a Sicn @sdZddlZddlZddlZddlZddlZ ddl Z ddl Z ddl Z ddlmZddlmZddlmZddlmZddlmZddlmZdd lmZdd lmZdd lmZdd lmZdd lmZ ddl!m"Z"ddl#m$Z$ddl#m%Z%ddl&m'Z(ddl)m*Z*ddl+m,Z-e.e/dddZ0dCe$j1e2e j3dddZ4dDe$j1e2e j3dddZ5ej6rbdZ7ne(8dZ7ej9:e7s~dZ7dd Z;e-d!Gd"d#d#ej<Z=eej>ej?d$d$e j@e$jAe$jAd$d$ddf d%d&ZBeej>ejCd'd(ZDd)d*ZEdEd,d-ZFeejGejHdFd.d/ZIdGd0d1ZJe$jAdejKejLdd+d$d$d$d$dd$dddd$d+d+d+d$dd$d$dd$d$dd$d$d$d$d$f d2d3ZMeejGejNd4d5ZOeejGejNd6d7ZPeejGejQd8d9ZReejGejSd:d;ZTe-dd?d@ZVdAdBZWdS)Hz1Converts a frozen graph into a TFLite FlatBuffer.N)lite_constants)util) wrap_toco) Component) convert_phase)ConverterError) SubComponent)metrics_wrapper)model_flags_pb2)toco_flags_pb2) types_pb2)flatbuffer_utils)dtypes) tensor_shape)resource_loader) deprecation) tf_export)conversion_flagsreturncCs(tjtjg}|j|vs |j|vo&|j S)a7Checks if the `quantized_input_stats` flag is required for conversion. Args: conversion_flags: A protocol buffer describing the conversion process. Returns: True, if the `inference_type` or the `inference_input_type` is a quantized type and it is not post training quantization, else False. ) _types_pb2QUANTIZED_UINT8QUANTIZED_INT8inference_typeinference_input_typepost_training_quantize)rZquantized_inference_typesrZ/var/www/html/django/DPS/env/lib/python3.9/site-packages/tensorflow/lite/python/convert.py"_is_quantized_input_stats_required.s  r)tf_typeusagercCstjtjtjtjtjtjtjtj tj tj tj tj tjtjtjtjtjtjtjtjtjtjtjtjtjtjtjtjtjtji}| |}|durt!d"|||S)aConvert tensor type from tf type to tflite type. Args: tf_type: TensorFlow type. usage: Text describing the reason for invoking this function. Raises: ValueError: If `tf_type` is unsupported. Returns: tflite_type: TFLite type. Refer to lite/toco/types.proto. N6Unsupported TensorFlow type `{0}` provided for the {1})#rfloat16rFLOAT16float32FLOATfloat64FLOAT64int8INT8int16INT16uint16UINT16int32INT32int64INT64uint8UINT8uint32UINT32uint64UINT64stringSTRINGboolBOOL complex64 COMPLEX64 complex128 COMPLEX128get ValueErrorformatrr mappingZ tflite_typerrr%convert_tensor_tf_type_to_tflite_typeAs0 rEcCsJtjtjtjtjtjtjtjtj i}| |}|durFt d |||S)a!Convert inference type from tf type to tflite type. Args: tf_type: TensorFlow type. usage: Text describing the reason for invoking this function. Raises: ValueError: If `tf_type` is unsupported. Returns: tflite_type: TFLite type. Refer to lite/toco/types.proto. Nr!) rr$rr%r2rr(rr*QUANTIZED_INT16r@rArBrCrrr(convert_inference_tf_type_to_tflite_typeks rGz../toco/python/toco_from_protosZtoco_from_protoscCs:|dur dSt|tr6z t|WSty4Yn0|S)Nr) isinstancebytessix ensure_textUnicodeDecodeError)outputrrr_try_convert_to_unicodes   rNz lite.OpsSetc@s4eZdZdZdZdZdZdZddZe dd Z d S) OpsSetzEnum class defining the sets of ops available to generate TFLite models. WARNING: Experimental interface, subject to change. TFLITE_BUILTINS SELECT_TF_OPSTFLITE_BUILTINS_INT8;EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8cCs t|jSN)strvalue)selfrrr__str__szOpsSet.__str__cCsddttDS)z6Returns a list of OpsSet options as a list of strings.cSsg|] }t|qSr)rU).0optionrrr z&OpsSet.get_options..)listrOrrrr get_optionsszOpsSet.get_optionsN) __name__ __module__ __qualname____doc__rPrQrRrSrX staticmethodr^rrrrrOsrOFc Cs$t||||t|t||||| S)aQuantize `input_data_str` with calibration results. Args: input_data_str: Input data in serialized form (e.g. a TFLITE model with calibration results). disable_per_channel: Bool indicating whether to do per-channel or per-tensor quantization fully_quantize: Bool indicating whether to fully quantize the model. Besides model body, the input/output will be quantized as well. inference_type: Data type for the activations. The default value is int8. input_data_type: Data type for the inputs. The default