BVhmw dZddlZddlZddlZddlZddlZddlmZddlmZddlm Z ddlm Z GddejZ Gd d ejZ e d Ze d ZGd dej ZGddej$ZGddej(ZdZdZdZ d-dZdZdZdZdZdZdZdZ dZ!dZ"d Z# d-d!Z$ d-d"Z%d#Z&d$Z'd%Z(d&Z)d'Z*d(Z+d)Z,d*Z-d+Z.d,Z/y).zBUpgrader for Python scripts from 1.* TensorFlow to 2.0 TensorFlow.N)all_renames_v2) ast_edits)module_deprecations_v2) reorders_v2ceZdZdZy)UnaliasedTFImportc<tj|_d|_y)NzThe tf_upgrade_v2 script detected an unaliased `import tensorflow`. The script can only run when importing with `import tensorflow as tf`.)rERROR log_level log_messageselfs \/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/tools/compatibility/tf_upgrade_v2.py__init__zUnaliasedTFImport.__init__#s__DNDDN__name__ __module__ __qualname__rrrrr!sErrceZdZdZy)VersionedTFImportcHtj|_d|zdz|_y)Nz*Not upgrading symbols because `tensorflow.z ` was directly imported as `tf`.)rINFOr r )rversions rrzVersionedTFImport.__init__,s"^^DNDwN:;DrNrrrrrr*sd8dd?id@d>d8dd?dAdBdCdDdEd'didFd'didGd'didHd'didId'didJd7d8idKdLdCidMdNdCidOdNdCidPd'didQd7d8idRdSdTdUdVdSdTdUdWdXdYidZdXdYiid[dd\d]d^dd\d]d_d`didad`didbdd8dcdddd8dcdedd8dcdfdd8dcdgdhdiidjdkd>idldkd>idmdkd>idndidoidpdidoidqdrdsdtdud>dvdwdxdydBdCdzid{d|dBid}d~d>idd~d>idd>dBddd>dBdddBdCddddidddidddidddidddidddidddidddidddiddddddddidddiddidoiddd>iddd>iddd>iddd>iddd8dddd8dddd8dddd8dddd8dddd8dddd8dddd8dddd8dddd8dddd8diddd8dddd8dddd8dddd8dddd8ddd8dddd8dddddidd7d8idd>ddddBdddddd>dBddddddddidddddƜddddddddddʜdddddiddiddiddidќ|_tj|jtj|_||_|j r!dt jdgdԢիi|_ni|_i|_ hd֣|_ gdעgdآgd٢gdڢgdۢgdڢdܜ|_ ttj|_|jj!|jt j"df}t j"df}t j"df}t j$df}t j"df}t j"df}t j"df} t j$df} t j&df} t j"df} t j$df} t j$df}t j$df}t j&df}t j&df}t j&df}t j$df}d}t j&d|zf}t j$d|zf}t j&d|zf}t j&d|zf}t j&d|zf}t j&d|zf}t j$df}t j$df}id|d|d|d|d|d|d|d|d|d|d|d|d|d|d|d|d|id|d |d |d |d | d| d|d| d |d| d|d| d|d| d|d|d|id|d|d|d|d|d|d|d|d|d|d|d|d |d!|d"|d#|d$|id%|d&|d'|d(|d)|d*|d| d+| d,| d-| d.| d/|d0|d1|d2|d3|d4|id5|d6|d7|d8|d9|d:|d;|d<|d=|d>|d?|d@|dA|dB|dC|dD|dE|idF|dG|dH|dI|dJ|dK|dL|dM|dN|dO|dP|dQ|dR|dS| dT| dU| dV| idW| dX| dY| dZ| d[| d\| d]| d^| d_| d`| da| db| dc|dd|de|df|dg|idh|di|dj|dk|dl|dm|dn|do|dp|dq|dr|ds|dt|du|dv|dw|dx|idy|dz|d{|d||d}|d~|d|d|d|d|d|dt j$dfdt j&dfd|d|d|d|||| | | | | | | | | | | ||d|_tj|j(tj*j-D]7\}}t j$dj/||f}||j(|<9iddt j$dfiddt j$dfiddt j$dfiddt j$dfid dt j"dfid dt j"dfiddt j"dfiddt j"dfiddt j$dfidd|iddt j$dfidt j$df|ddǐd|idt j$df|ddd|id#dt j$dfiddt j$dfidt j$dfidt j$dfidt j$dfid|_tj|j0idt2dt2dt4dt6dt6dt6dt6dt6dt6dt6d3t8dt:dt<dt<dt<dt<dt>idt>dt@dtBdtEjFtHdtJdd«d}tEjFtHdtLdĐdŐ«dtEjFtHdtLdĐdƐ«dtEjFtHdtLdĐdɐ«dtEjFtHdtNdĐd̐«dytEjFtPdtSjTdΫϫdAtEjFtPdtSjTdΫϫdtEjFtVdѐҫdtXdtXdtXdtXdtEjFtVdԐҫdtXtZt\t^t^t`tbtbtZt\t^t^t`tEjFtPdtSjTd֫ϫdל |_2tj|jdtfjh|_5y(Ntf.test.assert_equal_graph_def) checkpoint_v2hash_table_shared_nameztf.autograph.to_code) arg_types arg_values indentationztf.autograph.to_graph)r8r9tf.nn.embedding_lookupvalidate_indices&tf.image.sample_distorted_bounding_boxseed2 tf.gradientscolocate_gradients_with_ops tf.hessiansz *.minimizez*.compute_gradientstf.condtrue_fnfalse_fn)strictfn1fn2 tf.argmin dimensionaxis tf.argmaxz tf.arg_minz tf.arg_maxztf.math.argminztf.math.argmaxztf.image.crop_and_resizebox_ind box_indicesztf.extract_image_patchesksizessizesztf.image.extract_image_patchesztf.image.resize align_cornersztf.image.resize_imagesztf.expand_dimsdimtf.batch_to_space block_size block_shapeztf.space_to_batchtf.nn.space_to_batchz tf.constant verify_shapeverify_shape_is_now_always_truetf.convert_to_tensorpreferred_dtype dtype_hint'tf.nn.softmax_cross_entropy_with_logitsz*tf.nn.softmax_cross_entropy_with_logits_v2ztf.linalg.l2_normalizetf.linalg.norm keep_dimskeepdimstf.normztf.load_file_system_librarylibrary_filenamelibrary_locationztf.count_nonzeroinput) input_tensorr]reduction_indicesztf.math.count_nonzeroztf.nn.erosion2dfilters dilations)kernelratesztf.math.l2_normalizeztf.math.log_softmaxztf.math.softmaxztf.nn.l2_normalizeztf.nn.log_softmax tf.nn.moments tf.nn.pool dilation_ratetf.nn.separable_conv2dratetf.nn.depthwise_conv2dz tf.nn.softmaxztf.nn.sufficient_statisticstf.debugging.assert_all_finitexmessage)tmsgztf.verify_tensor_all_finite tf.sparse.addthresh threshold tf.sparse_addtf.sparse.concatexpand_nonconcat_dims) concat_dimexpand_nonconcat_dimtf.sparse_concatztf.sparse.split split_dimtf.sparse_splittf.sparse.reduce_max)reduction_axesr]tf.sparse_reduce_maxztf.sparse.reduce_sumztf.sparse_reduce_sumztf.nn.max_pool_with_argmaxTargmax output_dtypetf.nn.max_poolvaluetf.nn.avg_poolztf.nn.avg_pool2dtf.multinomialdtypetf.random.multinomialtf.reverse_sequenceseq_axis batch_axis)seq_dim batch_dimz*tf.nn.batch_norm_with_global_normalizationmeanvariance)rrmvztf.nn.dilation2d)filterrhz tf.nn.conv3drz tf.zeros_liketensorz tf.ones_likeztf.nn.conv2d_transpose)rrztf.nn.conv3d_transposetf.nn.convolution)rrkztf.gfile.Existsfilenamepathztf.gfile.Removez tf.gfile.Statz