ёiJ%SSKJr SSKrSSKJr SSKrSSKJr SSKJr \(a SSK J r SSK J r \RR\RRR \R"\RR$R&\R(\RR$R*0rS \S '\R0\R2/r"S S 5r"S S5rg)) annotationsN) TYPE_CHECKING)nn)ObserveWrapper)Layer)QuanterFactorydict[Layer, Layer]DEFAULT_QAT_LAYER_MAPPINGSc\\rSrSrSrS Sjr\S Sj5r\S Sj5rSr Sr g) SingleLayerConfig(a Configure how to quantize the activations and weights of a single layer. Args: activation(QuanterFactory): The factory to create instance of quanter used to quantize activations. weight(QuanterFactory): The factory to create instance of quanter used to quantize weights. cXlX lgN _activation_weightself activationweights Z/var/www/html/banglarbhumi/venv/lib/python3.13/site-packages/paddle/quantization/config.py__init__SingleLayerConfig.__init__1s& cUR$r)rrs rrSingleLayerConfig.activation7srcUR$r)rrs rrSingleLayerConfig.weight;s ||rc:SURSUR3$)Nz activation: z weight: rrs r__str__SingleLayerConfig.__str__?s d../z$,,HHrrN)rr rr returnNone)r$r ) __name__ __module__ __qualname____firstlineno____doc__rpropertyrrr"__static_attributes__rrr r (sR(2@    Irr c\rSrSrSrS SjrS!S"SjjrS!S#SjjrS!S$SjjrS%Sjr S&S jr \ S'S j5r S(S jr S(S jrS(S jrS(SjrS(SjrS(SjrS(SjrS(SjrS(Sjr\ S)Sj5r\ S)Sj5r\ S*Sj5rS*SjrS(SjrS+SjrS+SjrS,SjrS(SjrSrSr Sr!g)- QuantConfigCa Configure how to quantize a model or a part of the model. It will map each layer to an instance of SingleLayerConfig by the settings. It provides diverse methods to set the strategies of quantization. Args: activation(QuanterFactory | None): The global quantizer used to quantize the activations. weight(QuanterFactory | None): The global quantizer used to quantize the weights. Examples: .. code-block:: python >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import FakeQuanterWithAbsMaxObserver >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=quanter, weight=quanter) >>> print(q_config) Global config: activation: FakeQuanterWithAbsMaxObserver(name=None,moving_rate=0.9,bit_length=8,dtype=float32) weight: FakeQuanterWithAbsMaxObserver(name=None,moving_rate=0.9,bit_length=8,dtype=float32) cUc UcSUlO[X5Ul0Ul0Ul0UlSUl[ R"[5Ul 0Ul /Ul gr) _global_configr _layer2config_prefix2config _type2config_modelcopydeepcopyr _qat_layer_mapping_customized_qat_layer_mapping_customized_leavesrs rrQuantConfig.__init__\sd  &."&D "3J"GD   "&--0J"K-/*"$rNc[U[5(aUHnURXBUS9 M gURUR 5X#S9 g)a Set the quantization config by layer. It has the highest priority among all the setting methods. Args: layer(Layer|list[Layer]]): One or a list of layers. activation(QuanterFactory | None): Quanter used for activations. Default is None. weight(QuanterFactory | None): Quanter used for weights. Default is None. Examples: .. code-block:: python >>> import paddle >>> from paddle.nn import Linear >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import FakeQuanterWithAbsMaxObserver >>> class Model(paddle.nn.Layer): ... def __init__(self): ... super().__init__() ... self.fc = Linear(576, 120) >>> model = Model() >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=None, weight=None) >>> q_config.add_layer_config([model.fc], activation=quanter, weight=quanter) >>> # doctest: +SKIP('random memory address') >>> print(q_config) Global config: None Layer prefix config: {'linear_0': } rrN) isinstancelistadd_layer_configadd_name_config full_name)rlayerrr_elements rrAQuantConfig.add_layer_configlsUN eT " "!%%F&"  !j ! rc[U[5(a[X#5nX@RU'[U[5(aUHnUR XRUS9 M gg)a Set the quantization config by full name of layer. Its priority is lower than `add_layer_config`. Args: layer_name(str|list[str]): One or a list of layers' full name. activation(QuanterFactory | None): Quanter used for activations. Default is None. weight(QuanterFactory | None): Quanter used for weights. Default is None. Examples: .. code-block:: python >>> import paddle >>> from paddle.nn import Linear >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import FakeQuanterWithAbsMaxObserver >>> class Model(paddle.nn.Layer): ... def __init__(self): ... super().