value is float32. output_data_type: Data type for the outputs. The default value is float32. enable_numeric_verify: Experimental. Subject to change. Bool indicating whether to add NumericVerify ops into the debug mode quantized model. enable_whole_model_verify: Experimental. Subject to change. Bool indicating whether to add verification for layer by layer, or on whole model. When disabled (per-layer) float and quantized ops will be run from same input (output of previous quantized layer). When enabled, float and quantized ops will run with respective float and quantized output of previous ops. denylisted_ops: Experimental. Subject to change. Set of ops to denylist. denylisted_nodes: Experimental. Subject to change. Set of notes to denylist. Returns: Quantized model in serialized form (e.g. a TFLITE model) with floating-point inputs and outputs. )rZ"wrapped_experimental_mlir_quantizerE) input_data_strZdisable_per_channelZfully_quantizerZinput_data_typeZoutput_data_typeZenable_numeric_verifyZenable_whole_model_verifyZdenylisted_opsZdenylisted_nodesrrr mlir_quantizes%recCs t|S)zSparsify `input_data_str` to encode sparse tensor with proper format. Args: input_data_str: Input data in serialized form (e.g. a TFLITE model). Returns: Sparsified model in serialized form (e.g. a TFLITE model). )rZ"wrapped_experimental_mlir_sparsify)rdrrr mlir_sparsifys rfcCs t|S)aRegister the given custom opdefs to the TensorFlow global op registry. Args: custom_opdefs_list: String representing the custom ops OpDefs that are included in the GraphDef. Returns: True if the registration is successfully completed. )rZwrapped_register_custom_opdefs)Zcustom_opdefs_listrrrregister_custom_opdefss rgTc Cs||stsnzt|||||}|WStyl}z4tt|}tD]}||qD|WYd}~n d}~00t ||||S)acConverts `input_data_str` to a TFLite model. Args: model_flags_str: Serialized proto describing model properties, see `model_flags.proto`. conversion_flags_str: Serialized proto describing conversion properties, see `toco/toco_flags.proto`. input_data_str: Input data in serialized form (e.g. a graphdef is common, or it can be hlo text or proto) debug_info_str: Serialized `GraphDebugInfo` proto describing logging information. (default None) enable_mlir_converter: Enables MLIR-based conversion. (default True) Returns: Converted model in serialized form (e.g. a TFLITE model is common). Raises: ConverterError: When conversion fails in TFLiteConverter, usually due to ops not being supported. RuntimeError: When conversion fails, an exception is raised with the error message embedded. N) _deprecated_conversion_binaryrZwrapped_toco_convert ExceptionrrU_metrics_wrapperZretrieve_collected_errorsZ append_error!_run_deprecated_conversion_binary) model_flags_strconversion_flags_strrddebug_info_strenable_mlir_converterZ model_streZconverter_errorZ error_datarrrconverts"   rqcCstjtdurtdd\}}}}ztjdd}tjdd} tjdd} tjdd~} |j}| j}| j}| j} | |||| t ||r|nd}t |t s| | dn | |Wdn1s0YWdn1s0YWdn1s$0YWdn1sD0Ytjdd} | j}Wdn1sx0Yt||||d| g}d |}td k}tj|d tjtj| d }|\}}|j}|d krXt|dV} | WdW||||fD].}zt|Wnttfy<Yn0qS1sL0Yn t|}t|}td||fW||||fD].}zt|WnttfyYn0qn>||||fD].}zt|WnttfyYn0q0dS)aConvert `input_data_str` using deprecated conversion binary. Args: model_flags_str: Serialized proto describing model properties, see `model_flags.proto`. conversion_flags_str: Serialized proto describing TFLite converter properties, see `toco/toco_flags.proto`. input_data_str: Input data in serialized form (e.g. a graphdef is common) debug_info_str: Serialized `GraphDebugInfo` proto describing logging information. (default None) Returns: Converted model in serialized form (e.g. a TFLITE model is common). Raises: ConverterError: When cannot find the deprecated conversion