o Zha@sFddlZddlZddlZddlmZmZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZddlZddlmZddlmZmZmZmZdd lmZmZmZe rldd lm Z dd l!m"Z"dd l#m$Z$dd l%m&Z&eruddl'm(Z(e)e*Z+dZ,dZ-ej.GdddZ/GdddeZ0Gddde0eZ1Gddde1Z2dS)N)ABCabstractmethod) OrderedDict) TYPE_CHECKINGAnyCallableDictIterableListMappingOptionalTupleUnion)version) TensorTypeis_torch_availableis_vision_availablelogging)ParameterFormat compute_effective_axis_dimension"compute_serialized_parameters_size)PretrainedConfigFeatureExtractionMixinImageProcessingMixinPreTrainedTokenizerBase)Image lc@sJeZdZUdZeed<eed<eed<dZe eed<dZ e eed<dS) PatchingSpeca Data class that holds patching specifications. Args: o: Module / object where the op to patch is located name: Name of the op to monkey patch custom_op: Custom op that patches the original op orig_op: Original op that is being patched op_wrapper: Wrapper (optional) that wraps both the original and custom ops. It is useful for ops that are class or static methods for instance. oname custom_opNorig_op op_wrapper) __name__ __module__ __qualname____doc__r__annotations__strrr&r r'r.r.G/var/www/auris/lib/python3.10/site-packages/transformers/onnx/config.pyr"/s  r"c@sfeZdZdZdZdZdZedZ e dddd ie d ddd ie dddd ie ddd ddd ddd d e dddd ie dddd ie dd diie ddd ddd d e ddd ddd de ddddddie dddd ie dd diie dddd ie dddd ie dddd idZ d\ddde de eefddZed]ddde ddfddZeedee eee fffd d!Zedee eee fffd"d#Zede ee effd$d%Zedefd&d'Zedefd(d)Zedefd*d+Zedefd,d-Zedefd.d/Z ede!fd0d1Z"e#d2ede!fd3d4Z$ 6d^d7ed8ed9ed:efd;d<Z% ?d_d7ed@edAedBefdCdDZ& E E E F  5 6 6 = > ? d`dGe'dHd7edIedJedKe!dLe e(d8ed:ed9ed@edAedBedMe dNdee effdOdPZ)dQee efdee effdRdSZ*dTdUZ+dVdWZ,edXe dYe-ede.e effdZd[Z/dS)a OnnxConfigzv Base class for ONNX exportable model describing metadata on how to export the model through the ONNX format. rz1.8logitsbatchsequencerrZlast_hidden_state)r3 pred_boxesZ pred_masksr)r3r7)Z start_logitsZ end_logitsZ num_labelsheightwidth)rrrdecoder_sequence)z causal-lmdefaultzimage-classificationzimage-segmentationz masked-imz masked-lmmultiple-choicezobject-detectionzquestion-answeringzsemantic-segmentationz seq2seq-lmzsequence-classificationztoken-classificationz vision2seq-lmz speech2seq-lmr<Nconfigrtaskpatching_specscCs||_||jvrt|d|j||_g|_|dur |ngD]}|}|jdur7tj|t |j |j d}|j |q"dS)Nz+ is not a supported task, supported tasks: )r&) _config_tasks_to_common_outputs ValueErrorkeysr?_patching_specsr& dataclassesreplacegetattrr#r$append)selfr>r?r@specZ final_specr.r.r/__init__os  zOnnxConfig.__init__returncCs |||dS)z Instantiate a OnnxConfig for a specific model Args: config: The model's configuration to use when exporting to ONNX Returns: OnnxConfig for this model )r?r.clsr>r?r.r.r/from_model_configs zOnnxConfig.from_model_configcCst)z Mapping containing the axis definition of the input tensors to provide to the model Returns: For each input: its name associated to the axes symbolic name and the axis position within the tensor )NotImplementedErrorrJr.r.r/inputs zOnnxConfig.inputscCs|j|j}t|S)z Mapping containing the axis definition