U db@sddlZddlZddlmZddlmZddlZddlmZddlZddl Z ddl m Z ddl Z ddlZddlZddlmZddlmZmZddlmZddlmmZeeZejZej Z!ej"Z#ej$Z%ej&Z'ej(Z)ej*Z+ej,Z-ej.Z/ej0Z1ej2Z3ej4Z5ej6Z7ej8Z9ej:Z;ejZ?ej@Z@ejAZAejBZBejCZCejDZDeBrtejEZFejGZHejGZIejJZKejLZMd d ZNejOZPejQZRnd d ZHd d ZKdd ZMeCrejSZFejTZIejUZVdd ZNejOZPejQZReDsejEZFdd ZIdd ZPdd ZNdd ZRejWZXejYZZej[Z\ej]Z^ddZ_ddZ`d|ddZaddZbd}ddZcd~dd Zdd!d"Zed#d$Zfd%d&Zgd'd(Zhd)d*Zid+d,Zjd-d.Zkd/d0Zldd2d3Zmd4d5Znd6d7Zodd8d9Zpd:d;Zqdd?Zsd@dAZtddBdCZudDdEZvdFdGZwdHdIZxeydJdKdLdMgZzdNdOZ{dPdQZ|dRdSZ}e}ej~_dTdUZdVdWZdXdYZGdZd[d[eZeZdad\Zd]ad^d_Zejd`daZddbdcZdddeZdfdgZdhdiZdjdkZdldmZddndoZdpdqZddrdsZdtduZej~Z~ee~_ee~_ee~_ee~_ddvdwZeej_dxdyZeej_dzd{Zeej_dS)N)Message)Process) defaultdict) basestring) caffe2_pb2)scopeutils)TriggerLazyImportcCsttSN)npasarrayCZget_cuda_peer_access_patternrr;/tmp/pip-unpacked-wheel-ua33x9lu/caffe2/python/workspace.pyGetGpuPeerAccessPattern=srcCsdSNrrrrrrDrcCsdSrrrrrrrFrcCsdSrrrrrrrHrcCsttSr )r r r Zget_hip_peer_access_patternrrrrrPscCsdSrrrrrrrZrcCsdSr r)xrrrr[rcCs tgSr )r arrayrrrrr\rcCsdSr rrrrrr^rcCst|||dSr )r Zfill_random_network_inputsProtoSerializeToString)netZ input_dimsZ input_typesrrrFillRandomNetworkInputsfsrcCsTttjtj}|d}|dkr&dSt}|d|d}||SdS)zGet a free flask port.)z 127.0.0.1rr)rN)socketAF_INET SOCK_STREAM connect_exbind getsocknameclose)sockresultsportrrr_GetFreeFlaskPortjs   r(cCsfddlm}|dkrt}|dkr*t}t|jdt|d|gfd}|t d t ||S)zStart a mint instance. TODO(Yangqing): this does not work well under ipython yet. According to https://github.com/ipython/ipython/issues/5862 writing up some fix is a todo item. r)appNz-pz-r)targetargszMint running at http://{}:{}) Zcaffe2.python.mintr)r root_folderr(rmainstrstartprintformatrgetfqdn)r,r'r)processrrr StartMint|s r4cCsLt|tr|St|tr |St|dr6|Stdt|jdS)aStringify a protocol buffer object. Inputs: obj: a protocol buffer object, or a Pycaffe2 object that has a Proto() function. Outputs: string: the output protobuf string. Raises: AttributeError: if the passed in object does not have the right attribute. rz.Unexpected argument to StringifyProto of type N) isinstancerrrhasattrr ValueErrortype__name__)objrrrStringifyProtos    r;cCs:|dkrttStj|s,t|t|SdSr )r Zreset_workspacer,ospathexistsmakedirs)r,rrrResetWorkspaces   r@FcCsFt|dkrg}|D]}t|qttjtjjjt|t ||Sr ) r r create_blobCallWithExceptionIntercept create_net Workspacecurrent_last_failed_op_net_position GetNetNamer;)r overwriteZ input_blobsZ input_blobrrr CreateNets rIcCstt|t|Sr )r Predictorr;)init_netZ predict_netrrrrJsrJcCstt||Sr )r Zget_operator_costr;)operatorblobsrrrGetOperatorCostsrNcCstt|Sr )r Zrun_operator_oncer;)rLrrrRunOperatorOncesrOcCstt||Sr )r Zrun_operator_multipler;)rLZnum_runsrrrRunOperatorMultiplesrPcCs |D]}t|}|sdSqdS)NFT)rO) operatorsopsuccessrrrRunOperatorsOnces rTcCstSr )r