tf.gfile.Globpatternztf.gfile.MkDirdirnameztf.gfile.MakeDirsztf.gfile.DeleteRecursivelyztf.gfile.IsDirectoryztf.gfile.ListDirectoryz tf.gfile.Copysrcdst)oldpathnewpathztf.gfile.Rename)oldnamenewnamez tf.gfile.Walkin_ordertopdownztf.random.stateless_multinomialztf.string_to_number string_tensorztf.strings.to_numberztf.string_to_hash_bucketztf.strings.to_hash_bucket tf.reduce_all)rdr]tf.math.reduce_all tf.reduce_anytf.math.reduce_any tf.reduce_mintf.math.reduce_min tf.reduce_maxtf.math.reduce_max tf.reduce_sumtf.math.reduce_sumtf.reduce_meantf.math.reduce_meantf.reduce_prodtf.math.reduce_prodtf.reduce_logsumexptf.math.reduce_logsumexptf.reduce_join)r]rdtf.strings.reduce_joinz tf.squeeze squeeze_dimstf.nn.weighted_moments tf.nn.conv1d)ruse_cudnn_on_gpu tf.nn.conv2d)rrtf.nn.conv2d_backprop_input output_shape)r input_sizes out_backproprtf.contrib.summary.audiodata)rfamily%tf.contrib.summary.create_file_writerr+tf.contrib.summary.generictag)r+rrtf.contrib.summary.histogramtf.contrib.summary.image max_outputs)r bad_color max_imagesrtargetslabelsbytes input_bytescheckpoint_dirckpt_dir_or_file)tf.contrib.summary.scalarz(tf.nn.weighted_cross_entropy_with_logitsz tf.decode_rawztf.io.decode_rawz"tf.contrib.framework.load_variablerr")ztensorflow.contribztensorflow.flagsr r"ztensorflow.google)excluded_prefixes>`tf.padtf.sizetf.shapetf.tuple tf.substr tf.nn.crelu tf.gather_nd tf.transposetf.nn.in_top_ktf.quantize_v2tf.boolean_masktf.io.decode_csvtf.math.in_top_ktf.parse_exampletf.sparse_matmultf.random.poissontf.strings.lengthtf.strings.substrtf.confusion_matrixtf.io.parse_exampletf.nn.depth_to_spacetf.nn.space_to_depthtf.sparse.segment_sumtf.io.serialize_sparsetf.sparse.segment_meantf.parse_single_exampletf.math.confusion_matrixtf.sparse.segment_sqrt_ntf.io.parse_single_exampletf.io.serialize_many_sparse$tf.data.experimental.TensorStructure)tf.data.experimental.TensorArrayStructure*tf.data.experimental.RaggedTensorStructure*tf.data.experimental.SparseTensorStructure9tf.feature_column.categorical_column_with_vocabulary_filerBr_rKrHrjrAr?rrrirrrrrrtrw tf.while_loopr\rrrrrrr~rxr|rRrrrrrrrrrrrXrUrrrrnr;rlrrrtf.nn.fractional_avg_pooltf.nn.fractional_max_poolrtf.nn.ctc_beam_search_decodertf.nn.embedding_lookup_sparseror5#tf.uniform_unit_scaling_initializer$tf.initializers.uniform_unit_scalingr=r[)r+r sample_raterrstep)logdir max_queue flush_millisfilename_suffixr+)r+rmetadatarr)r+rrr)r+rrrrr)rrrrrrzvTo use learning rate decay schedules with TensorFlow 2.0, switch to the schedules in `tf.keras.optimizers.schedules`. z has been changed to return None, the data argument has been removed, and arguments have been reordered. The calls have been converted to compat.v1 for safety (even though they may already have been correct).z has been changed to return None, and the data and summarize arguments have been removed. The calls have been converted to compat.v1 for safety (even though they may already have been correct).ak(Manual edit required) `tf.contrib.layers.layer_norm` has been deprecated, and its implementation has been integrated with `tf.keras.layers.LayerNormalization` in TensorFlow 2.0. Note that, the default value of `epsilon` is changed to `1e-3` in the new API from `1e-12`, and this may introduce numerical differences. Please check the new API and use that instead.zInitializers no longer have the dtype argument in the constructor or partition_info argument in the __call__ method. The calls have been converted to compat.v1 for safety (even though they may already have been correct).ztf.metrics have been replaced with object oriented versions in TF 2.0 and after. The metric function calls have been converted to compat.v1 for backward compatibility. Please update these calls to the TF 2.0 versions.ztf.losses have been replaced with object oriented versions in TF 2.0 and after. The loss function calls have been converted to compat.v1 for backward compatibility. Please update these calls to the TF 2.0 versions.zh`partition_strategy` has been removed from . The 'div' strategy will be used by default.zuniform_unit_scaling_initializer has been removed. Please use tf.initializers.variance_scaling instead with distribution=uniform to get equivalent behaviour.aThe TF 1.x summary API cannot be automatically migrated to TF 2.0, so symbols have been converted to tf.compat.v1.summary.* and must be migrated manually. Typical usage will only require changes to the summary writing logic, not to individual calls like scalar(). For examples of the new summary API, see the Effective TF 2.0 migration document or check the TF 2.0 TensorBoard tutorials.aHtf.contrib.summary.* functions have been migrated best-effort to tf.compat.v2.summary.