of the output tensors to provide to the model Returns: For each output: its name associated to the axes symbolic name and the axis position within the tensor )rBr?copydeepcopyrJcommon_outputsr.r.r/outputss  zOnnxConfig.outputscCst|jdr ddiSdS)z Dictionary of keys to override in the model's config before exporting Returns: Dictionary with the keys (and their corresponding values) to override use_cacheFN)hasattrrArRr.r.r/values_overrides zOnnxConfig.values_overridecCtjS)zp The default batch size to use if no other indication Returns: Integer > 0 )r0default_fixed_batchrRr.r.r/default_batch_sizerTzOnnxConfig.default_batch_sizecCr])zu The default sequence length to use if no other indication Returns: Integer > 0 )r0default_fixed_sequencerRr.r.r/default_sequence_lengthz"OnnxConfig.default_sequence_lengthcCr])zw The default number of choices to use if no other indication Returns: Integer > 0 )r0default_fixed_num_choicesrRr.r.r/default_num_choicesrbzOnnxConfig.default_num_choicescCstS)z{ Which onnx opset to use when exporting the model Returns: Integer ONNX Opset version )DEFAULT_ONNX_OPSETrRr.r.r/default_onnx_opsetzOnnxConfig.default_onnx_opsetcCsdS)z What absolute tolerance value to use during model conversion validation. Returns: Float absolute tolerance value. gh㈵>r.rRr.r.r/atol_for_validationrgzOnnxConfig.atol_for_validationcCs(trddlm}t||jkSdS)z The minimum PyTorch version required to export the model. Returns: `bool`: Whether the installed version of PyTorch is compatible with the model. r)get_torch_versionF)rZtransformers.utilsrirparsetorch_onnx_minimum_version)rJrir.r.r/is_torch_support_availables z%OnnxConfig.is_torch_support_availablenum_parameterscCst|tjtkS)a Flag indicating if the model requires using external data format Args: num_parameters: Number of parameter on the model Returns: True if model.num_parameters() * size_of(float32) >= 2Gb False otherwise )rrFloatEXTERNAL_DATA_FORMAT_SIZE_LIMIT)rmr.r.r/use_external_data_formats z#OnnxConfig.use_external_data_formatr:( batch_size num_channels image_height image_widthcCsFg}t|D]}tj|||d}|t|ddq|S)NZuint8RGB) rangenprandomZrandrIr Z fromarrayZastypeconvert)rJrrrsrtruimages_datar.r.r/_generate_dummy_imagess  z!OnnxConfig._generate_dummy_images"V@ sampling_rate time_duration frequencycCsRg}t|D] }tjd|t||dd}|dtdtj||q|S)NrF)Zendpointg?r)rxryZlinspaceintrIsinpi)rJrrrrrZ audio_datar}tr.r.r/_generate_dummy_audio s  $z OnnxConfig._generate_dummy_audioF preprocessor)rrr seq_length num_choicesis_pair framework tokenizerrcsLddlm}ddlm}ddlm}t||r| durtd| dur0t dt t d| }t||rt |tjd d }||}t |tj|d }|jdurYt|jd krY|jnd }d |g|g|}|jd krt tjd d |}|||d}|D]\}fddtd tD||<qt|j|dSt|||dSt||r|jd dkrtd|jjd|jd t |tjd}|||| |}t|||dSt||r|jd dkrt |tjd}|||| |}t|||dSt||r"|jd dkr"t |tjd}||| | | }t|||dStd)am Generate inputs to provide to the ONNX exporter for the specific framework Args: preprocessor: ([`PreTrainedTokenizerBase`], [`FeatureExtractionMixin`], or [`ImageProcessingMixin`]): The preprocessor associated with this model configuration. batch_size (`int`, *optional*, defaults to -1): The batch size to export the model for (-1 means dynamic axis). num_choices (`int`, *optional*, defaults to -1): The number of candidate answers provided for multiple choice task (-1 means dynamic axis). seq_length (`int`, *optional*, defaults to -1): The sequence length to export the model for (-1 means dynamic axis). is_pair (`bool`, *optional*, defaults to `False`): Indicate if the input is a pair (sentence 1, sentence 2) framework (`TensorType`, *optional*, defaults to `None`): The framework (PyTorch or TensorFlow) that the tokenizer will generate tensors for. num_channels (`int`, *optional*, defaults to 3): The number of channels of the generated images. image_width (`int`, *optional*, defaults to 40): The width of the generated images. image_height (`int`, *optional*, defaults to 40): The height of the generated images. sampling_rate (`int`, *optional* defaults to 22050) The sampling rate for audio data generation. time_duration (`float`, *optional* defaults to 5.0) Total seconds of sampling for audio data generation. frequency (`int`, *optional* defaults to 220) The desired natural frequency of generated audio. Returns: Mapping[str, Tensor] holding the kwargs to provide to the model's forward function rrrrNzPYou cannot provide both a tokenizer and a preprocessor to generate dummy inputs.ztThe `tokenizer` argument is deprecated and will be removed in version 5 of Transformers. Use `preprocessor` instead.zROverwriting the `preprocessor` argument with `tokenizer` to generate dummy inputs.r)fixed_dimensionZnum_token_to_add0 r=)Z text_paircsg|] }||qSr.r.).0irvr.r/ tz4OnnxConfig.generate_dummy_inputs..)Z tensor_type)return_tensorsZ pixel_valuesz*The `preprocessor` is an image processor (zC) and expects `model_input_names[0]` to be "pixel_values", but got )r)r|rZinput_featuresz\Unable to generate dummy inputs for the model. Please provide a tokenizer or a preprocessor.) feature_extraction_utilsrimage_processing_utilsrtokenization_utils_baser isinstancerCwarningswarn FutureWarningloggerwarningrr0r^Znum_special_tokens_to_addr`Z unk_tokenlenjoinr?rcitemsrxdictZconvert_to_tensorsZmodel_input_names __class__r(rr)rJrrrrrrrrsrurtrrrrrrrZ token_to_addZ input_tokenZ dummy_inputZtokenized_inputkr.rr/generate_dummy_inputsst 0       &  z OnnxConfig.generate_dummy_inputsreference_model_inputscCs|S)a Generate inputs for ONNX Runtime using the reference model inputs. Override this to run inference with seq2seq models which have the encoder and decoder exported as separate ONNX files. Args: reference_model_inputs ([`Mapping[str, Tensor]`): Reference inputs for the model. Returns: `Mapping[str, Tensor]`: The mapping holding the kwargs to provide to the model's forward function r.)rJrr.r.r/!generate_dummy_inputs_onnxruntimes z,OnnxConfig.generate_dummy_inputs_onnxruntimecC<|jD]}|jdur |jn||j}t|j|j|qdSN)rEr'r%setattrr#r$)rJrKr%r.r.r/ patch_ops zOnnxConfig.patch_opscCrr)rEr'r&rr#r$)rJrKr&r.r.r/ restore_opsrzOnnxConfig.restore_opsr$fieldcs(ddlm}fddt||DS)a Flatten any