h0SSKJrJrJrJrJrJr SSKJr SSK J r SSK J r SSK JrJr SSKJrJrJr SSKJrJrJrJrJrJr \ R6"S 5S(S \S \\S \\S\ \\44Sjj5r\ R6"S5S(S \S \\S \\S\ \\44Sjj5r\ R6"S5S)S \S \\S \\S\S\\S\SS\ \\44Sjj5r\ R6"S5S*S SS \\S \\S\S\\S\SS\\\/\ 4S\\\\ S/\4S\ \\44Sjj5r!S\ S\S\S\\\44S jr"S\ S\S\S\\\\/\444S!jr#S\ S"\S\4S#jr$S\ S$\S\4S%jr%S\ S&\4S'jr&g )+)AnyCallableDictOptionalTuplecast)torch)registry)Model)PyTorchGradScaler PyTorchShim) ArgsKwargsFloats3dPadded)convert_recursiveis_torch_array is_xp_arraypartialtorch2xpxp2torchzPyTorchRNNWrapper.v1N pytorch_modelconvert_inputsconvert_outputsreturnc rUc[nUc[n[[[[4[ UUUS95$)z*Wrap a PyTorch RNN model for use in Thinc.rr)convert_rnn_inputsconvert_rnn_outputsrr rPyTorchWrapperrrrs U/home/james-whalen/.local/lib/python3.13/site-packages/thinc/layers/pytorchwrapper.pyPyTorchRNNWrapperr#sD+-  ffn )+  zPyTorchWrapper.v1c fUc[nUc[n[S[XS.[ U5/SSS.S9$)aWrap a PyTorch model, so that it has the same API as Thinc models. To optimize the model, you'll need to create a PyTorch optimizer and call optimizer.step() after each batch. See examples/wrap_pytorch.py Your PyTorch model's forward method can take arbitrary args and kwargs, but must return either a single tensor as output or a tuple. You may find the PyTorch register_forward_hook helpful if you need to adapt the output. The convert functions are used to map inputs and outputs to and from your PyTorch model. Each function should return the converted output, and a callback to use during the backward pass. So: Xtorch, get_dX = convert_inputs(X) Ytorch, torch_backprop = model.shims[0](Xtorch, is_train) Y, get_dYtorch = convert_outputs(Ytorch) To allow maximum flexibility, the PyTorchShim expects ArgsKwargs objects on the way into the forward and backward passed. The ArgsKwargs objects will be passed straight into the model in the forward pass, and straight into `torch.autograd.backward` during the backward pass. NpytorchrnInOattrsshimsdimsconvert_pytorch_default_inputsconvert_pytorch_default_outputsr forwardrr!s r"r r 'sF679 !/T=)* %  r$zPyTorchWrapper.v2mixed_precision grad_scalerdevicez torch.devicec hUc[nUc[n[S[XS.[ UUUUS9/SSS.S9$)aWrap a PyTorch model, so that it has the same API as Thinc models. To optimize the model, you'll need to create a PyTorch optimizer and call optimizer.step() after each batch. See examples/wrap_pytorch.py Your PyTorch model's forward method can take arbitrary args and kwargs, but must return either a single tensor as output or a tuple. You may find the PyTorch register_forward_hook helpful if you need to adapt the output. The convert functions are used to map inputs and outputs to and from your PyTorch model. Each function should return the converted output, and a callback to use during the backward pass. So: Xtorch, get_dX = convert_inputs(X) Ytorch, torch_backprop = model.shims[0](Xtorch, is_train) Y, get_dYtorch = convert_outputs(Ytorch) To allow maximum flexibility, the PyTorchShim expects ArgsKwargs objects on the way into the forward and backward passed. The ArgsKwargs objects will be passed straight into the model in the forward pass, and straight into `torch.autograd.backward` during the backward pass. mixed_precision: Enable mixed-precision. This changes whitelisted ops to run in half precision for better performance and lower memory use. grad_scaler: The gradient scaler to use for mixed-precision training. If this argument is set to "None" and mixed precision is enabled, a gradient scaler with the default configuration is used. device: The PyTorch device to run the model on. When this argument is set to "None", the default device for the currently active Thinc ops is used. Nr&r)r2r3r4r'r*r.)rrrr2r3r4s r"PyTorchWrapper_v2r6OsWT79 !/T  /'    %  r$zPyTorchWrapper.v3ztorch.nn.Moduleserialize_modeldeserialize_modelclUc[nUc[n[S[XS.