograd.grad()``, then set ``disable_autograd_tracking=False``. Otherwise, if you're only planning on using functorch's gradient transforms, then please set ``disable_autograd_tracking=True`` to avoid unnecessarily tracking history with PyTorch autograd. rzdmake_functional(model): `model` has buffers. Please use make_functional_with_buffers(model) instead.rh)rWrJr(rr|rm)rLrhrJrrrmake_functional]s;  r~cCstj||dS)amake_functional_with_buffers(model, disable_autograd_tracking=False) -> func, params, buffers Given a ``torch.nn.Module``, make_functional_with_buffers extracts the state (params and buffers) and returns a functional version of the model ``func`` that can be invoked like a function. ``func`` can be invoked as follows: .. code-block:: python import torch import torch.nn as nn from functorch import make_functional_with_buffers x = torch.randn(4, 3) model = nn.Linear(3, 3) func, params, buffers = make_functional_with_buffers(model) func(params, buffers, x) And here is an example of applying the grad transform over the parameters of a model: .. code-block:: python import torch import torch.nn as nn from functorch import make_functional_with_buffers, grad x = torch.randn(4, 3) t = torch.randn(4, 3) model = nn.Linear(3, 3) func, params, buffers = make_functional_with_buffers(model) def compute_loss(params, buffers, x, t): y = func(params, buffers, x) return nn.functional.mse_loss(y, t) grad_weights = grad(compute_loss)(params, buffers, x, t) Args: model (torch.nn.Module): Input model. disable_autograd_tracking (bool): Flag to disable gradients tracking for output parameters. The returned params are unrelated to the set of params from the original model. If False (default), the params will have ``requires_grad=True`` on them (aka they will be trackable with regular PyTorch autograd), matching the requires_grad-ness of the params from the original model. Otherwise, the returned params will have ``requires_grad=False``. Default, False. If you plan on using regular PyTorch autograd (e.g., if you want to call ``.backward()`` or ``torch.autograd.grad()``, then set ``disable_autograd_tracking=False``. Otherwise, if you're only planning on using functorch's gradient transforms, then please set ``disable_autograd_tracking=True`` to avoid unnecessarily tracking history with PyTorch autograd. r})r[rm)rLrhrrrmake_functional_with_bufferss8r)tuple_of_tuple_of_tensorsrcCs"tt|}tdd|D}|S)Ncss|]}t|VqdSr7r&stackdetachr=Zshardsrrr sz"transpose_stack..)r%r))rresultsrrrtranspose_stacks  r)modelsrcst|dkrtdtdd|Ds@tdd|Ds@tdt|dtfdd|Dsjtdtd d |D\}}}t|}t|}|d||fS) a(combine_state_for_ensemble(models) -> func, params, buffers Prepares a list of torch.nn.Modules for ensembling with :func:`vmap`. Given a list of ``M`` ``nn.Modules`` of the same class, stacks all of their parameters and buffers together to make ``params`` and ``buffers``. Each parameter and buffer in the result will have an additional dimension of size ``M``. :func:`combine_state_for_ensemble` also returns ``func``, a functional version of one of the models in :attr:`models`. One cannot directly run ``func(params, buffers, *args, **kwargs)`` directly, you probably want to use ``vmap(func, ...)(params, buffers, *args, **kwargs)`` Here's an example of how to ensemble over a very simple model: .. code-block:: python num_models = 5 batch_size = 64 in_features, out_features = 3, 3 models = [torch.nn.Linear(in_features, out_features) for i in range(num_models)] data = torch.randn(batch_size, 3) fmodel, params, buffers = combine_state_for_ensemble(models) output = vmap(fmodel, (0, 0, None))(params, buffers, data) assert output.shape == (num_models, batch_size, out_features) .. warning:: All of the modules being stacked together must be the same (except for the values of their parameters/buffers). For example, they should be in the same mode (training vs eval). This API is subject to change -- we're investigating better ways to create ensembles and would love your feedback how to improve this. rz?combine_state_for_ensemble: Expected at least one model, got 0.css|] }|jVqdSr7Ztrainingr=mrrrrsz-combine_state_for_ensemble..css|]}|j VqdSr7rrrrrrszTcombine_state_for_ensemble: Expected all models to have the same training/eval mode.c3s|]}t|kVqdSr7)typerZ model0_typrrrszHcombine_state_for_ensemble: Expected all models to be of the same class.cSsg|] }t|qSr)rr=rLrrrr>!sz.combine_state_for_ensemble..)r(rallrr)r)rfuncsr1rJrrrcombine_state_for_ensembles&( $  rcpu) model_classensemble_shaper$csfdd}|S)Nc stdkrtdtdkr8}t|Sd}|dkrXtd|dtfddt|D}t\}}}tdd|D}tt|}td d|D}|||fS) N,NYI: ensemble_shape with more than 1 elementr num_models  should be > 0c3s|]}VqdSr7tor=rDrnr$rorrrr7sz3functional_init..wrapped..css|]}t|dVqdS)rN)rXrrrrr;scss|]}t|VqdSr7rrrrrr=s)r( ValueErrorrrXr%ranger)) rnrorL num_modelsrrDfnr0rMr$rrrnrorwrapped-s    z functional_init..wrappedrrrr$rrrrfunctional_init(srcsfdd}|S)Nc stdkrtdtdkr8}t|Sd}|dkrXtd|dtfddt|D}t\}}}}}ttdd|D\} } tt| } td d| D} tt| } td d| D} | | |||fS) Nrrrrrc3s|]}VqdSr7rrrrrrRsz@functional_init_with_buffers..wrapped..css|]}t|ddVqdS)Nr)rZrrrrr]scss|]}t|VqdSr7rrrrrrcscss|]}t|VqdSr7rrrrrres)r(rrrXr%rrZr)) rnrorLrrrDrrNrOrMrJrrrrHs8     z-functional_init_with_buffers..wrappedrrrrrfunctional_init_with_buffersCs r)F)F)rr)F)F)rr)rr)+rQtypingrrrrrrrr r r r&Ztorch.nnr3r Z%torch.nn.utils._named_member_accessorr rrxrrwr2r6r;rzr@rIrKrPrXrZr[r|r~rrrinttypesZDevicerrrrrrs0       ( ( (     *%8' G >  @