U 4Af3R@sddlZddlZddlmZmZddlmZddlZddlZ ddl m Z ddl m Z mZmZmZmZmZmZmZmZmZmZmZddlmZmZmZdd l mZerddlZej j!"d erddl#Z#erdd l$m%Z%d Z&ed ddZ'GdddeZ(dS)N)ArgumentParser Namespace) import_module)version) FEATURE_EXTRACTOR_MAPPINGIMAGE_PROCESSOR_MAPPINGPROCESSOR_MAPPINGTOKENIZER_MAPPING AutoConfigAutoFeatureExtractorAutoImageProcessor AutoProcessor AutoTokenizeris_datasets_availableis_tf_availableis_torch_available)TF2_WEIGHTS_INDEX_NAMETF2_WEIGHTS_NAMElogging)BaseTransformersCLICommandF) load_datasetg-C6 ?argsc Cs&t|j|j|j|j|j|j|j|jS)z~ Factory function used to convert a model PyTorch checkpoint in a TensorFlow 2 checkpoint. Returns: ServeCommand ) PTtoTFCommand model_name local_dir max_error new_weightsno_prpushextra_commit_descriptionoverride_model_classrr$B/tmp/pip-unpacked-wheel-zw5xktn0/transformers/commands/pt_to_tf.pyconvert_command_factory9sr&c @sVeZdZeedddZeddZeeee e e eedddZ d d Z d d Z d S)r)parsercCs|jddd}|jdtddd|jdtd d d |jd ttd tdd |jdddd|jdddd|jdddd|jdtd dd |jdtddd |jtddS)z Register this command to argparse so it's available for the transformer-cli Args: parser: Root parser to register command-specific arguments zpt-to-tfzCLI tool to run convert a transformers model from a PyTorch checkpoint to a TensorFlow checkpoint. Can also be used to validate existing weights without opening PRs, with --no-pr.)helpz --model-nameTzAThe model name, including owner/organization, as seen on the hub.)typerequiredr(z --local-dirzOOptional local directory of the model repository. Defaults to /tmp/{model_name})r)defaultr(z --max-errorz%Maximum error tolerance. Defaults to z4. This flag should be avoided, use at your own risk.z --new-weights store_truezKOptional flag to create new TensorFlow weights, even if they already exist.)actionr(z--no-prz6Optional flag to NOT open a PR with converted weights.z--pushzKOptional flag to push the weights directly to `main` (requires permissions)z--extra-commit-descriptionzXOptional additional commit description to use when opening a PR (e.g. to tag the owner).z--override-model-classNzIf you think you know better than the auto-detector, you can specify the model class here. Can be either an AutoModel class or a specific model class like BertForSequenceClassification.)func) add_parser add_argumentstrfloat MAX_ERROR set_defaultsr&)r'Z train_parserr$r$r%register_subcommandLsd z!PTtoTFCommand.register_subcommandcsPt|}t|}||kr6td|d|ddfdd ||iS)zr Compares the TensorFlow and PyTorch outputs, returning a dictionary with all tensor differences. zAThe model outputs have different attributes, aborting. (Pytorch: z, TensorFlow: )r+c st|tjr2tt||}|||<n`|}t|D]R\}}t|trj||}||} ||}n|d|d}||} || ||}q>|S)N[]) isinstancetorchZTensornpmaxabsnumpy enumerater2) Zpt_outZtf_outZ differences attr_nameZtensor_difference root_nameiZpt_item branch_nameZtf_item_find_pt_tf_differencesr$r%rFs    zEPTtoTFCommand.find_pt_tf_differences.._find_pt_tf_differences)r+)setkeys ValueError) pt_outputs tf_outputsZ pt_out_attrsZ tf_out_attrsr$rEr%find_pt_tf_differencess  z$PTtoTFCommand.find_pt_tf_differences)rrrrr r!r"r#c GsRtd|_||_|r|n tjd||_||_||_ ||_ ||_ ||_ ||_ dS)Nztransformers-cli/pt_to_tfz/tmp)rZ get_logger_logger _model_nameospathjoin _local_dir _max_error _new_weights_no_pr_push_extra_commit_description_override_model_class) selfrrrrr r!r"r#rr$r$r%__init__s zPTtoTFCommand.