[ UUUUUUS9/SSS.S9$)aWrap a PyTorch model, so that it has the same API as Thinc models. To optimize the model, you'll need to create a PyTorch optimizer and call optimizer.step() after each batch. See examples/wrap_pytorch.py Your PyTorch model's forward method can take arbitrary args and kwargs, but must return either a single tensor or a tuple. You may find the PyTorch register_forward_hook helpful if you need to adapt the output. The convert functions are used to map inputs and outputs to and from your PyTorch model. Each function should return the converted output, and a callback to use during the backward pass. So: Xtorch, get_dX = convert_inputs(X) Ytorch, torch_backprop = model.shims[0](Xtorch, is_train) Y, get_dYtorch = convert_outputs(Ytorch) To allow maximum flexibility, the PyTorchShim expects ArgsKwargs objects on the way into the forward and backward passed. The ArgsKwargs objects will be passed straight into the model in the forward pass, and straight into `torch.autograd.backward` during the backward pass. mixed_precision: Enable mixed-precision. This changes whitelisted ops to run in half precision for better performance and lower memory use. grad_scaler: The gradient scaler to use for mixed-precision training. If this argument is set to "None" and mixed precision is enabled, a gradient scaler with the default configuration is used. device: The PyTorch device to run the model on. When this argument is set to "None", the default device for the currently active Thinc ops is used. serialize_model: Callback that receives the wrapped PyTorch model as its argument and returns a "bytes" representation of the same. The representation should contain all the necessary information to fully deserialize the model. When set to "None", the default serializer serializes the model's parameters. deserialize_model: Callback that receives the default PyTorch model (passed to the constructor), the serialized "bytes" representation and a PyTorch device. It should return a fully deserialized model on the target device as its result. When set to "None", the default deserializer deserializes the model's parameters. Nr&r)r2r3r4r7r8r'r*r.)rrrr2r3r4r7r8s r"PyTorchWrapper_v3r:s]l79 !/T  /' /"3     % r$modelXis_trainc^ ^ ^ URSnURSnU"XU5unm URS"XR5unm U"XU4U5unm S[S[4U U U 4SjjnXx4$)zwReturn the output of the wrapped PyTorch model for the given input, along with a callback to handle the backward pass. rrrdYrc8>T"U5nT"U5nT"U5nU$N)r?dYtorchdXtorchdXget_dX get_dYtorchtorch_backprops r"backpropforward..backprops$b/ ) G_ r$)r+r,r) r;r<r=rrXtorchYtorchYrIrFrGrHs @@@r"r1r1s[[!12Nkk"34O#Eh7NFF"[[^F=FN$UKBNA{SS ;r$c^^[[URS5mUU4Sjn[[X15n[ U[ 5(aSnXE4$[ U[5(aSn[ [5US9U4$[ U[[45(aSn[ [U50S9U4$Sn[ U40S9U4$)Nrc.>[UTTRS9$)N requires_gradr4)rr4)xr=shims r"0convert_pytorch_default_inputs..s(1HT[[Qr$c,[[[U5$rA)rrr)rDs r"reverse_conversion:convert_pytorch_default_inputs..reverse_conversions$^XwG Gr$cD[[[U5nUR$rA)rrrkwargsrDrEs r"rWrXs">8WEB99 r$argsrZcD[[[U5nUR$rArrrr]r[s r"rWrXs">8WEB77Nr$cJ[[[U5nURS$Nrr_r[s r"rWrX s">8WEB771: r$) rrr,rr isinstancerdicttuplelist)r;r<r= xp2torch_ convertedrWrSs ` @r"r/r/s  U[[^ ,DQI!+y[[[[TRS9U5n[ T44SU0S9$)N)r4 grad_tensorsr\)rrrrr4r)r?rCrLrSs r"rW;convert_pytorch_default_outputs..reverse_conversions7# $++>  |^W4MNNr$)rrr,rrrrr)r;rhr=r<rMrWrLrSs @@r"r0r0sU  U[[^ ,DIAv.(F;AOsOzOO  r$Xpc^^^[[URS5nURmURmUR mS[ S[4UUU4Sjjn[ [URSURS9S40S9nXT4$)Nrd_inputsrcP>[URS5n[UTTT5$ra)rr]r)rorEindiceslengths size_at_ts r"convert_from_torch_backward7convert_rnn_inputs..convert_from_torch_backward*s' hmmA& 'b)Wg66r$TrPr\) rrr,rsrrrqrrrdatar4) r;rmr=rSrtoutputrqrrrss @@@r"rr$s}  U[[^ ,D IjjGjjG7j7V77rwwd4;;G NF  ..r$inputs_outputsc^^ [[URS5m UunumnS[S[4UU 4Sjjn[[ [ T55n[XcRURUR5nXu4$)NrdYprc\>[URSTRS9n[T4SU0S9$)NTrPrkr\)rrvr4r)rzrCrLrSs r"convert_for_torch_backward7convert_rnn_outputs..convert_for_torch_backward9s-3884 Ly.'1JKKr$) rrr,rrrrrsrrrq) r;rxr=rm_r|rMYprLrSs @@r"rr5sv  U[[^ ,D$OB LL:LL Xx'(A <<RZZ 8B ))r$)NN)NNFNN)NNFNNNN)'typingrrrrrrcompatr configr r;r r,r rtypesrrrutilrrrrrrlayersr#r boolr6bytesr:r1r/r0rrrBr$r"rs==200 '(*.*.X&h' 66> )( $%*.*.$$X&$h'$ 38_ $&$N $%*.*.!/3'+ ::X&:h': : +, : ^ $ : 38_:&:z $%*.*.!/3'+8