o Zh%@sddlmZmZddlZddlmZddlmZddlm Z m Z m Z ddl m Z mZmZgdZGdddejjjZGd d d eejZGd d d eejZGd ddeejZdS))ClassVarUnionN) _FusedModule) _size_1_t _size_2_t _size_3_t)_pair_single_triple)Conv1dConv2dConv3dc@seZdZUeeejjje d<   dde de de e dfde e dfde e e e dffd e e dfd ed e e dfd e d ede ddfddZddZedddZddZdS)_ConvNd _FLOAT_MODULEN in_channels out_channels kernel_size.stridepaddingdilation transposedoutput_paddinggroupsbias padding_modereturnc Cs\| |d}tjjjj|||||||||| | | f i|| s"Jd| |_| j|d|_dS)N)devicedtypez'qconfig must be provided for QAT module)factory_kwargs)nnmodulesconvr__init__qconfigweightweight_fake_quant)selfrrrrrrrrrrrr#rrrr'K/var/www/auris/lib/python3.10/site-packages/torch/ao/nn/qat/modules/conv.pyr"s(    z_ConvNd.__init__cC||||j|jSNZ _conv_forwardr%r$rr&inputr'r'r(forward6z_ConvNd.forwardFc Cst||jksJd|jd|jjt|dsJd|js$Jdtt|tr/|d}|j}||j|j|j |j |j |j |j |jdu|j|d }|j|_|j|_|S) zCreate a qat module from a float module Args: `mod`: a float module, either produced by torch.ao.quantization utilities or directly from user zqat.z.from_float only works for r#z,Input float module must have qconfig definedz,Input float module must have a valid qconfigrN)rrrrrrr#)typer__name__hasattrr# issubclassrrrrrrrrrrr$)clsmoduse_precomputed_fake_quantr#Zqat_convr'r'r( from_float9s: z_ConvNd.from_floatc Cst|}||j|j|j|j|j|j|j|j du|j }t j |j|_|j dur5t j |j |_ t|trZ|g}t|dsDJ|}|||j|}||j|S|S)zThis works for both single qat conv, and the qat conv - relu modules to convert the qat module to a floating point module N_FLOAT_RELU_MODULE)r0_FLOAT_CONV_MODULErrrrrrrrrtorchr Parameterr$detachr3rr2r8appendrtrainZtraining)r&r4r!r ZreluZfusedr'r'r(to_float\s0     z_ConvNd.to_float)NNNF)r1 __module__ __qualname__rr0rr r!r__annotations__inttuplerstrboolr"r. staticmethodr7r?r'r'r'r(rsD         %  "rcseZdZUdZejZeeeje d<ejZ eeeje d<         dd e d e d e d e d e ee fde de dededdffdd Zedfdd ZZS)r aZ A Conv1d module attached with FakeQuantize modules for weight, used for quantization aware training. We adopt the same interface as :class:`~torch.nn.Conv1d` Similar to :class:`~torch.nn.Conv2d`, with FakeQuantize modules initialized to default. Attributes: weight_fake_quant: fake quant module for weight rr9rTzerosNrrrrrrrrrrc ^t|} t|}t|tr|nt|}t|}tj||| |||dtd||| | | | ddSNFr) rrrrrrrrr#rr)r isinstancerFsuperr"r&rrrrrrrrrr#rrZ kernel_size_Zstride_Zpadding_Z dilation_ __class__r'r(r"( zConv1d.__init__Fctj|||dSN)r6rNr7r4r5r6rPr'r(r7zConv1d.from_float rIrrIrITrJNNNr@)r1rArB__doc__rr rrr0rCr9rDrrrFrGr" classmethodr7 __classcell__r'r'rPr(r |sH     $r ceZdZUdZejZeeeje d<ejZ eeeje d<         dd e d e d e d e d e ee fde de dededdffdd ZddZedfdd ZZS)r a A Conv2d module attached with FakeQuantize modules for weight, used for quantization aware training. We adopt the same interface as `torch.nn.Conv2d`, please see https://pytorch.org/docs/stable/nn.html?highlight=conv2d#torch.nn.Conv2d for documentation. Similar to `torch.nn.Conv2d`, with FakeQuantize modules initialized to default. Attributes: weight_fake_quant: fake quant module for weight rr9rIrTrJNrrrrrrrrrrc rKrL)rrMrFrNr"rOrPr'r(r"rRzConv2d.__init__cCr)r*r+r,r'r'r(r.r/zConv2d.forwardFcrSrTrUrVrPr'r(r7rWzConv2d.from_floatrXr@)r1rArBrYrr rrr0rCr9rDrrrFrGr"r.rZr7r[r'r'rPr(r J    $r cr\)r a A Conv3d module attached with FakeQuantize modules for weight, used for quantization aware training. We adopt the same interface as `torch.nn.Conv3d`, please see https://pytorch.org/docs/stable/nn.html?highlight=conv3d#torch.nn.Conv3d for documentation. Similar to `torch.nn.Conv3d`, with FakeQuantize modules initialized to default. Attributes: weight_fake_quant: fake quant module for weight rr9rIrTrJNrrrrrrrrrrc rKrL)r rMrFrNr"rOrPr'r(r" rRzConv3d.__init__cCr)r*r+r,r'r'r(r.-r/zConv3d.forwardFcrSrTrUrVrPr'r(r70rWzConv3d.from_floatrXr@)r1rArBrYrr rrr0rCr9rDrrrFrGr"r.rZr7r[r'r'rPr(r r]r )typingrrr:Ztorch.nnrZtorch.ao.nn.intrinsicrZtorch.nn.common_typesrrrZtorch.nn.modules.utilsrr r __all__r r!rr r r r'r'r'r(s  n;@