__init__() ... self.fc = Linear(576, 120) >>> model = Model() >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=None, weight=None) >>> q_config.add_name_config([model.fc.full_name()], activation=quanter, weight=quanter) >>> # doctest: +SKIP('random memory address') >>> print(q_config) Global config: None Layer prefix config: {'linear_0': } r>N)r?strr r4r@rB)r layer_namerrconfigrEs rrBQuantConfig.add_name_configs^N j# & &&z:F.4   + j$ ' '&$$F%' (rc[U[5(aB[U[RR 5(a[ X#5nX@RU'[U[5(aUHnURXRUS9 M gg)a Set the quantization config by the type of layer. The `layer_type` should be subclass of `paddle.nn.Layer`. Its priority is lower than `add_layer_config` and `add_name_config`. Args: layer_type(type[Layer] | list[type[Layer]]): One or a list of layers' type. It should be subclass of `paddle.nn.Layer`. Python built-in function `type()` can be used to get the type of a layer. activation(QuanterFactory | None): Quanter used for activations. Default is None. weight(QuanterFactory | None): Quanter used for weights. Default is None. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.nn import Linear >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import ( ... FakeQuanterWithAbsMaxObserver, ... ) >>> class Model(paddle.nn.Layer): ... def __init__(self): ... super().__init__() ... self.fc = Linear(576, 120) >>> model = Model() >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=None, weight=None) >>> q_config.add_type_config([Linear], activation=quanter, weight=quanter) >>> # doctest: +SKIP('random memory address') >>> print(q_config) Global config: None Layer type config: {: } r>N) r?type issubclasspaddlerrr r5r@add_type_config)r layer_typerrrJrEs rrPQuantConfig.add_type_configsyV j$ ' 'J  - - 'z:F,2  j ) j$ ' '&$$F%' (rcP[U[5(a)[U[RR 5(dS5e[U[5(a)[U[RR 5(dS5eX R U'X RU'g)ah Add rules converting layers to simulated quantization layers before quantization-aware training. It will convert layers with type `source` to layers with type `target`. `source` and `target` should be subclass of `paddle.nn.Layer`. And a default mapping is provided by property `default_qat_layer_mapping`. Args: source(type[Layer]): The type of layers that will be converted. target(type[Layer]): The type of layers that will be converted to. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.nn import Conv2D >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import ( ... FakeQuanterWithAbsMaxObserver, ... ) >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=None, weight=None) >>> class CustomizedQuantedConv2D(paddle.nn.Layer): ... def forward(self, x): ... pass ... # add some code for quantization simulation >>> q_config.add_qat_layer_mapping(Conv2D, CustomizedQuantedConv2D) zEThe source layer to be placed should be a subclass of paddle.nn.Layerz&$''J FIIOO- -   T  &$''J FIIOO- -  J I J +1'5;**62rc:URRU5 g)aB Declare the customized layer as leaf of model for quantization. The leaf layer is quantized as one layer. The sublayers of leaf layer will not be quantized. Args: layer_type(type[Layer]): The type of layer to be declared as leaf. Examples: .. code-block:: python >>> from paddle.nn import Sequential >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import FakeQuanterWithAbsMaxObserver >>> q_config = QuantConfig(activation=None, weight=None) >>> q_config.add_customized_leaf(Sequential) N)r;append)rrQs radd_customized_leafQuantConfig.add_customized_leaf-s& &&z2rcUR$)z Get all the customized leaves. )r;rs rcustomized_leavesQuantConfig.customized_leavesBs &&&rcTURU5=(a URU5$)z3 Whether the layer should be observed by observer. )_is_leaf_has_observer_configrrDs r _need_observeQuantConfig._need_observeIs#}}U#H(A(A%(HHrcURU5nURR[U5URR[U555nU"X5$r)_get_config_by_layerr:getrMqat_layer_mappings)rrDq_config target_types r_get_qat_layerQuantConfig._get_qat_layerOsS,,U388<< K0044T%[A 5++rcVURU5nUSL=(a URSL$)zD Whether the layer has been configured for activation quantization. N)rfr)rrD_configs rra QuantConfig._has_observer_configWs/++E2d"Ew'9'9'EErcURU5=(d) URU5=(d URU5$r)_is_default_leaf _is_real_leaf_is_customized_leafrbs rr`QuantConfig._is_leaf^s;  ! !