* equivalents where possible, but the resulting code is not guaranteed to work, so please check carefully. For more information about the new summary API, see the Effective TF 2.0 migration document or check the updated TensorBoard tutorials.z replacement does not accept a 'family' argument; instead regular name scoping should be used. This call site specifies a family argument that has been removed on conversion, so the emitted tag names may be incorrect without manual editing.a3tf.contrib.summary.create_file_writer() has been ported to the new tf.compat.v2.summary.create_file_writer(), which no longer re-uses existing event files for the same logdir; instead it always opens a new writer/file. The python writer objects must be re-used explicitly if the reusing behavior is desired.ak(Manual edit required) tf.contrib.summary.record_summaries_every_n_global_steps(n, step) should be replaced by a call to tf.compat.v2.summary.record_if() with the argument `lambda: tf.math.equal(0, global_step % n)` (or in graph mode, the lambda body can be used directly). If no global step was passed, instead use tf.compat.v1.train.get_or_create_global_step().a(Manual edit required) tf.contrib.summary.graph() has no direct equivalent in TF 2.0 because manual graph construction has been superseded by use of tf.function. To log tf.function execution graphs to the summary writer, use the new tf.compat.v2.summary.trace_* functions instead.a&(Manual edit required) tf.contrib.summary.import_event() has no direct equivalent in TF 2.0. For a similar experimental feature, try tf.compat.v2.summary.experimental.write_raw_pb() which also accepts serialized summary protocol buffer input, but for tf.Summary protobufs rather than tf.Events.z(This warning is only applicable if the code saves a tf.Keras model) Keras model.save now saves to the Tensorflow SavedModel format by default, instead of HDF5. To continue saving to HDF5, add the argument save_format='h5' to the save() function.aIf you're using the strategy with a custom training loop, note the following changes in methods: make_dataset_iterator->experimental_distribute_dataset, experimental_make_numpy_iterator->experimental_make_numpy_dataset, extended.call_for_each_replica->run, reduce requires an axis argument, unwrap->experimental_local_results experimental_initialize and experimental_finalize no longer needed aS(Manual edit required) tf.contrib.distribute.MirroredStrategy has been migrated to tf.distribute.MirroredStrategy. Things to note: Constructor arguments have changed. If you are using MirroredStrategy with Keras training framework, the input provided to `model.fit` will be assumed to have global batch size and split across the replicas. zN(Manual edit may be required) tf.distribute.MirroredStrategy API has changed. zu(Manual edit required) tf.contrib.distribute.OneDeviceStrategy has been migrated to tf.distribute.OneDeviceStrategy. z(Manual edit required) tf.contrib.distribute.TPUStrategy has been migrated to tf.distribute.TPUStrategy. Note the slight changes in constructor. z(Manual edit required) tf.contrib.distribute.CollectiveAllReduceStrategy has been migrated to tf.distribute.experimental.MultiWorkerMirroredStrategy. Note the changes in constructor. a(Manual edit required) tf.contrib.distribute.ParameterServerStrategy has been migrated to tf.compat.v1.distribute.experimental.ParameterServerStrategy (multi machine) and tf.distribute.experimental.CentralStorageStrategy (one machine). Note the changes in constructors. ztf.keras.experimental.export_saved_model and tf.keras.experimental.load_from_saved_model have been deprecated.Please use model.save(path, save_format='tf') (or alternatively tf.keras.models.save_model), and tf.keras.models.load_model(path) instead.ztf.saved_model.load works differently in 2.0 compared to 1.0. See migration information in the documentation of tf.compat.v1.saved_model.load. The calls have been converted to compat.v1.z*.saveztf.assert_equalztf.assert_none_equalztf.assert_negativeztf.assert_positiveztf.assert_non_negativeztf.assert_non_positiveztf.assert_nearztf.assert_lessztf.assert_less_equalztf.assert_greaterztf.assert_greater_equalztf.assert_integerztf.assert_typeztf.assert_scalarztf.assert_rankztf.assert_rank_at_leastztf.assert_rank_inztf.contrib.layers.layer_normz'tf.contrib.saved_model.load_keras_modelz'tf.contrib.saved_model.save_keras_modelz"tf.contrib.summary.all_summary_opsztf.contrib.summary.graphztf.contrib.summary.import_eventz8tf.contrib.summary.record_summaries_every_n_global_stepsrztf.debugging.assert_equalztf.debugging.assert_greaterz!tf.debugging.assert_greater_equalztf.debugging.assert_integerztf.debugging.assert_lessztf.debugging.assert_less_equalztf.debugging.assert_nearztf.debugging.assert_negativez tf.debugging.assert_non_negativez