|jJPUddlmZddlZddlmZddlZddlmZddlmZer ddl m Z ddl m Z ej jej jjejej jjejej jjiZd ed <ejejgZGd d ZGd dZdS)) annotationsN) TYPE_CHECKING)nn)ObserveWrapper)Layer)QuanterFactorydict[Layer, Layer]DEFAULT_QAT_LAYER_MAPPINGScPeZdZdZd dZed dZed d Zd Zd S)SingleLayerConfiga- 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. activationr weightreturnNonec"||_||_dSN _activation_weightselfrrs j/lsinfo/ai/hellotax_ai/data_center/backend/venv/lib/python3.11/site-packages/paddle/quantization/config.py__init__zSingleLayerConfig.__init__1s& c|jSr)rrs rrzSingleLayerConfig.activation7s rc|jSr)rrs rrzSingleLayerConfig.weight;s |rc&d|jd|jS)Nz activation: z weight: rrs r__str__zSingleLayerConfig.__str__?sHd.HH$,HHHrN)rr rr rr)rr ) __name__ __module__ __qualname____doc__rpropertyrrr rrr r (s    X XIIIIIrr c8eZdZdZd2dZ d3d4d Z d3d5dZ d3d6dZd7dZd8dZ e d9dZ d:dZ d:dZ d:dZd:dZd:dZd:dZd:d Zd:d!Zd:d"Ze d;d$Ze d;d%Ze dd.Zd:d/Zd0Zd1ZdS)? QuantConfiga 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) rQuanterFactory | Nonerrrc| |d|_nt|||_i|_i|_i|_d|_t jt|_ i|_ g|_ dSr) _global_configr _layer2config_prefix2config _type2config_modelcopydeepcopyr _qat_layer_mapping_customized_qat_layer_mapping_customized_leavesrs rrzQuantConfig.__init__\ss  &."&D  "3J"G"GD   "&-0J"K"K-/*"$rNlayerLayer | list[Layer]ct|tr|D]}||||dS||||dS)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)rr5rr_elements rr;zQuantConfig.add_layer_configlsN eT " " !  %%F&    !!j !     r layer_namestr | list[str]ct|trt||}||j|<t|tr|D]}||||dSdS)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': } r8N)r9strr r-r:r<)rr?rrconfigr>s rr<zQuantConfig.add_name_configsN j# & & 5&z6::F.4D  + j$ ' ' &  $$F%    r layer_typetype[Layer] | list[type[Layer]]c t|tr9t|tjjrt ||}||j|<t|tr|D]}| |||dSdS)a Set the quantization config by the type of layer. The `layer_type` should be subclass of `paddle.nn.Layer`. Its priority is lower than `add_layer_config` and `add_name_config`. Args: layer_type(type[Layer] | list[type[Layer]]): One or a list of layers' type. It should be subclass of `paddle.nn.Layer`. Python built-in function `type()` can be used to get the type of a layer. activation(QuanterFactory | None): Quanter used for activations. Default is None. weight(QuanterFactory | None): Quanter used for weights. Default is None. Examples: .. code-block:: pycon >>> import paddle >>> from paddle.nn import Linear >>> from paddle.quantization import QuantConfig >>> from paddle.quantization.quanters import ( ... FakeQuanterWithAbsMaxObserver, ... ) >>> class Model(paddle.nn.Layer): ... def __init__(self): ... super().__init__() ... self.fc = Linear(576, 120) >>> model = Model() >>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9) >>> q_config = QuantConfig(activation=None, weight=None) >>> q_config.add_type_config([Linear], activation=quanter, weight=quanter) >>> # doctest: +SKIP('random memory address') >>> print(q_config) Global config: None Layer type config: {: } r8N) r9type issubclasspaddlerrr r.r:add_type_config)rrDrrrCr>s rrJzQuantConfig.add_type_configsV j$ ' ' 3J  - -  3'z6::F,2D j ) j$ ' ' &  $$F%    rsource type[Layer]targetc&t|trt|tjjs Jdt|trt|tjjs Jd||j|<||j|<dS)a( 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&$'' J FIO- -    T   &$'' JJ FIO- -  J J I J J +1'5;*6222rc:|j|dS)a 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)r4append)rrDs radd_customized_leafzQuantConfig.add_customized_leaf-s!