% ( /!!%( /''. rc&[U5[;$r)rMDEFAULT_LEAVESrbs rrqQuantConfig._is_default_leafesE{n,,rc\URSL=(d [UR5S:H$)z1 The leaf is real leaf when it has no sublayers. Nr) _sub_layerslenrbs rrrQuantConfig._is_real_leafhs*  D(GC0A0A,Ba,GGrc2[U5UR;$r)rMr]rbs rrsQuantConfig._is_customized_leafnsE{d4444rcrURU5nUcSO URnUcS$URU5$)z_ Create an instance of observer or quanter according to the given layer's quantization config. N)rfr _instance)rrDrn _observers r _get_observerQuantConfig._get_observerqs@ ++E2#OD1C1C  (tHi.A.A%.HHrc:URU5n[X!5$r)rr)rrDrs r_get_observe_wrapper QuantConfig._get_observe_wrapperzs&&u- i//rcUR$r)r9rs rrhQuantConfig.qat_layer_mappings~s&&&rc[$r)r rs rdefault_qat_layer_mapping%QuantConfig.default_qat_layer_mappings))rcUR$r)r2rs r global_configQuantConfig.global_configs"""rc:URRUS5$r)r3rgrbs rrf QuantConfig._get_config_by_layers!!%%eT22rcXR;$)zi The layer is quantifiable when it configured by activation quanter/observer or weight quanter/observer. )r3rbs r_is_quantifiableQuantConfig._is_quantifiables ****rcFXlURUR5 g)ay Specify the quantization config of each sublayer in model. For each layer in sublayers of mode, 1. Set the config by global config 2. Overwrite the config with parents' config 3. Overwrite the config with config set by layer's type 4. Overwrite the config with config set by layer's full name 5. Overwrite the config with config set by layer Args: model(Layer): The model to be specified by the config. Examples: .. code-block:: python >>> import paddle >>> from paddle.nn import Linear, Sequential >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import FakeQuanterWithAbsMaxObserver >>> class Model(paddle.nn.Layer): ... def __init__(self): ... super().__init__() ... self.fc = Sequential(Linear(576, 120),Linear(576, 120)) >>> model = Model() >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=None, weight=None) >>> q_config.add_layer_config([model.fc], activation=quanter, weight=quanter) >>> q_config._specify(model) N)r6_specify_helper)rmodels r_specifyQuantConfig._specifys>  T[[)rcbUR5HnUR5nURRXR5nUR R[ U5U5nURRX45nUbX@RU'URU5 M U$r) childrenrCr3rgrr5rMr4r)rrchild layer_prefixrJs rrQuantConfig._specify_helpers^^%E ??,L''++E3E3EFF&&**4;?F((,,\BF!,2""5)   '& rcrURcUR5$URUR5$)z. Get the formatted details of current config. )r6r"_details_helperrs rdetailsQuantConfig.detailss/ ;; <<> !##DKK00rc/nUR5Houp4URU5nURUS5nX@R;dM9UR SU-S-U-S-[ URU5-5 Mq UR RS-nU(aUSSRU5-S-- nUS- nU$)N(z): z, z  )) named_childrenr _addindentr3rYrH __class__r&join)rrDsublayer_linesnamesublayer sublayer_str final_strs rrQuantConfig._details_helpers#224ND//9L??<;L---%%##  $,,X67 8 5OO,,s2  &++n"==D DIS rcURS5n[U5S:XaU$/n[U5H-upVUS:dM UR[ US-U-55 M/ USS-SR U5-$)Nrrr )splitrz enumeraterYrHr)rstringindents1s2idxlines rrQuantConfig._addindentss \\$  r7a<M "2ICQw #v|t345'!ut|diim++rcSnUSURS3- n[UR5S:aUSURS3- n[UR5S:aUSURS3- nU$)NzGlobal config: rrzLayer type config: zLayer prefix config: )r2rzr5r4)rresults rr"QuantConfig.__str__s$T%8%8$9<< t  !A % ,T->->,?rB BF t"" #a ' /0C0C/DBG GF r)r;r:r2r3r6r4r9r5)rQuanterFactory | Nonerrr$r%)NN)rDzLayer | list[Layer]rrrrr$r%)rIzstr | list[str]rrrrr$r%)rQztype[Layer] | list[type[Layer]]rrrrr$r%)rT type[Layer]rUrr$r%)rQrr$r%)r$zlist[type[Layer]])rDr)r$r )r$r )rr)r$rH)"r&r'r(r)r*rrArBrPrVrZr+r]rcrkrar`rqrrrsrrrhrrrfrrrrrrr"r,r-rrr/r/Cs0%/%9N% %&-1(, /"/*/& /  /h-1(, .#.*.& .  .f-1(, 434*4& 4  4l(<!(<+6(< (rs# #' HHMM288==,,IIrxx||))IIrxx||))2. ''2<<(II6nnr