binary. RuntimeError: When conversion fails, an exception is raised with the error message embedded. NazCould not find `toco_from_protos` binary, make sure your virtualenv bin directory or pip local bin directory is in your path. In particular, if you have installed TensorFlow with --user, make sure you add the install directory to your path. For example: Linux: export PATH=$PATH:~/.local/bin/ Mac: export PATH=$PATH:~/Library/Python//bin Alternative, use virtualenv.)NNNNF)deleterzutf-8z--debug_proto_file={} WindowsT)shellstdoutstderr close_fdsrrbzSee console for info. %s %s ) distutilsspawnfind_executablerhr _tempfileNamedTemporaryFilenamewriterJ ensure_binaryrHrIencoderBjoin _platformsystem _subprocessPopenPIPESTDOUT communicate returncodeopenread_osunlinkOSError TypeErrorrN)rlrmrdrnZconversion_filenameZmodel_filenameZinput_filenameZoutput_filenameZ fp_conversionZfp_modelZfp_inputZfp_debugdebug_filenamefpcmdcmdlineZ is_windowsprocrvrwexitcodefilenamerrrrk>s       &       rkcKsHt}||_||_|r||_||_|r4|j||rD|j||S)aBuilds the model flags object from params. Args: change_concat_input_ranges: Boolean to change behavior of min/max ranges for inputs and outputs of the concat operator for quantized models. Changes the ranges of concat operator overlap when true. (default False) allow_nonexistent_arrays: Allow specifying array names that don't exist or are unused in the final graph. (default False) saved_model_dir: Filepath of the saved model to be converted. This value will be non-empty only when the saved model import path will be used. Otherwises, the graph def-based conversion will be processed. saved_model_version: SavedModel file format version of The saved model file to be converted. This value will be set only when the SavedModel import path will be used. saved_model_tags: Set of string saved model tags, formatted in the comma-separated value. This value will be set only when the SavedModel import path will be used. saved_model_exported_names: Names to be exported (default: export all) when the saved model import path is on. This value will be set only when the SavedModel import path will be used. Returns: model_flags: protocol buffer describing the model. ) _model_flags_pb2 ModelFlagschange_concat_input_rangesallow_nonexistent_arrayssaved_model_dirsaved_model_versionsaved_model_tagsextendsaved_model_exported_names)rrrrrr_ model_flagsrrrbuild_model_flagss  rc "KsPt}!t|dd|!_|r*t|dd|!_n|!j|!_||!_||!_|rV|d|!_|d|!_||!_ ||!_ ||!_ ||!_ | |!_ | r~| |!_| |!_| rtj| vrd|!_t| tjhkrd|!_| r| |!_|r|!j|||!_||!_||!_||!_||!_|rt|dd|!_||!_||!_ |r|!j!|||!_"||!_#||!_$||!_%||!_&|r@||!_'||!_(||!_)|!S)aXBuilds protocol buffer describing a conversion of a model. Typically this is to convert from TensorFlow GraphDef to TFLite, in which case the default `input_format` and `output_format` are sufficient. Args: inference_type: Data type of numeric arrays, excluding the input layer. (default tf.float32, must be in {tf.float32, tf.int8, tf.uint8}) inference_input_type: Data type of the numeric arrays in the input layer. If `inference_input_type` is in {tf.int8, tf.uint8}, then `quantized_input_stats` must be provided. (default is the value assigned to `inference_type`, must be in {tf.float32, tf.int8, tf.uint8}) input_format: Type of data to read. (default TENSORFLOW_GRAPHDEF, must be in {TENSORFLOW_GRAPHDEF}) output_format: Output file format. (default TFLITE, must be in {TFLITE, GRAPHVIZ_DOT}) default_ranges_stats: Tuple of integers representing (min, max) range values for all arrays without a specified range. Intended for experimenting with quantization via "dummy quantization". (default None) drop_control_dependency: Boolean indicating whether to drop control