U d1@sbddlZddlZddlmZmZddlmZddlm Z ddl m Z ddZ Gdd d e e Z dS) N)coreschema)get_current_scope) ModelLayer)SamplingTrainableMixincCs|dkstd||S)N)fp32zeOnly support fp32 for the fully connected layer in the predictor net, the provided FC precision is {})AssertionErrorformat) fc_versionr J/tmp/pip-unpacked-wheel-ua33x9lu/caffe2/python/layers/fc_with_bootstrap.pyget_fc_predictor_versions  r c sbeZdZdfdd ZddZdd Zd d Zd d ZddZddZ ddZ e ddZ Z S)FCWithBootstrapNfc_with_bootstrapc stt|j|| |f|t|tjs2td|t| dj dksPtd| dksftd| | |_ t | dj | dd}|dkstd|d|_||_||_g|_d|_d|_d|_d|_| dk rht| dkstd| \}}|dk s|dk std |dks:|dks:||ks:td i|_|dk rT||jd <|dk rh||jd <td |}|r|nd| |df}|r|nd| |df}g}t}t|D]}|td||d|fd||d|f7}|j|jd|||g||| d|jd||g||| dgq||_| dkr\|f}n,t| dj d| d}t||g}dS)NzIncorrect input type {}rz1FC expects limited dimensions of the input tensorrzaxis {} should >= 1.z'FC expects input dimensions > 0, got {}zUclip_param must be a tuple / list of length 2 and in the form of (clip_min, clip max)z8clip_min, and clip_max in clip_param cannot both be Nonez?clip_param = [clip_min, clip_max] must have clip_min < clip_maxminmaxg?Z UniformFill)rrzbootstrap_iteration_{}/indiceszbootstrap_iteration_{}/predszbootstrap_iteration_{}/w) param_nameshapeZ initializerZ optimizerZ regularizerzbootstrap_iteration_{}/b) superr__init__ isinstancerScalarrr lenZ field_typesraxisnpprod clip_args num_bootstrap input_dimsbootstrapped_FCs batch_sizeoutput_dim_vec lower_bound upper_boundmathsqrtStructrangeZget_next_blob_referenceextendZ create_param output_schemalisttuple)selfmodel input_recordZ output_dimsrZ weight_initZ bias_initZ weight_optimZ bias_optimnameZ weight_regZbias_regZ clip_paramrkwargsr Zclip_minZclip_maxZscaler!r+iZ output_shape __class__r r rs        zFCWithBootstrap.__init__c Cstd||dkr||d}|tdgd}|||gd}||_|jtdgdtj d}|jtdgd tj d}|j |jgd tj j d } | | |gd } ||_| |_|j|j|j|jg|j|d ddd} | W5QRSQRXdS)a Args: net: the caffe2 net to insert operator copied_cur_layer: blob of the bootstrapped features (make sure this blob has a stop_gradient on) iteration: the bootstrap interation to generate for. Used to correctly populate the output_schema Return: A blob containing the generated indices of shape: (batch_size,) zbootstrap_iteration_{}r input_shapebatch_size_indexr"r$)Zdtyperoffsetint_batch_size)tor%r)Zinput_as_shapeN)rZ NameScoper ZShapeZConstrarrayGatherr"Zint32ZCastZDataTypeZINT32ZSubr$r%ZUniformIntFillr+ field_blobs) r.netcopied_cur_layer iterationr6r7r"r$r8r9r%indicesr r r _generate_bootstrapped_indicess, z.FCWithBootstrap._generate_bootstrapped_indicescCs4|||g|d|}ttj|jf|}|S)a This method contains all the bootstrapping logic used to bootstrap the features. Only used by the train_net. Args: net: the caffe2 net to insert bootstrapping operators copied_cur_layer: the blob representing the current features. Note, this layer should have a stop_gradient on it. Returns: bootstrapped_features: blob of bootstrapped version of cur_layer with same dimensions zbootstrapped_features_{})r<NextScopedBlobr rrrZfloat32r )r.r>r?rAr@bootstrapped_featuresr r r _bootstrap_opss zFCWithBootstrap._bootstrap_opscCsB|dkr0|j|||fd|ji|j}|Std|dS)a= Args: net: the caffe2 net to insert operator features: Scalar containing blob of the bootstrapped features or actual cur_layer features params: weight and bias for FC outputs: the output blobs version: support fp32 for now. rrzunsupported FC type version {}N)ZFCr=rr2 Exceptionr )r.r>featuresparamsoutputsversionZ pred_blobr r r _insert_fc_opss zFCWithBootstrap._insert_fc_opsc s|jdk rJfdd|D}t||D]\}}j|g|gf|jq&|}|jdksbt|jdkr|j|||j|ddg|ddS)a0 Args: params: the weight and bias, passed by either add_ops or add_train_ops function features: feature blobs to predict on. Can be the actual cur_layer or the bootstrapped_feature blobs. version: currently fp32 support only Ncsg|]}dt|qS)z clipped_%s)rCstr).0pr>r r sz,FCWithBootstrap._add_ops..rr)r>rGrHrIrJ)rzipZClipr#rrKr+r=) r.r>rGr@rHrJZclipped_paramsrNcpr rOr _add_opss zFCWithBootstrap._add_opscCszttjddi}|d}t|jD]N}|j||jd|d}|j |d|dd}|j ||j|||dq&dS)a} Both the predict net and the eval net will call this function. For bootstrapping approach, the goal is to pass the cur_layer feature inputs through all the bootstrapped FCs that are stored under self.bootstrapped_FCs. Return the preds in the same output_schema with dummy indices (because they are not needed). r rrr>r?r@r)r>rGrHr@rJN) rgetr __name__r)rrBr0r=r!rS)r.r> version_infoZpredictor_fc_fp_versionr3rArHr r r add_ops"s&  zFCWithBootstrap.add_opsc Cstt|jD]d}|j||jd|d}|j||jd||d}|j||||j|d|ddddq dS)NrrT)r>r?rAr@rr)rGr@rHrJ)r)rrBr0r=rErSZtrain_param_blobs)r.r>r3rArDr r r add_train_opsCs&  zFCWithBootstrap.add_train_opscCs8|jdkst|jdkr,ddt|jDStddS)NrcSs g|]\}}|ddkr|qS)rrr )rMidxZblobr r r rP]s zBFCWithBootstrap.get_fp16_compatible_parameters..z=Currently only supports functionality for output_dim_vec == 1)r#r enumerater!rFr.r r r get_fp16_compatible_parameters[sz.FCWithBootstrap.get_fp16_compatible_parameterscCs*|jdkst|jdkr|jStddS)Nrz5FCWithBootstrap layer only supports output_dim_vec==1)r#rr!rFr\r r r param_blobsfszFCWithBootstrap.param_blobs) NNNNrNNNr)rV __module__ __qualname__rrBrErKrSrXrYr]propertyr^ __classcell__r r r4r rs((! r)r&ZnumpyrZ caffe2.pythonrrZcaffe2.python.helpers.arg_scoperZcaffe2.python.layers.layersrZ-caffe2.python.layers.sampling_trainable_mixinrr rr r r r s