potential nested structure expanding the name of the field with the index of the element within the structure. Args: name: The name of the nested structure field: The structure to, potentially, be flattened Returns: (Dict[str, Any]): Outputs with flattened structure and key mapping this new structure. r)chaincs i|] \}}d||qS).r.)ridxitemr$r.r/ s zAOnnxConfig.flatten_output_collection_property..) itertoolsr enumerate from_iterable)rOr$rrr.rr/"flatten_output_collection_propertys z-OnnxConfig.flatten_output_collection_property)r<Nr<)rr:rqrq)rrrr) rrrFNr:rqrqrrrN)0r(r)r*r+r^r`rcrrjrkrrBr-r r r"rL classmethodrPpropertyrr rrSrYrr\r_rardrffloatrhboolrl staticmethodrprrrrrrrrr rrr.r.r.r/r0Ds #   "                  "x(r0csVeZdZ   d*dddedeeedeffd d Ze d+ddded dfd d Z e d e ee e effffdd Ze d ee eeffddZe d e fddZe d e fddZ    d,ddde de dedeed e eeff fdd Z d-de ee e effd ed!efd"d#Zd$d%Zd&ed'eed eeefffd(d) ZZS).OnnxConfigWithPastr<NFr>rr?r@use_pastcstj|||d||_dS)N)r?r@)superrLr)rJr>r?r@rrr.r/rLs zOnnxConfigWithPast.__init__rMcCs|||ddS)z Instantiate a OnnxConfig with `use_past` attribute set to True Args: config: The underlying model's config to use when exporting to ONNX Returns: OnnxConfig with `.use_past = True` T)r?rr.rNr.r.r/ with_pasts zOnnxConfigWithPast.with_pastcs tj}|jr|j|dd|S)NrY direction)rrYrfill_with_past_key_values_rWrr.r/rYszOnnxConfigWithPast.outputscCst|jdr d|jiSdS)NrZ)r[rArrRr.r.r/r\s  z"OnnxConfigWithPast.values_overridecCt|jds td|jjS)z The number of layers attribute retrieved from the model config. Override this for model configs where the number of layers attribute is not called `num_layers`. num_layerszcould not find the number of layers attribute in the model configuration, override the num_layers property of the model OnnxConfig to solve this)r[rAAttributeErrorrrRr.r.r/r zOnnxConfigWithPast.num_layerscCr)z The number of attention heads attribute retrieved from the model config. Override this for model configs where the number of attention heads attribute is not called `num_attention_heads`. num_attention_headszcould not find the number of attention heads attribute in the model configuration, override the num_attention_heads property of the model OnnxConfig to solve this)r[rArrrRr.r.r/rrz&OnnxConfigWithPast.num_attention_headsrrrrrrrrcstj|||||d}|jrgtstdddl}|dj\}} | d} ||j| |jj |jf} d|vrL|dj } |j |d|j || | dgdd |d<g|d <t |jD]} |d || || fqU|S) NrrrrrACannot generate dummy past_keys inputs without PyTorch installed.r input_idsrZattention_mask)dtyper)dimpast_key_values)rrrrrCtorchshaperrA hidden_sizercatZonesrxrrIzeros)rJrrrrrr common_inputsrr4ZseqlenZpast_key_values_lengthrZ mask_dtyper}rr.r/rs0     z(OnnxConfigWithPast.generate_dummy_inputsinputs_or_outputsrinverted_values_shapecCs|dvr td|d|dkrdnd}t|jD],}ddd ||d |d <|r8ddd ||d |d <qddd ||d |d <qdS)a Fill the input_or_outputs mapping with past_key_values dynamic axes considering. Args: inputs_or_outputs: The mapping to fill. direction: either "inputs" or "outputs", it specifies whether input_or_outputs is the input mapping or