tf.debugging.assert_non_positiveztf.debugging.assert_none_equalztf.debugging.assert_positiveztf.debugging.assert_typeztf.debugging.assert_scalarztf.debugging.assert_rankz!tf.debugging.assert_rank_at_leastztf.debugging.assert_rank_inztf.train.exponential_decayz!tf.train.piecewise_constant_decayztf.train.polynomial_decayztf.train.natural_exp_decayztf.train.inverse_time_decayztf.train.cosine_decayztf.train.cosine_decay_restartsztf.train.linear_cosine_decayz"tf.train.noisy_linear_cosine_decayrztf.nn.nce_lossz"tf.nn.safe_embedding_lookup_sparseztf.nn.sampled_softmax_lossz(tf.keras.experimental.export_saved_modelz+tf.keras.experimental.load_from_saved_modelztf.keras.initializers.Zerosztf.keras.initializers.zerosztf.keras.initializers.Onesztf.keras.initializers.onesztf.keras.initializers.Constantztf.keras.initializers.constantz%tf.keras.initializers.VarianceScalingz tf.keras.initializers.Orthogonalz tf.keras.initializers.orthogonalztf.keras.initializers.Identityztf.keras.initializers.identityz$tf.keras.initializers.glorot_uniformz#tf.keras.initializers.glorot_normalztf.initializers.zerosztf.zeros_initializerztf.initializers.onesztf.ones_initializerztf.initializers.constantztf.constant_initializerztf.initializers.random_uniformztf.random_uniform_initializerztf.initializers.random_normalztf.random_normal_initializerz tf.initializers.truncated_normalztf.truncated_normal_initializerz tf.initializers.variance_scalingztf.variance_scaling_initializerztf.initializers.orthogonalztf.orthogonal_initializerztf.initializers.identityztf.glorot_uniform_initializerztf.initializers.glorot_uniformztf.glorot_normal_initializerztf.initializers.glorot_normalztf.losses.absolute_differenceztf.losses.add_lossztf.losses.compute_weighted_lossztf.losses.cosine_distanceztf.losses.get_lossesz!tf.losses.get_regularization_lossz#tf.losses.get_regularization_lossesztf.losses.get_total_lossztf.losses.hinge_lossztf.losses.huber_lossztf.losses.log_lossz%tf.losses.mean_pairwise_squared_errorztf.losses.mean_squared_errorztf.losses.sigmoid_cross_entropyztf.losses.softmax_cross_entropyz&tf.losses.sparse_softmax_cross_entropyztf.metrics.accuracyztf.metrics.aucz!tf.metrics.average_precision_at_kztf.metrics.false_negativesz(tf.metrics.false_negatives_at_thresholdsztf.metrics.false_positivesz(tf.metrics.false_positives_at_thresholdsztf.metrics.meanztf.metrics.mean_absolute_errorztf.metrics.mean_cosine_distanceztf.metrics.mean_iouz"tf.metrics.mean_per_class_accuracyztf.metrics.mean_relative_errorztf.metrics.mean_squared_errorztf.metrics.mean_tensorztf.metrics.percentage_belowztf.metrics.precisionztf.metrics.precision_at_kz"tf.metrics.precision_at_thresholdsztf.metrics.precision_at_top_kztf.metrics.recallztf.metrics.recall_at_kztf.metrics.recall_at_thresholdsztf.metrics.recall_at_top_kz"tf.metrics.root_mean_squared_errorz%tf.metrics.sensitivity_at_specificityz(tf.metrics.sparse_average_precision_at_kz tf.metrics.sparse_precision_at_kz%tf.metrics.specificity_at_sensitivityztf.metrics.true_negativesz'tf.metrics.true_negatives_at_thresholdsztf.metrics.true_positivesz'tf.metrics.true_positives_at_thresholdsztf.get_variablez returns ResourceVariables by default in 2.0, which have well-defined semantics and are stricter about shapes. You can disable this behavior by passing use_resource=False, or by calling tf.compat.v1.disable_resource_variables().ztf.pywrap_tensorflowz cannot be converted automatically. `tf.pywrap_tensorflow` will not be distributed with TensorFlow 2.0, please consider an alternative in public TensorFlow APIs.z&tf.contrib.distribute.MirroredStrategyztf.distribute.MirroredStrategyz'tf.contrib.distribute.OneDeviceStrategyz!tf.contrib.distribute.TPUStrategy)z1tf.contrib.distribute.CollectiveAllReduceStrategyz-tf.contrib.distribute.ParameterServerStrategyztf.summary.FileWriterztf.summary.FileWriterCacheztf.summary.Summaryztf.summary.audioztf.summary.histogramztf.summary.imageztf.summary.mergeztf.summary.merge_allztf.summary.scalarztf.summary.tensor_summaryztf.summary.textztf.saved_model.loadztf.saved_model.loader.loadz(Manual edit required) `{}` has been migrated to `{}` in TensorFlow Addons. The API spec may have changed during the migration. Please see https://github.com/tensorflow/addons for more info.)rzSuse_cudnn_on_gpu has been removed, behavior is now equivalentto setting it to True.ztf.nn.conv2d_backprop_filter)r)r@rzitf.gradients no longer takes 'colocate_gradients_with_ops' argument, it behaves as if it was set to True.)r@zhtf.hessians no longer takes 'colocate_gradients_with_ops' argument, it behaves as if it was set to True.)r@rzoOptimizer.minimize no longer takes 'colocate_gradients_with_ops' argument, it behaves as if it was set to True.zxOptimizer.compute_gradients no longer takes 'colocate_gradients_with_ops' argument, it behaves as if it was set to True.)rErzLtf.cond no longer takes 'strict' argument, it behaves as if was set to True.rr)r+ratf.contrib.summary.create_file_writer() no longer