& &&z22222rlist[type[Layer]]c|jS)z0 Get all the customized leaves. )r4rs rcustomized_leaveszQuantConfig.customized_leavesBs &&rrcV||o||S)zC Whether the layer should be observed by observer. )_is_leaf_has_observer_configrr5s r _need_observezQuantConfig._need_observeIs)}}U##H(A(A%(H(HHrc||}|jt||jt|}|||Sr)_get_config_by_layerr3getrGqat_layer_mappings)rr5q_config target_types r_get_qat_layerzQuantConfig._get_qat_layerOsc,,U338<< KK044T%[[AA  {5(+++rcF||}|duo|jduS)zT Whether the layer has been configured for activation quantization. N)r\r)rr5_configs rrXz QuantConfig._has_observer_configWs/++E22d"Ew'9'EErc||p)||p||Sr)_is_default_leaf _is_real_leaf_is_customized_leafrYs rrWzQuantConfig._is_leaf^sD  ! !% ( ( /!!%(( /''.. rc.t|tvSr)rGDEFAULT_LEAVESrYs rrezQuantConfig._is_default_leafesE{{n,,rcD|jdupt|jdkS)zA The leaf is real leaf when it has no sublayers. Nr) _sub_layerslenrYs rrfzQuantConfig._is_real_leafhs' D(GC0A,B,Ba,GGrc.t||jvSr)rGrUrYs rrgzQuantConfig._is_customized_leafnsE{{d444rct||}|dn|j}|dn||S)zw Create an instance of observer or quanter according to the given layer's quantization config. N)r\r _instance)rr5rc _observers r _get_observerzQuantConfig._get_observerqsD ++E22#ODD1C  (tti.A.A%.H.HHrcL||}t||Sr)rqr)rr5rps r_get_observe_wrapperz QuantConfig._get_observe_wrapperzs%&&u-- i///rr c|jSr)r2rs rr^zQuantConfig.qat_layer_mappings~s &&rctSr)r rs rdefault_qat_layer_mappingz%QuantConfig.default_qat_layer_mappings))rr c|jSr)r+rs r global_configzQuantConfig.global_configs ""rc8|j|dSr)r,r]rYs rr\z QuantConfig._get_config_by_layers!%%eT222rc||jvS)z The layer is quantifiable when it configured by activation quanter/observer or weight quanter/observer. )r,rYs r_is_quantifiablezQuantConfig._is_quantifiables ***rmodelcH||_||jdS)aA 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)r/_specify_helper)rr|s r_specifyzQuantConfig._specifys'>  T[)))))rcd|D]}|}|j||j}|jt ||}|j||}| ||j|<|||Sr) childrenr=r,r]rxr.rGr-r~)rr|child layer_prefixrCs rr~zQuantConfig._specify_helpers^^%% ( (E ??,,L'++E43EFFF&**4;;??F(,,\6BBF!,2"5)   ' ' ' ' rrBcl|j|S||jS)z> Get the formatted details of current config. )r/r _details_helperrs rdetailszQuantConfig.detailss0 ; <<>> !##DK000rcg}|D]u\}}||}||d}||jvr<|d|zdz|zdzt |j|zv|jjdz}|r|dd|zdzz }|dz }|S)N(z): z, z  )) named_childrenr _addindentr,rQrB __class__r!join)rr5sublayer_linesnamesublayer sublayer_str final_strs rrzQuantConfig._details_helpers#2244  ND(//99L??<;;L4---%%##  $,X677 8O,s2  E &++n"="==D DIS rc"|d}t|dkr|Sg}t|D]3\}}|dkr(|t |dz|z4|ddzd|zS)Nrrr )splitrl enumeraterQrBr)rstringindents1s2idxlines rrzQuantConfig._addindents \\$   r77a<<M "2 6 6ICQww #v|t344555!ut|diimm++rcd}|d|jdz }t|jdkr|d|jdz }t|jdkr|d|jdz }|S)NzGlobal config: rrzLayer type config: zLayer prefix config: )r+rlr.r-)rresults rr zQuantConfig.__str__sBBB BF t" # #a ' ' G0CGGG GF r)rr)rr)rr)NN)r5r6rr)rr)rr)r?r@rr)rr)rr)rDrErr)rr)rr)rKrLrMrLrr)rDrLrr)rrS)r5r)rr )rr )r|r)rrB) r!r"r#r$rr;r<rJrOrRr%rUrZrarXrWrerfrgrqrsr^rvrxr\r{rr~rrrr r&rrr(r(Cs0%%%%&-1(, /////h-1(, .....f-1(, 44444l(<(<(<(rsV#"""""" ######('''''' HM28=,Irx|)Irx|)2 '2<(IIIIIIII6nnnnnnnnnnr