the output mapping, this is important for axes naming. inverted_values_shape: If `True`, store values on dynamic axis 1, else on axis 2. rSrY4direction must either be "inputs" or "outputs", but was givenrSrpresentr4zpast_sequence + sequencerrr.keyr6.valueN)rCrxr)rJrrrr$rr.r.r/r(sz-OnnxConfigWithPast.fill_with_past_key_values_cCs4|d||d|d<|d||d|d<dS)Nrrrrrr.rJflattened_outputr$rrr.r.r/_flatten_past_key_values_Asz,OnnxConfigWithPast._flatten_past_key_values_r$rcsDi}|dvrt|D] \}}|||||q |St||}|S)N)rr)rrrr)rJr$rrrrrr.r/rEsz5OnnxConfigWithPast.flatten_output_collection_property)r<NFrrrFN)F)r(r)r*r-r listr"rrLrrrr rrYrr\rrrrrrr rr __classcell__r.r.rr/rsd   $  + .rcseZdZedeeeeeffffdd Zedeeffdd Z edeeffdd Z   dd e d d edede de e deeeff fdd ZdeeeeeffdefddZddZZS)OnnxSeq2SeqConfigWithPastrMcsrtt|j}|D]"\}}d|vrdnd}|D]\}}d|vr'|||<q|||<qq |jr7|j|dd|S)Nencoderencoder_sequencer;r5rYr)rrrYrrr)rJrXr$Z axes_namesZ sequence_nameZaxis_idxrr.r/rYQs   z!OnnxSeq2SeqConfigWithPast.outputsc\z tj}||f}W|Sty-t|jdr)t|jdr)|jj|jjf}Y|Stdw)Nencoder_layersdecoder_layerszcould not find the number of encoder and decoder layers attributes in the model configuration, override the num_layers property of the model OnnxConfig to solve this)rrrr[rArr)rJrrr.r/rb  z$OnnxSeq2SeqConfigWithPast.num_layerscr)Nencoder_attention_headsdecoder_attention_headszcould not find the number of attention heads for the encoder and the decoder attributes in the model configuration, override the num_attention_heads property of the model OnnxConfig to solve this)rrrr[rArr)rJrrr.r/rrrz-OnnxSeq2SeqConfigWithPast.num_attention_headsrFNrrrrrrrc stt|j|||||d}|js|nd}tt|j|||||d}dd|D}td i||} |jrts=tdddl} | dj d} | dj d} | dj d}|j \} }| | | |j j | f}| ||d |j j |f}g| d <|j \}}t||}t|||}||krd nd }t|D]}| d | || || || |fq|d kr|n|}t||D]}| d | || |fq| S)NrrcSsi|] \}}d||qS)Zdecoder_r.)rr$Ztensorr.r.r/rrzCOnnxSeq2SeqConfigWithPast.generate_dummy_inputs..rrrZdecoder_input_idsr:rrdecoderr.)rrrrrrrrCrrrrArrminmaxrxrIr)rJrrrrrrZencoder_inputsZdecoder_seq_lengthZdecoder_inputsrrr4Zencoder_seq_lengthZnum_encoder_attention_headsZnum_decoder_attention_headsZ encoder_shapeZ decoder_shapenum_encoder_layersnum_decoder_layersmin_num_layersmax_num_layersremaining_side_namer}rrr.r/rsX            z/OnnxSeq2SeqConfigWithPast.generate_dummy_inputsrrc Cs0|dvr td|d|dkrdnd}|j\}}t||}t|||}||kr+dnd}d } |dkr5d nd } t|D]6} d | d ||d| d<d | d ||d| d<d | d ||d| d<d | d ||d| d<q;t||D]} |dkrd | d } nd | d } | ||d| d|d<qwdS)NrrrrSrrrrZpast_encoder_sequenceZpast_decoder_sequencez past_decoder_sequence + sequencer4rr .decoder.key.decoder.value .encoder.key.encoder.valuer)rCrrrrx) rJrrr$rrrrrrr;rZ axes_infor.r.r/rs(     z4OnnxSeq2SeqConfigWithPast.fill_with_past_key_values_cCsd|d||d|d<|d||d|d<|d||d|d<|d||d|d <dS) Nrrrrrrrr:rr.rr.r.r/rsz3OnnxSeq2SeqConfigWithPast._flatten_past_key_values_r)r(r)r*rr r-rrYr rrr rrrrrrrr.r.rr/rPs4$ "Er)3rUrFrabcrr collectionsrtypingrrrrr r r r r rnumpyry packagingrutilsrrrrrrrZconfiguration_utilsrrrrrrrZPILr Z get_loggerr(rrero dataclassr"r0rrr.r.r.r/s8  0       {