supports implicit writer re-use based on shared logdirs or resource names; this call site passed a 'name' argument that has been removed. The new tf.compat.v2.summary.create_file_writer() replacement has a 'name' parameter but the semantics are the usual ones to name the op itself and do not control writer re-use; writers must be manually re-used if desired.atf.contrib.summary.generic() takes a 'name' argument for the op name that also determines the emitted tag (prefixed by any active name scopes), but tf.compat.v2.summary.write(), which replaces it, separates these into 'tag' and 'name' arguments. The 'name' argument here has been converted to 'tag' to preserve a meaningful tag, but any name scopes will not be reflected in the tag without manual editing.))r+r)rr)rztf.contrib.summary.image no longer takes the 'bad_color' argument; caller must now preprocess if needed. This call site specifies a bad_color argument so it cannot be converted safely.))rrr)rPrzalign_corners is not supported by tf.image.resize, the new default transformation is close to what v1 provided. If you require exactly the same transformation as before, use compat.v1.image.resize.ztf.image.resize_bilinear)rPrzalign_corners is not supported by tf.image.resize, the new default transformation is close to what v1 provided. If you require exactly the same transformation as before, use compat.v1.image.resize_bilinear.zalign_corners is not supported by tf.image.resize, the new default transformation is close to what v1 provided. If you require exactly the same transformation as before, use compat.v1.image.resize_area.zalign_corners is not supported by tf.image.resize, the new default transformation is close to what v1 provided. If you require exactly the same transformation as before, use compat.v1.image.resize_bicubic.zalign_corners is not supported by tf.image.resize, the new default transformation is close to what v1 provided. If you require exactly the same transformation as before, use compat.v1.image.resize_nearest_neighbor.)tf.image.resize_areatf.image.resize_bicubic tf.image.resize_nearest_neighborz*.make_initializable_iteratorz*.make_one_shot_iteratorz tf.nn.dropoutztf.to_bfloat16ztf.to_complex128ztf.to_complex64z tf.to_doublez tf.to_floatz tf.to_int32z tf.to_int64ztf.image.extract_glimpserrr rrz tf.name_scopeztf.strings.splitz tf.device device_nameFztf.device no longer takes functions as an argument. We could not determine that the argument value is a string, so the call was converted to compat.v1.)arg_namearg_ok_predicate remove_if_okrqoptimizeTztf.zeros_like no longer takes an optimize argument, and behaves as if optimize=True. This call site specifies something other than optimize=True, so it was converted to compat.v1.ztf.ones_like no longer takes an optimize argument, and behaves as if optimize=True. This call site specifies something other than optimize=True, so it was converted to compat.v1.rreturn_same_structureztf.while_loop no longer takes 'return_same_structure' argument and behaves as if return_same_structure=True. This call site specifies something other than return_same_structure=True, so it was converted to compat.v1.rmerge_repeatedztf.nn.ctc_beam_search_decoder no longer takes the 'merge_repeated' argument and behaves as if merge_repeated=False. This call site specifies something other than merge_repeated=False, so it was converted to compat.v1. data_formatNHWC)r  arg_value_astz*tf.contrib.summary.always_record_summariesTrue)condz)tf.contrib.summary.never_record_summariesFalse save_formath5) z tf.contrib.layers.l1_regularizerz tf.contrib.layers.l2_regularizerz$tf.contrib.layers.xavier_initializerz+tf.contrib.layers.xavier_initializer_conv2dz.tf.contrib.layers.variance_scaling_initializerrrzslim.l1_regularizerzslim.l2_regularizerzslim.xavier_initializerzslim.xavier_initializer_conv2dz!slim.variance_scaling_initializerztf.keras.models.save_model)6upgrade_compat_v1_importfunction_keyword_renamesr!add_contrib_direct_import_supportsymbol_renames import_renamer ImportRenameimport_renameschange_to_functionreordered_function_namesmanual_function_reordersdictrreordersfunction_reordersupdaterWARNINGr function_warningsaddons_symbol_mappingsitemsformatfunction_arg_warnings_iterator_transformer_dropout_transformer_cast_transformer._softmax_cross_entropy_with_logits_transformer_extract_glimpse_transformer_image_resize_transformer_pool_seed_transformer_name_scope_transformer_string_split_rtype_transformer functoolspartial _rename_if_arg_found_transformer _is_ast_str _is_ast_true _is_ast_false_add_argument_transformerastStr'_add_summary_recording_cond_transformer_add_summary_step_transformer*_contrib_layers_l1_regularizer_transformer*_contrib_layers_l2_regularizer_transformer._contrib_layers_xavier_initializer_transformer8_contrib_layers_variance_scaling_initializer_transformer,_add_uniform_scaling_initializer_transformerfunction_transformersrMODULE_DEPRECATIONSmodule_deprecations) rrrdecay_function_commentassert_return_type_commentassert_rank_comment!contrib_layers_layer_norm_commentinitializers_no_dtype_commentmetrics_commentlosses_comment$deprecate_partition_strategy_comment(uniform_unit_scaling_initializer_commentsummary_api_commentcontrib_summary_comment"contrib_summary_family_arg_comment"contrib_create_file_writer_comment&contrib_summary_record_every_n_commentcontrib_summary_graph_comment$contrib_summary_import_event_comment!keras_default_save_format_commentdistribute_strategy_api_changes!contrib_mirrored_strategy_warningcore_mirrored_strategy_warning#contrib_one_device_strategy_warningcontrib_tpu_strategy_warning#contrib_collective_strategy_warningcontrib_ps_strategy_warning!keras_experimental_export_commentsaved_model_load_warningsymbol replacementwarnings rrzTFAPIChangeSpec.__init__Xs$  0"  0  9)% C%D!   O   N  7,8(  '0(,  HI  IJ  =*>&  .*/&  N.O*  %!  +,,(  <)=% N$  #B  B)C%  5 6&7"  =' (+)'  *,K L$M   #&E  E+F' 7 ( (#)  4)5%  8 9L -L  & L  & L  &L  &L ! &L ! &L  &!L"  &%L&  &)L*  &-L. " &1L2  &5L6  &9L:  &=L>  ALB " ELF  ILJ ' -MLN 2 -QLR 2 -ULV - #YLZ # #]L^ 0 .aLb % #eLf # )iLj ' #mLn * 0qLr # #uLv C 2yLz $ #}L~ $ &ALB & &ELF , &ILJ & &MLN # &QLR ) &ULV # &YLZ ' &]L^ + &aLb + &eLf ) &iLj ' &mLn # &qLr % &uLv # yLz , }L~ & ALB % "ELF , "ILJ $ "MLN % "QLR & "ULV "YLZ ) "]L^ ' "aLb - "eLf ! 0iLj ( 0mLn  0qLr - 0uLv % 0yLz 3 -}L~ 6 -ALB & )ELF & )ILJ % )MLN % )QLR ) )ULV ) )YLZ 0 )]L^ + )aLb + )eLf ) )iLj ) )mLn / )qLr . )uLv )yLz  )}L~  )ALB  )ELF # )ILJ " )MLN ) )QLR ( )ULV ( )YLZ ' )]L^ + )aLb * )eLf + )iLj * )mLn % )qLr $ )uLv # )yLz ( )}L~ ) )ALB ' )ELF ( )ILJ ( MLN  QLR * ULV $ YLZ  ]L^ , aLb . eLf # iLj  mLn  qLr  uLv 0 yLz ' }L~ * ALB * ELF 1 ILJ  MLN  QLR , ULV % YLZ 3 ]L^ % aLb 3 eLf  iLj ) mLn * qLr  uLv - yLz ) }L~ ( ALB ! ELF & ILJ  MLN $ QLR - ULV ( YLZ  ]L^ ! aLb * eLf % iLj - mLn 0 qLr 3 uLv + yLz 0 }L~ $ A LB 2 E LF $ I LJ 2 M LN    E FQ LZ  __  !] Lf 1 -i Lj ) *m Ln 2 /q Lr , (u Lx 0 '!4&91/ 3// 30%8.7&>W LDZ 44T5K5KL-DDJJL/   !'v{ ; =g (/dV$/H" #""*+ H"  #""*+ H" ' #""*+) H"& & #""*+( 'H"2  ."$% 3H">  ."$% ?H"J  ."$% KH"V  ."56 WH"b  ""'( cH"n # =% oH"t 0 ""OP 2 uH"J %!!?@> ' KH"b ' =) cH"h #!! >% iH"z $ =& {H"@  ""+, AH"P # ""45% QH"b !""01!  !""34$  !""<=- AH"DR44T5O5OP*g"')>g""$9g" -g" + g" - g" , g" )g" (g" (g" (g" 2 :g" #$@g"  9g" "#<g" #$=!g"" +,E#g"$ $%;%g"& $%;'g"( 0)g". ;/g"0 Y&& ,}(u341g"< ** ,z)JK=g"H  )) ,z)JKIg"T ** ,,)0 1Ug"d ():): ,%*F *Geg"t I-- %"''&/+ug"| 9,, %"''&/+}g"D 5i6G6G 3&7BEg"H #$AIg"J %&CKg"L '(EMg"N #$AOg"P 4Y5F5F 3'6CQg"T $%BUg"X 7 6 : : D 8 8 6 6 : : D&/&7&7 %"''$-')Gg"DP44T5O5OP5IIDrctjt}|j|t |j }t |vr:|jr.|s,tj||j|dS|rMi|_ i|_ i|_ i|_ i|_i|_t j"|_i|_i|_||j*|j,fS)NT)after_compat_v1_upgrade)rPastaAnalyzeVisitorrvisitsetresultsr$rr( preprocessfunction_handler%rrr(r rrGrHrFrlogwarnings_and_errors)r root_nodergvisitor detectionss rrlzTFAPIChangeSpec.preprocesss++,C,EFG MM)W__%J J&4+H+H #$$Y/ __Y_ EE d!d&(d#d!d "d!7!K!Kd#%d d gkk7#>#> >>rcR|j|j|jy)N)rr)rrrr s rclear_preprocessingz#TFAPIChangeSpec.clear_preprocessings%MM 2 2+/+H+HJrN)FF)F)rrrr1rrlrtrrrr3r3Us;DJL,?<Jrr3ctjg}ttdr|tjgz }ttdr|tjgz }ttdr|tj gz }t ||S)z0Determine whether this node represents a string.Bytes JoinedStrFormattedValue)r=r>hasattrrvrwrx isinstance)r. allowed_typess rr9r9sl77)- S'cii[ M S+cmm_$M S"#c(())M D- ((rcttdr*t|tjxr|jduSt|tj xr|j dk(S)N NameConstantTrryr=rzr}rNameidr.s rr:r: sL S.! dC,, - D$**2DD dCHH % ;$''V*;;rcttdr*t|tjxr|jduSt|tj xr|j dk(S)Nr}Frr~rs rr;r;sL S.! dC,, - E$**2EE dCHH % <$''W*<._replace_keep_prob_nodes ''A,CCJCN sswwy )+I OO!!&)Y?i+ JJi37 JJi37r keep_probz0Changing keep_prob arg of tf.nn.dropout to rate rmrztf.nn.dropout called without arguments, so automatic fix was disabled. tf.nn.dropout has changed the semantics of the second argument.rrzHChanging keep_prob arg of tf.nn.dropout to rate, and recomputing value. N) rrrrrrrrlenrr r=r)rr.rr+rrrrate_args rr.r.s 8 ==i}} # kk9>>4;;FHIimi9 k  ^aKK$++t9:; {{vTYYq\:HHhnn5MM" ! KKdoo()* Krc |dd}|dk(rd}n|dk(rd}tjdtjtjdtj |tj  }t |j d k(rStjd |j d }|j dd|_|jj||j|j_ |jdz|j_ |jj|t|jtjrd|j_n7t|jtjsJd|j_|jt j"|j|jd|d|df|S)z7Transforms to_int and to_float to cast(..., dtype=...).rNfloatfloat32doublefloat64rr!rctxrattrrrrr+rdcastChanged z call to tf.cast(..., dtype=tf.).)r=rrrLoadrrrrrrrrzrrrrr)rr.rr+r dtype_strnew_argname_args rr/r/sj12h)'IHI KKG!mm#((d?Bxxz3K1: L M'  ^q{{v!%20H #2DIMM"'--!__S0'----w 3==)DIIN dii ** *DIIL++y~~t{{DOOBKBKMNO +rcd}|jD]_}|jdk(s|||jr7|jtj |j |jdf|cS|S)z)Wrap labels argument with stop_gradients.ct|tjxrt|jtjxrn|jj dk(xrSt|jj tjxr#|jj jdk(}|ry tjtjdtj|gg}tjj|||tj||y#t$rDtjtjdtj|ggdd}YwxYw)z"Wrap labels with tf.stop_gradient. stop_gradientr!Fztf.stop_gradientrNT)rzr=Callrrrrrrr TypeErrorrrrr)rralready_stop_gradrs r _wrap_labelzC_softmax_cross_entropy_with_logits_transformer.._wrap_labels"#Isxx89#INNCMMB9",,?9$INN$8$8#((C9#--00D8   '(( (((chhj 9 +ri OO!!