dependencies silently. This is due to TFLite not supporting control dependencies. (default True) reorder_across_fake_quant: Boolean indicating whether to reorder FakeQuant nodes in unexpected locations. Used when the location of the FakeQuant nodes is preventing graph transformations necessary to convert the graph. Results in a graph that differs from the quantized training graph, potentially causing differing arithmetic behavior. (default False) allow_custom_ops: Boolean indicating whether to allow custom operations. When false any unknown operation is an error. When true, custom ops are created for any op that is unknown. The developer will need to provide these to the TensorFlow Lite runtime with a custom resolver. (default False) post_training_quantize: Boolean indicating whether to quantize the weights of the converted float model. Model size will be reduced and there will be latency improvements (at the cost of accuracy). (default False) quantize_to_float16: Boolean indicating whether to convert float buffers to float16. (default False) dump_graphviz_dir: Full filepath of folder to dump the graphs at various stages of processing GraphViz .dot files. Preferred over --output_format=GRAPHVIZ_DOT in order to keep the requirements of the output file. (default None) dump_graphviz_video: Boolean indicating whether to dump the graph after every graph transformation. (default False) target_ops: Experimental flag, subject to change. Set of OpsSet options indicating which converter to use. (default set([OpsSet.TFLITE_BUILTINS])) conversion_summary_dir: A string, the path to the generated conversion logs. select_user_tf_ops: List of user's defined TensorFlow ops need to be supported in the TensorFlow Lite runtime. These ops will be supported as select TensorFlow ops. allow_all_select_tf_ops: If True, automatically add all TF ops (including custom TF ops) to the converted model as flex ops. enable_tflite_resource_variables: Experimental flag, subject to change. Enables conversion of resource variables. (default False) unfold_batchmatmul: Whether to unfold tf.BatchMatMul to a set of tfl.fully_connected ops. If not, translate to tfl.batch_matmul. lower_tensor_list_ops: Whether to lower tensor list ops to builtin ops. If not, use Flex tensor list ops. default_to_single_batch_in_tensor_list_ops: Whether to force to use batch size one when the tensor list ops has the unspecified batch size. accumulation_type: Data type of the accumulators in quantized inference. Typically used for float16 quantization and is either fp16 or fp32. allow_bfloat16: Whether the converted model supports reduced precision inference with the bfloat16 type. unfold_large_splat_constant: Whether to unfold large splat constant tensors in the flatbuffer model to reduce size. supported_backends: List of TFLite backends which needs to check compatibility. disable_per_channel_quantization: Disable per-channel quantized weights for dynamic range quantization. Only per-tensor quantization will be used. enable_mlir_dynamic_range_quantizer: Enable MLIR dynamic range quantization. If False, the old converter dynamic range quantizer is used. tf_quantization_mode: Indicates the mode of TF Quantization when the output model is used for TF Quantization. disable_infer_tensor_range: Disable infering tensor ranges. use_fake_quant_num_bits: Allow quantization parameters to be calculated from num_bits attribute. enable_dynamic_update_slice: Enable to convert to DynamicUpdateSlice op. (default: False). preserve_assert_op: Whether to preserve `TF::AssertOp` (default: False). guarantee_all_funcs_one_use: Whether to clone functions so that each function only has a single use. This option will be helpful