__init__cCs6dd}t|j}|tkrHt|j}|tkr|jjdkr|jj |j_nd|t kr^t |j}nN|t krtt |j}n8|tkrt|j}|jdkr|j |_ntd|dtt|jj}i}d|kr|ddgd d d d |krtd d ddddd} |d| id|krxt|jj} d| krb| dj} | dk r\| dk r\| } nd } nd } ||| dd|kr||d d|f|ddi} |f|ddi}|jst|drt|dsd|kr.tjdgdggtd|jjpd }| dt !|i|dt"#|i| |fS)!zQ Returns the right inputs for the model, based on its signature. cSs@tdddd}|dtdddd}dd |D}|S) Nz)hf-internal-testing/librispeech_asr_dummycleanZ validationsplitidraudiocSsg|] }|dqS)arrayr$).0xr$r$r% szFPTtoTFCommand.get_inputs.._get_audio_input..)rsortselectrange)ZdsZspeech_samplesZ raw_samplesr$r$r%_get_audio_inputs z2PTtoTFCommand.get_inputs.._get_audio_inputNz1Unknown data processing type (model config type: r7Z input_idsz Hi there!z@I am a batch with more than one row and different input lengths.T)textpaddingZ truncationZ pixel_valueszuoft-cs/cifar10Z plain_texttestr\rimgZimagesZinput_featuresriF)r_riZ input_valuesZreturn_tensorspttfencoderdecoderdecoder_input_idsr)Zdtyper)$r)configr rfrom_pretrainedrRr tokenizerZ pad_tokenZ eos_tokenrr rr rrIrGinspect signatureZforward parametersrHupdaterZfeature_extractorr,Zis_encoder_decoderhasattrr<ZasarrayintZdecoder_start_token_idr;ZtensorrmZconvert_to_tensor)rYpt_modeltf_dummy_inputsrqrgZmodel_config_class processorZmodel_forward_signatureZprocessor_inputsZ sample_imagesZfeature_extractor_signatureZdefault_strategyZpadding_strategypt_inputtf_inputrpr$r$r% get_inputssh          $zPTtoTFCommand.get_inputsc! sjdttjtdkr,tdnddlm}m}ddl m }|j j d}t j }|j}jdk r jdrjd dg}njg}zttd |d}Wn&tk rtd jd YnXzttd d|d}Wn(tk rtd jd YnXn|dkrTttd d}ttd d}jdnt|dkrrtd|djd|dttd |d}zttd d|d}Wn*tk rtd|ddYnX||} | j} ~ |j } | | | |\} } t| f| ddi}W5QRX~ |jj dd} | f| ddd}||}dd|D}dd|D}t|dkr|dk rtd|d |rt| nd!}t| }|j!ks|j!krRtd"d#j!d$d%"fd&d'|Dd(j!d$d%"fd)d'|Dt#j$"j t%}t#j$"j t&}t#j$'|st#j$'|rj(r| )j ~ |j }|f| ddi}||}d*d|D}d+d|D}t|dkr|dk rtd|d |r,t| nd!}t| }|j!ksT|j!krtd,d#j!d$d%"fd-d'|Dd(j!d$d%"fd.d'|Dj(rd/nd0}j*r|j+dd1|,||j-dd2jd3j nj.sjd4d5|d6d7|d6d8|d6d9|d6d: }j!t/krX|d;j!d<7}j0rn|d=j07}t#j$'|r|t&|d>g}t1j2j34j d?D] }||t#j$5||d>g7}qn|t%|d>g}|j |||d@ddAj6} jdB| dS)CNak Converting PyTorch weights to TensorFlow is deprecated and will be removed in v4.43. Instead, we recommend that you convert PyTorch weights to Safetensors, an improved format that can be loaded by any framework, including TensorFlow. For more information, please see the Safetensors conversion guide: https://huggingface.co/docs/safetensors/en/convert-weights z0.9.0zrThe huggingface_hub version must be >= 0.9.0 to use this