&)Y?i+  '(( (((chhj 9 +r4'i's5?D,,A E98E9rz~Changing labels arg of tf.nn.softmax_cross_entropy_with_logits to tf.stop_gradient(labels). Please check this transformation. )rrrrrrrr)rr.rr+rrkargs rr0r0sl0mmd xx8 T4:: & Y^^T[[$//)* + k +rc|ddj}tjdtjtjtjtjdtj dtj dtj |tj  }t |jd k(rStjd |jd  }|jdd |_|jj|t |jd k(r8tjd|jd  }|jdd |_g}|jD]#} | jdk7s|j| %||_|j|j_ |jdz|j_ |jj|t|jtjrd|j_n7t|jtjsJd|j_|jt$j&|j|jd|d|df|S)z>Transforms image.resize_* to image.resize(..., method=*, ...).Nmethodr!rimager ResizeMethodrrpreserve_aspect_ratiorrrPrresizerz; call to tf.image.resize(..., method=tf.image.ResizeMethod.r)upperr=rrrrrrrrrrrrrzrrrrr) rr.rr+r resize_methodrpos_arg new_keywordsrs rr2r2sqr(.."- KKH!mm"%--&)mm*-((d *K)0chhj'B&4 #E "/CHHJ @ A' ^qkk5 $ " /G #2DIMM!^qkko $ " /G #2DI, MMb vv "$-'--!__S0'----w 3==)DIIN dii ** *DIIL++y~~t{{DOO6?6CEFG +rc d}d}d}g}|jD]} tjdddk\rt| tj rn| j dk(r| }n| j dk(s| j dk(rt| d|j} t| d |j} |jtj| | d | j d |xs|f| j dk(rt| jsd }d }|j| |r|x|jt jdt j d |j#tj|j|jd|xs|zfn7|j#tj$|j|jdf|r ||_|Sy)zBRemoves seed2 and deterministic, and adds non-zero seed if needed.NFr)rrseedr> deterministicrrrrT*rz>6:vvy0D0234 ? "RXX&-h!$#++& DE hh ..$++t H4 !  hh   T[[$// L  DM K rcd}|jD]v}|jdk(s|jtj|j |j dfd|_|jrdnd}|||j|cSt|jdk\rO|||jd|jtj|j |j df|Sy) Nctjd}tjd}tj|||}tjj |||tj ||tjjj|jddtjjj|jddy ) z0Replaces old_value with 'uniform' or 'gaussian'.uniform)sgaussian)bodytestorelserrrrN) r=r>IfExprrrrrrrjr)rrrrrs r_replace_uniform_noise_nodezA_extract_glimpse_transformer.._replace_uniform_noise_nodehsgg "Gww$H wYxHI OO!!&)Y?i+ JJinnh< JJinnh>4;;'()"m(..iJe!-1D1DE k  ^qdiil3KKdoo345 K rc |jD]}|jdk(s|cSd}tj|jdj }|`|jjtjd||jtj|j |jd|xs|d|df|S)zAdds a step argument to the summary API call if not specified. The inserted argument value is tf.compat.v1.train.get_or_create_global_step(). rz.tf.compat.v1.train.get_or_create_global_step()rrzSummary API writing function z6 now requires a 'step' argument; inserting default of r) rrr=parserrrrrrr'r)rr.rr+r keyword_arg default_value ast_values rr@r@s ]]k& kC-ii &++A.44)--s{{vY?@++doo$-$5$5}FGH +rc |jjtj||jtj |j |jd|d|xs|df|S)zAdds cond argument to tf.contrib.summary.xxx_record_summaries(). This is in anticipation of them being renamed to tf.summary.record_if(), which requires the cond argument. zAdding `z` argument to zH in anticipation of it being renamed to tf.compat.v2.summary.record_if())rrrrrrrr)rr.rr+rrs rr?r?s_))5;;t$%++nndkk4??,0)2Ct2CEFG +rc F|D]} t|||||| ||| } | r| n|}|S)aReplaces the given call with tf.compat.v1 if any of the arg_names is found. Args: parent: Parent of node. node: ast.Call node to modify. full_name: full name of function to modify. name: name of function to modify. logs: list of logs to append to. arg_names: list of names of the argument to look for. arg_ok_predicate: predicate callable with the ast of the argument value, returns whether the argument value is allowed. remove_if_ok: remove the argument if present and ok as determined by arg_ok_predicate. message: message to print if a non-ok arg is found (and hence, the function is renamed to its compat.v1 version). Returns: node, if it was modified, else None. r8 rr.rr+r arg_namesr r rqr  rename_nodes r$_rename_if_any_arg_found_transformerr sF:0h2643._get_distribution's} ;;K LD YYq\  F OO!!&&++yA JJdHc2 JJdHc2 KrFrTrrrrrz1.0rrz "fan_avg"moderrrrrrrrrz}Changing tf.contrib.layers xavier initializer to a tf.compat.v1.keras.initializers.VarianceScaling and converting arguments. )rrrrrrrrrjrrr=rrrrrrrrrr)rr.rr+rrfound_distributionrrrrrrr uniform_distrrs rrCrCs ]] Bk)#&ko##i#K$5$56i oo##KIF jj 1 18SA jj 1 18SA B, ++e %ckkgU;< ] #$ckkfD9:^q TYYq\ *D dCD ;;}-L lKL^q diilCD^q tyy|DE$).$- 99?? ! !&yy))*,,-NO$))/!$))//)$))//$$)).++y~~t{{DOO*+, +rc2d}d}d}|jD]f}|jdk(r d|_d}|jdk(r|||j|jdk(sMd |_|||jht|jd k\rd}t|jd k\r|||jd t|jd k\r|||jd |s?t j d } tjd| g|jz|_|jjj} |jjj} tjd|j_| |jj_ | |jj_ d|j_|jtj |j|jdf|S)aUpdates references to contrib.layers.variance_scaling_initializer. Transforms: tf.contrib.layers.variance_scaling_initializer( factor, mode, uniform, seed, dtype ) to tf.compat.v1.keras.initializers.VarianceScaling( scale=factor, mode=mode.lower(), distribution=("uniform" if uniform else "truncated_normal"), seed=seed, dtype=dtype) And handles the case where no factor is provided and scale needs to be set to 2.0 to match contrib's default instead of tf.keras.initializer's default of 1.0 ctjd}|jdj}tjj ||j |tjj |||tjjj|ddtjjj|ddy)zFReplaces old_value: ("uniform" if (old_value) else "truncated_normal")rrrrrrNr)rrrrs r_replace_distributionzW_contrib_layers_variance_scaling_initializer_transformer.._replace_distributionys <>I ^^A  $ $F OO!!