if the conversion fails when the `PartitionedCall` or `StatefulPartitionedCall` can't be properly inlined (default: False). Returns: conversion_flags: protocol buffer describing the conversion process. Raises: ValueError, if the input tensor type is unknown. zinference_type flagr zinference_input_type flagrTzaccumulation_type flag)*_conversion_flags_pb2 TocoFlagsrGrr input_format output_formatdefault_ranges_mindefault_ranges_maxdrop_control_dependencyreorder_across_fake_quantallow_custom_opsrquantize_to_float16dump_graphviz_dirdump_graphviz_include_videorOrQenable_select_tf_opssetforce_select_tf_opsconversion_summary_dirselect_user_tf_opsrallow_all_select_tf_ops enable_tflite_resource_variablesunfold_batchmatmullower_tensor_list_ops*default_to_single_batch_in_tensor_list_opsrEaccumulation_typeallow_bfloat16unfold_large_splat_constantsupported_backends disable_per_channel_quantization#enable_mlir_dynamic_range_quantizerenable_dynamic_update_slicepreserve_assert_opguarantee_all_funcs_one_usetf_quantization_modedisable_infer_tensor_rangeuse_fake_quant_num_bits)"rrrrZdefault_ranges_statsrrrrrrZdump_graphviz_videoZ target_opsrrrrrrrrrrrrrrrrrrrrrrrrbuild_conversion_flagssvz     rcKstfi|}tfi|}|dd}|dd}t|D]Z\} \} } |j} t|rx|rp|| \| _| _nt d| | _ | j j ttt| q<|r|D]} |j| q|r|D]} |j| qt|||d|d} | S)a"Convert a frozen GraphDef that can't be loaded in TF. Conversion can be customized by providing arguments that are forwarded to `build_model_flags` and `build_conversion_flags` (see documentation). Args: input_data: Input data (i.e. often `sess.graph_def`), input_arrays_with_shape: Tuple of strings representing input tensor names and list of integers representing input shapes (e.g., [("foo" : [1, 16, 16, 3])]). Use only when graph cannot be loaded into TensorFlow and when `input_tensors` is None. output_arrays: List of output tensors to freeze graph with. Use only when graph cannot be loaded into TensorFlow and when `output_tensors` is None. control_output_arrays: Control output node names. This is used when converting a Graph with no output tensors. For example, if the graph's last operation is a Print op, just specify that op's name in this field. This can be used together with the `output_arrays` parameter. **kwargs: See `build_model_flags` and `build_conversion_flags`. Returns: The converted data. For example if TFLite was the destination, then this will be a tflite flatbuffer in a bytes array. Raises: Defined in `build_conversion_flags`. roTquantized_input_statsNzThe `quantized_input_stats` flag must be defined when either `inference_type` flag or `inference_input_type` flag is set to tf.int8 or tf.uint8.rnro)rrr@ enumerate input_arraysaddr mean_value std_valuerArshapedimsrr]mapint output_arraysappendcontrol_output_arraysrqSerializeToString) input_dataZinput_arrays_with_shaperrkwargsrrroridxrr input_arraydatarrrconvert_graphdef_with_arrayss:    rcKstfi|}tfi|}|dd}|dd}|dd}|dd} |dd} t|D]\} } |j} |r| j| _n t| | _t | j dd | _ t |r| r| | \| _ | _n td |dur| j}n|| }|jdurLg}|D]B}|dust|tjr$|jdur$|d q|t|q| jj|d | j_q`d| j_q`|D].}|rt|j|jn|jt|qZt|||| r| nd|d }|S)aConvert a frozen GraphDef model using the TF Lite converter. Conversion can be customized by providing arguments that are forwarded to `build_model_flags` and `build_conversion_flags` (see documentation). Args: input_data: Input data (i.e. often `sess.graph_def`), input_tensors: List of input tensors. Type and shape are computed using `foo.shape` and `foo.dtype`. output_tensors: List of output tensors (only .name is used from this). **kwargs: See `build_model_flags` and `build_conversion_flags`. Returns: The converted data. For example if TFLite was the destination, then this will be a tflite flatbuffer