Zclear_global_net_observerrrrrClearGlobalNetObserversrUc Osz |||WStk r|}t|d}td|||r||kr||}t|D]$}td|d|d|dq\YnXdS)NzgOriginal python traceback for operator `{}` in network `{}` in exception above (most recent call last):z File "{}", line {}, in {}rr) Exceptionoperator_tracebacksgetloggerwarningr1reversed) funcZ op_id_fetcherZnet_namer+kwargsZop_idZnet_tracebackstblinerrrrBs(    rBcCsttjtjjjt|t|Sr )rBr Z run_net_oncerDrErFrGr;rrrr RunNetOnces rbrcCs"ttjtjjjt|t|||S)aRuns a given net. Inputs: name: the name of the net, or a reference to the net. num_iter: number of iterations to run allow_fail: if True, does not assert on net exec failure but returns False Returns: True or an exception. )rBr Zrun_netrDrErFrGStringifyNetName)nameZnum_iterZ allow_failrrrRunNets recCs6ddlmm}t||jr(||}tt|Sr) caffe2.python.corepythoncorer5 ExecutionStepPlanr Zrun_planr;Z plan_or_steprhrrrRunPlans  rlcCs6ddlmm}t||jr(||}tt|Sr) rfrgrhr5rirjr Zrun_plan_in_backgroundr;rkrrrRunPlanInBackgrounds  rmc Cs|rdd|D}ndd|D}|dkrB|dks6tt|}n$|dkrXt||}nt|||}t}||i}i}|jD]&} | jst | j || j <| j || j <q||fS)aInfers the shapes and types for the specified nets. Inputs: nets: the list of nets blob_dimensions (optional): a dictionary of blobs and their dimensions. If not specified, the workspace blobs are used. nets_proto (optional): a boolean flag indicating whether the protobuffer representation is passed to the routine. Returns: A tuple of (shapes, types) dictionaries keyed by blob name. cSsg|] }t|qSr)r;.0nrrr ,sz'InferShapesAndTypes..cSsg|]}t|qSr)r;rrnrrrrq.sN) AssertionErrorr Z%infer_shapes_and_types_from_workspaceZinfer_shapes_and_types_from_maprZ TensorShapesParseFromStringshapesZ unknown_shapelistZdimsrdZ data_type) netsZblob_dimensionsZ nets_protoZ blob_typesZ net_protosZblobdesc_prototxtZblobdesc_protorttypestsrrrInferShapesAndTypess2     rycCs0t|tr|St|j|ks(td|t|S)NzExpected a string or %s)r5rr8r9rrr.)rdZ expected_typerrr_StringifyNameFs  rzcCs t|dS)NZ BlobReferencerzrdrrrStringifyBlobNameNsr}cCs t|dS)NNetr{r|rrrrcRsrccCsZt|tr|St|jdks*t|jdkr2|St|tjrD|jStd t |dS)Nr~ZNetWithShapeInferencezNot a Net object: {}) r5rr8r9NamerNetDefrdrWr1r.rarrrrGVs  rGcCstjj}t||||S)aNFeeds a blob into the workspace. Inputs: name: the name of the blob. arr: either a TensorProto object or a numpy array object to be fed into the workspace. device_option (optional): the device option to feed the data with. Returns: True or False, stating whether the feed is successful. )r rDrE_Workspace_feed_blob)rdarr device_optionwsrrrFeedBlob`s rcCsdd|DS)zFetches a list of blobs from the workspace. Inputs: names: list of names of blobs - strings or BlobReferences Returns: list of fetched blobs cSsg|] }t|qSr FetchBlob)rordrrrrqwszFetchBlobs..r)namesrrr FetchBlobsosrcCs.tt|}t|tr*tdt||S)zFetches a blob from