o ٷi"@sddlZddlZddlZddlmZddlZddlZddlZddl m Z ddlm Z ddl m Z ddlmZddlmZmZddlmZeeZGd d d ejjZdS) N)Path)convert_float_to_float16) ModelProto) OnnxModel) WhisperConfig)get_model_dynamic_axesget_sample_encoder_inputs)InferenceSessionc seZdZdZdedejjdeffdd Z dej fdd Z d d Z d d Z ddZdedefddZ   ddededededef ddZdededefddZZS)WhisperEncoderzWhisper encoder componentconfigmodel model_implcs@t||_|j|_||_|dkr|j|_dS|jj|_dSNopenai)super__init__r devicer encoderr )selfr r r  __class__k/home/ubuntu/.local/lib/python3.10/site-packages/onnxruntime/transformers/models/whisper/whisper_encoder.pyrs "zWhisperEncoder.__init__audio_featurescCs||}|jdkr |S|jSr)rr last_hidden_state)rroutputsrrrforward%s zWhisperEncoder.forwardcC dg}|S)Nrr)r input_namesrrrr)zWhisperEncoder.input_namescCr)Nencoder_hidden_statesr)r output_namesrrrr!-rzWhisperEncoder.output_namescCst|j||}|S)N)rr )rrr! dynamic_axesrrrr"1szWhisperEncoder.dynamic_axesuse_fp16_inputscCs|jdkr |r t|}|Sr)r r)rr r#rrrfix_layernorm_weights5sz$WhisperEncoder.fix_layernorm_weightsTFonnx_model_pathproviderverboseuse_external_data_formatc Cst|j|jd|d}|}|}|||} t|jjdddt G} t j | d} t| jjddd|r<| n|} tjj||d| d||| dd|d tj| |d } || |} tj| ||dd Wd n1spwY||||d S) aExport encoder to ONNX Args: onnx_model_path (str): path to save ONNX model provider (str): provider to use for verifying parity on ONNX model verbose (bool, optional): print verbose information. Defaults to True. use_external_data_format (bool, optional): use external data format or not. Defaults to False. use_fp16_inputs (bool, optional): use float16 inputs for the audio_features. Defaults to False.  batch_sizeuse_fp16T)parentsexist_okz encoder.onnxr) argsf export_paramsrr!r" opset_versiondo_constant_foldingr')load_external_data)save_as_external_dataall_tensors_to_one_fileN)rr rrr!r"rparentmkdirtempfileTemporaryDirectoryospathjointorchonnxexport load_modelr$rsave verify_onnx)rr%r&r'r(r#inputsrr!r" tmp_dir_nametemp_onnx_model_pathout_pathr rrr export_onnxAsH    zWhisperEncoder.export_onnxc Cst|j|jd|d}||d}t||gd}|dd|did}t ||}t dt dt |dS) a"Verify ONNX model outputs and PyTorch model outputs match Args: onnx_model_path (str): path to save ONNX model provider (str): execution provider for ONNX model use_fp16_inputs (bool, optional): use float16 inputs for the audio_features r)r*r) providersNrz"Comparing encoder_hidden_states...z Max diff: )rr rrdetachcpunumpyr runnpabsloggerwarningmax) rr%r&r#rE pt_outputssess ort_outputsdiffrrrrDs$ zWhisperEncoder.verify_onnx)TFF)__name__ __module__ __qualname____doc__rr?nnModulestrrTensorrrr!r"rboolr$rIrD __classcell__rrrrr s: @r )loggingr<r:pathlibrrMrOr@r?float16rr onnx_modelr transformersrwhisper_inputsrr onnxruntimer getLoggerrXrQr\r]r rrrrs