in a bytes array. Raises: Defined in `build_conversion_flags`. rN input_shapesroTr debug_infoz"input type of the TensorFlow modelrzTStatistics for quantized inputs were expected, but not specified; continuing anyway.Fr)rrr@rrrrrZget_tensor_namerEdtype data_typerrrwarningswarnrrankrHr DimensionrVrrrr unknown_rankrrqr)r input_tensorsoutput_tensorsrrrrrrorrr input_tensorrrrdim output_tensorrrrrconvert_graphdefs`                rcKs:tfi|}tfi|}t||dddd}|S)z.Converts a SavedModel using TF Lite converter.NTrdrnro)rrrqr)rrrrrrrconvert_saved_model#src Ksnt}d|_|rtjj|_n tjj|_|D]}|j}||_q,t fi|}t | | |ddd}|S)z6Converts a Jax hlo-based model using TFLite converter.TNr) rrZuse_hlo_importZ HLO_PROTOZ hlo_file_typeZHLO_TEXTrrrrrqr) Z input_content input_namesZis_proto_formatrr input_namerrrrrrconvert_jax_hlo2s"   rzlite.toco_convert)v1z#Use `lite.TFLiteConverter` instead.cOs*|dd|d<t|||g|Ri|S)a Convert a TensorFlow GraphDef to TFLite. This function is deprecated. Please use `tf.lite.TFLiteConverter` API instead. Conversion can be customized by providing arguments that are forwarded to `build_model_flags` and `build_conversion_flags` (see documentation for details). Args: input_data: Input data (i.e. often `sess.graph_def`). input_tensors: List of input tensors. Type and shape are computed using `foo.shape` and `foo.dtype`. output_tensors: List of output tensors (only .name is used from this). *args: See `build_model_flags` and `build_conversion_flags`. **kwargs: See `build_model_flags` and `build_conversion_flags`. Returns: The converted TensorFlow Lite model in a bytes array. Raises: Defined in `convert`. roF)r@r)rrrargsrrrr toco_convertLsrcst|t}jD] }t}|jD]`}|jdur8q(t|jD]D\}}|jdurn|t|jkrn|j|rnqB|j |j r|qB| |qBq(|jD]D}|j dur|j D]}| |q|jdur|jD]}| |qq|jdur|jD]}| |q|j dur|j D]}| |q|D]} | |j | jqq|D]T} | dksj| jdustj| jtsj| jjdkrB| | qBt|}t|fdddi} t|D]\}} | | vr֐qj| } ||ddD]T}|| vrqj|}| jj|jjkr$q| jj|jjkr:q| | |<qqjD]X}|jD]J}|jdurnqZ|jD],}|j |j} | | vrt| | |j |_qtqZqP| D]}dj|_qtS)a"Generates a new model byte array after deduplicating readonly buffers. This function should be invoked after the model optimization toolkit. The model optimization toolkit assumes that each tensor object owns its each buffer separately. Args: tflite_model: TFLite flatbuffer in a byte array to be deduplicated. Returns: TFLite flatbuffer in a bytes array, processed with the deduplication method. Nrcsj|jjSrT)buffersrtobytes)rmodelrrr\z.deduplicate_readonly_buffers..)keyr)r Zconvert_bytearray_to_objectr subgraphs operatorsinputsrZmutatingVariableInputslentensorsZ isVariableroutputsdiscard intermediatesbuffercopyrrrHr]sizesortedZconvert_object_to_bytearray)Z tflite_modelZread_only_buffer_indicessubgraphZread_only_input_tensor_indicesopiZinput_tensor_idxZoutput_tensor_idxZintermediate_tensor_idx tensor_idx buffer_idxZduplicate_buffer_mapZ buffer_i_idxZbuffer_iZ buffer_j_idxZbuffer_jrrrrrdeduplicate_readonly_buffershs                             r)r)r)NT)N)FFNrNN)Xrbdistutils.spawnrzenumosrplatformr subprocessrtempfiler}rrJtensorflow.lite.pythonrrrZ$tensorflow.lite.python.convert_phaserrrrZ&tensorflow.lite.python.metrics.wrapperr rjtensorflow.lite.tocor rr rr rZtensorflow.lite.toolsr tensorflow.python.frameworkrrtensorflow.python.platformrZ_resource_loadertensorflow.python.utilr tensorflow.python.util.tf_exportr _tf_exportrr:rDTyperU IODataTyperErG"EXPERIMENTAL_USE_TOCO_API_DIRECTLYrhget_path_to_datafilepathexistsrNEnumrOZOPTIMIZE_TFLITE_MODELZQUANTIZErr$reZSPARSIFYrfrgrqZCONVERT_TF_TO_TFLITE_MODELZ+CONVERT_GRAPHDEF_USING_DEPRECATED_CONVERTERrkrTENSORFLOW_GRAPHDEFTFLITErZCONVERT_GRAPHDEFrrZCONVERT_SAVED_MODELrZCONVERT_JAX_HLOr deprecatedrrrrrrs                   +  (  ,   / m , 6 @ Q