the workspace. Inputs: name: the name of the blob - a string or a BlobReference Returns: Fetched blob (numpy array or string) if successful z'Use FetchInt8Blob to fetch Int8 Blob {})r fetch_blobr}r5tuple TypeErrorr1rdr%rrrrzs rcCstjj}|j|Sr )r rDrErMto_torch)rdrrrr FetchTorchsr Int8Tensordatascale zero_pointcCs2tt|}t|ts*tdt|t|S)mFetches an Int8 blob from the workspace. It shared backend implementation with FetchBlob but it is recommended when fetching Int8 Blobs Inputs: name: the name of the Int8 blob - a string or a BlobReference Returns: data: int8 numpy array, data scale: float, fake quantization scale zero_point: int, fake quantization offset 9You are not fetching an Int8Blob {}. Please use FetchBlob)r rr}r5rrrr1rrrrr FetchInt8Blobs  rcCsXtt|}t|ts*tdt|t|}|j t j t |j  t j|jS)zFetches an Int8 blob from the workspace and return its real value representation. Inputs: name: the name of the Int8 blob - a string or a BlobReference Returns: real value representation of int8 numpy array r)r rr}r5rrrr1rrastyper Zint32intrZfloat32r)rdr%Z int8_blobrrrFetchInt8BlobRealVals rcCs.||}t|ts&tdt|t|S)rz:You are not fetching an Int8Blob {}. Please use fetch_blob)rr5rrrr1r}r)rrdr%rrr_Workspace_fetch_int8_blobs  rcCs0t}tt|d|}|||S)aApply a Transform to a NetDef protobuf object, and returns the new transformed NetDef. Inputs: transform_key: the name of the transform, as it is stored in the registry net: a NetDef protobuf object Returns: Transformed NetDef protobuf object. utf-8)rrr Zapply_transformr.encoderrs) transform_keyrtransformed_nettransformed_strrrrApplyTransforms   rc Ks|d|kr|dnd}d|kr$|dnd}d|kr8|dnd}t}tt|d||||t|}|||S)aApply a Transform to a NetDef protobuf object, and returns the new transformed NetDef, only if it runs faster than the original. The runs are performed on the current active workspace (gWorkspace). You should initialize that workspace before making a call to this function. Inputs: transform_key: the name of the transform, as it is stored in the registry net: a NetDef protobuf object init_net: The net to initialize the workspace. warmup_runs (optional): Determines how many times the net is run before testing. Will be 5 by default. main_runs (optional): Determines how many times the net is run during testing. Will be 10 by default. improvement_threshold (optional): Determines the factor which the new net needs to be faster in order to replace the old. Will be 1.01 by default. Returns: Either a Transformed NetDef protobuf object, or the original netdef. warmup_runs main_runs improvement_thresholdg)\(?r) rrr Zapply_transform_if_fasterr.rrfloatrs) rrrKr^rrrrrrrrApplyTransformIfFasters"   rcCstS)z>Return the current namescope string. To be used to fetch blobs)rZCurrentNameScoperrrr GetNameScope src@s8eZdZdZddZddZddZdd Zd d Zd S) _BlobDictz>Provides python dict compatible way to do fetching and feedingcCst|Sr r)selfkeyrrr __getitem__sz_BlobDict.