command. Please update your huggingface_hub installation.r) Repository create_commit)CommitOperationAdd)rZ clone_fromZTFrZ transformersz Model class z not found in transformers.zTF model class TFZ AutoModelZ TFAutoModelz5No detected architecture, using AutoModel/TFAutoModelrzAMore than one architecture was found, aborting. (architectures = r7zDetected architecture: zThe TensorFlow equivalent of z doesn't exist in transformers.output_hidden_statesT)Zfrom_ptF)rZtrainingcSsi|]\}}d|kr||qShiddenr$rakvr$r$r% Usz%PTtoTFCommand.run..cSsi|]\}}d|kr||qSrr$rr$r$r%rVsz;Something went wrong -- the config file has architectures (zF), but no model head output was found. All outputs start with 'hidden'gzOThe cross-loaded TensorFlow model has different outputs, something went wrong! z9 List of maximum output differences above the threshold (z):  cs*g|]"\}}|jkr|d|dqSz: .3erSrrYr$r%rcbs z%PTtoTFCommand.run..z@ List of maximum hidden layer differences above the threshold (cs*g|]"\}}|jkr|d|dqSrrrrr$r%rcds cSsi|]\}}d|kr||qSrr$rr$r$r%rrscSsi|]\}}d|kr||qSrr$rr$r$r%rsszLThe converted TensorFlow model has different outputs, something went wrong! cs*g|]"\}}|jkr|d|dqSrrrrr$r%rcs cs*g|]"\}}|jkr|d|dqSrrrrr$r%rcs zUpdate TF weightszAdd TF weights)Zauto_lfs_track)blockingzTF weights pushed into z&Uploading the weights into a new PR...aModel converted by the [`transformers`' `pt_to_tf` CLI](https://github.com/huggingface/transformers/blob/main/src/transformers/commands/pt_to_tf.py). All converted model outputs and hidden layers were validated against its PyTorch counterpart. Maximum crossload output difference=rz,; Maximum crossload hidden layer difference=z'; Maximum conversion output difference=z-; Maximum conversion hidden layer difference=z; zB CAUTION: The maximum admissible error was manually increased to !z )Z path_in_repoZpath_or_fileobjz/tf_model-*.h5model)Zrepo_id operationscommit_messageZcommit_descriptionZ repo_typeZ create_prz PR open in )7rMwarningrparsehuggingface_hub __version__ ImportErrorrrZhuggingface_hub._commit_apirrRrNr rr architecturesrX startswithgetattrrAttributeErrorrIlenZ from_configZ dummy_inputsevalrr;Zno_gradrLitemsr=valuesrSrQrOrPrrexistsrTZsave_pretrainedrVZgit_addZ git_commitZgit_pushrUr4rWrmioZgfileglobbasenameZpr_url)!rYrrrreporqrZpt_classZtf_classZtf_from_pt_modelr{rzr}r~rJZtf_from_pt_outputsZcrossload_differencesZoutput_differencesZhidden_differencesZmax_crossload_output_diffZmax_crossload_hidden_diffZtf_weights_pathZtf_weights_index_pathZtf_modelrKZconversion_differencesZmax_conversion_output_diffZmax_conversion_hidden_diffrZcommit_descritionrZ shard_pathZ hub_pr_urlr$rr%run s              $           $    zPTtoTFCommand.runN) __name__ __module__ __qualname__ staticmethodrr6rLr2r3boolrZrrr$r$r$r%rKs> ' Fr))rtrOargparserr importlibrrr?r< packagingrr+rrr r r r r rrrrrutilsrrrrZ tensorflowrmrqZ experimentalZ enable_tensor_float_32_executionr;Zdatasetsrr4r&rr$r$r$r%s&  8