__getitem__cCs t||Sr )r)rrvaluerrr __setitem__sz_BlobDict.__setitem__cCs ttSr )lenr rMrrrr__len__sz_BlobDict.__len__cCs tSr )r rM__iter__rrrrrsz_BlobDict.__iter__cCs t|Sr )r has_blob)ritemrrr __contains__ sz_BlobDict.__contains__N) r9 __module__ __qualname____doc__rrrrrrrrrrs rZ_CAFFE2_IMMEDIATErcCstSr )_immediate_moderrrr IsImmediate@srccs"t}t|ddVt|dS)NT)CurrentWorkspaceSwitchWorkspace)Zworkspace_namerErrrWorkspaceGuardDs rc CsHtr tdatttattW5QRX|r|jjdkr>| tj }|dkrNt }|r|j tjkr|jtdkrtdd|dt|jt|}|dk r||||S|||SdS)NZSUZfloat64zeCUDA operators do not support 64-bit doubles, please use arr.astype(np.float32) or np.int32 for ints.z Blob: {}z type: {})r8rZ TensorProtorZCaffe2TensorToNumpyArrayr ZndarrayZdtypekindrobjectrZCurrentDeviceScopeZ device_typeCUDArZr[r1r.r}rAfeed)rrdrrrrrrs&  rcCs|t|dSr )Z _remove_blobr.)rblobrrr_Workspace_remove_blobsrcCsvt|jdkrZt|jdkrZddl}t||jrZ|dks@td|j|}| |dS|dk rjt |}| ||S)NTensortorchrz5device_option doesn't make sense with PyTorch tensorsT) r8r9rrr5rrr_C_tensor_impl_raw_handle_wrap_tensor_implr;Z_feed)rargrrhandlerrr _Blob_feeds    rcCsddl}|}|j|S)z PyTorch tensor interop (TensorCPU methods) Can be accessed as: workspace.Workspace.current.blobs['foo'].tensor().to_torch() rN)rrrr)Ztensorrrrrr_Tensor_to_torchsrcCs|std|S)NzBlob has to be a tensor)Z is_tensor RuntimeErrorZ as_tensorr)rrrr_Blob_to_torchsr)NN)N)FN)rF)NFN)N)F)F)N)N) collections contextlibZgoogle.protobuf.messagermultiprocessingrr<rloggingZnumpyr Z past.builtinsrrrrZ caffe2.protorZ caffe2.pythonrrZcaffe2.python.lazyr Z!caffe2.python._import_c_extensionrgZ_import_c_extensionr getLoggerr9rZrMrZ reset_blobZ ResetBlobrAZ CreateBlobZcurrent_workspacerZdeserialize_blobZDeserializeBlobZ global_initZ GlobalInitrZHasBlobZregistered_operatorsZRegisteredOperatorsZserialize_blobZ SerializeBlobZswitch_workspacerr,Z RootFolderZ workspacesZ WorkspacesZ benchmark_netZ BenchmarkNetZbenchmark_net_onceZBenchmarkNetOnceZ get_statsZGetStatsZcreate_offline_tensorZCreateOfflineTensordictrXZis_asanZ has_fbgemmZhas_cuda_supportZhas_hip_supportZhas_gpu_supportrZ GpuDeviceTypeZnum_cuda_devicesZNumCudaDevicesZ NumGpuDevicesZget_cuda_versionZGetCUDAVersionZget_cudnn_versionZGetCuDNNVersionrZget_device_propertiesZGetDevicePropertiesZget_gpu_memory_infoZGetGPUMemoryInfoZHIPZnum_hip_devicesZget_hip_versionZ GetHIPVersionZis_numa_enabledZ IsNUMAEnabledZget_num_numa_nodesZGetNumNUMANodesZget_blob_numa_nodeZGetBlobNUMANodeZget_blob_size_bytesZGetBlobSizeBytesrr(r4r;r@rIrJrNrOrPrTrUrBrbrerlrmryrzr}rcrGrrrr namedtuplerrrrrDZfetch_int8_blobrrrrrrrrrcontextmanagerrrrrrrrrrrrrCrunZ feed_blobZ remove_blobrZBlobrrZ TensorCPUrrrrrrs           (  +  2