&&++yA OO!!&)Y? JJi37 JJi37rctjd}|jdjj}tj j ||j|tj j |||tjjj|ddtjjj|ddy)z,Replaces old_value with (old_value).lower().z mode.lower()rrrrrN) rrrrrrrrrrj)rrrrs r _replace_modezO_contrib_layers_variance_scaling_initializer_transformer.._replace_modes N+I >>!  " " ' 'D OO!!$ I> OO!!&)Y? JJi37 JJi37rFrrTrrrrrrz2.0rrrzChanging tf.contrib.layers.variance_scaling_initializer to a tf.compat.v1.keras.initializers.VarianceScaling and converting arguments. )rrrrrrrr=rrrrrrrrr) rr.rr+rr!r# found_scaler scale_valuerrs rrDrDhs" 8 8 +]]|jtj|j |j dfd|_|jdk(s`|}c|rR|jtj|j |j df|jj|t|jdkDrK|jdd|_ |jtj|j |j df|jjj }|jjj }tjd|j_ ||jj_||jj_d |j_ |S) zReplace slim l1 regularizer with Keras one. This entails renaming the 'scale' arg to 'l' and dropping any provided scope arg. Nrz"Renaming scale arg of regularizer lscopezpDropping scope arg from tf.contrib.layers.l1_regularizer, because it is unsupported in tf.keras.regularizers.l1 rtf.keras.regularizersl1)rrrrrrrremoverrrrrr) rr.rr+r scope_keywordrrrs rrArAse-g{{g kk9>>4;;8:;gk{{gm KKdooLMN MM'^a "1 DIKKdooLMN 99?? ! !&yy))*,,-DE$))/!$))//)$))//$)). +rc6d}d}|jD]=}|jdk(rd|_|||j|jdk(s<|}?t|jdk\r|||jd|rR|j t j|j|jdf|jj|t|jdkDrK|jdd|_|j t j|j|jdf|j t j|j|jd f|jjj}|jjj} t jd |j_||jj_| |jj_ d |j_ |S) zbReplace slim l2 regularizer with Keras one, with l=0.5*scale. Also drops the scope argument. ctjd}d|_d|_tj|tj |}t jj|||t jjj|ddt jjj|ddy ) z(Replaces old_value with 0.5*(old_value).g?rrrrrrrN) r=rrrrMultrrrrrrj)rrhalfrs r_replace_scale_nodezG_contrib_layers_l2_regularizer_transformer.._replace_scale_nodes 77S>DDKDO t )+I OO!!&)Y? JJi37 JJi37rNrr'r(rrzpDropping scope arg from tf.contrib.layers.l2_regularizer, because it is unsupported in tf.keras.regularizers.l2 zlMultiplying scale arg of tf.contrib.layers.l2_regularizer by half to what tf.keras.regularizers.l2 expects. r)l2)rrrrrrrrrrr+rrr) rr.rr+rr1r,rrrs rrBrBs 8 -g{{ggk'7==1{{gm  ^qdiil+KKdooLMN MM'^a "1 DIKKdooLMN++y~~t{{DOOFGH 99?? ! !&yy))*,,-DE$))/!$))//)$))//$)). +rc tj|dd\}}tj|dd\}}|rtj|dk7r|j tj |j |jdfd}|j tj |jj |jjd|d |ftj||jj} tj| |jtjj||j| |S|r]|j tj |j |jd fg|_tj d| g|_|S|j tj$|j |jd fy )zGFix name scope invocation to use 'default_name' and omit 'values' args.r+r default_namerNonez`name` passed to `name_scope`. Because you may be re-entering an existing scope, it is not safe to convert automatically, the v2 name_scope does not support re-entering scopes by name. ztf.compat.v1.name_scopeRenamed rz2Using default_name as name in call to name_scope. rzPname_scope call with neither name nor default_name cannot be converted properly.N)rrrrrrrrrrrr=rrrrrrr ) rr.rr+r name_found default_foundr4r new_name_nodes rr4r4! sv,,T61=*d ) 7 7na P-EJJt$.KKdoo )HKK!1!14993G3G'0(;=>,,Xtyy}}EMmTYY/ OO!!$ =A K KKdooFHIDI[[V<@ADM K++y T__%&'rcF|jddd}t|||||S)Nrrr)r _rename_func)r.rrreasonrs r_rename_to_compat_v1r=G s)   uoq 9( dIxv >>rc p|jtj|j|jd|d|d|ftj ||j j}tj||j tjj||j ||S)Nr6rz: ) rrrrrrrrr=rrrr)r.rrrr<r9s rr;r;L s~++y~~t{{DOO)2HfEGH**8TYY]]C-M499-//dii? +rct|jD]\}}|jdk(st|jrT|j t j|j|jdf|jj|nt|||dcSd}t|jD]\}}|jdk(sd}t|jtjrE|jjdk(sVt!||d|d }|jj|t|||d cS|st|||d St#|||||S) zGUpdate tf.string_split arguments: skip_empty, sep, result_type, source. skip_emptyz0removed argument skip_empty for tf.string_split.zFtf.string_split's replacement no longer takes the skip_empty argument.FsepTrztf.strings.bytes_splitz;Splitting bytes is not handled by tf.strings.bytes_split().zThe semantics for tf.string_split's sep parameter have changed when sep is the empty string; but sep is not a string literal, so we can't tell if it's an empty string.zThe semantics for tf.string_split's sep parameter have changed when sep unspecified: it now splits on all whitespace, not just the space character.)rrrr;rrrrrrrr=rzr=r>rr;r5)rr.rr+rrr found_seps r_string_split_transformerrCU s[' .ea vv rxx  Y^^T[[$//GI J ! # )T$-. . .)'9ea vvi BHHcgg & 88:: I7KM$ --  A # )T 89 99   i   )y$ MMrc $d}t|jD]\}}|jdk(st|jt j r|jjdvr|jtj|j|jd|jjd|xs|f|jj||jjdk(rd}nt|||d|zcSnt|jD]\}}|jd k(sd |_|rt|t jr|j d k(ry |jtj|j|jd ft jt#j$|d } t j&|gg}|S|S#t($rt j&|ggd d }Y|SwxYw)z7Update tf.strings.split arguments: result_type, source.T result_type) RaggedTensor SparseTensorzRemoved argument result_type=rrFFz-%s no longer takes the result_type parameter.sourcerb to_sparseNzhAdding call to RaggedTensor.to_sparse() to result of strings.split, since it now returns a RaggedTensor.)rr)rrrrzrr=r>rrrrrrrr=rrcopydeepcopyrr)rr.rr+rneed_to_sparserrs rr5r5 s.' ea vv RXXsww ' ((**8 8 Y^^T[[$//hhjj)"3t"356 7 ! 88:: ' .# )T ;i GI I  'ea vvbf 63==)fkk[.H KK doo 0 12 ==t}}T2 ED0 XXdB #d ++ 0 XXdBD$ /d +0sG))"HH)NNFN)0r1r=rJr6rrtensorflow.tools.compatibilityrrrrAnalysisResultrrr$r%APIAnalysisSpecr NodeVisitorr( NoUpdateSpecr3r9r:r;r8r<r-r.r/r0r2r3r1r@r?r r rErCrDrArBr4r=r;rCr5rrrrRs\I  94A6 E 00E< 00<%[1$[1i77S__(iJi,,iJX- )<=/36:27-1 BJ 4)X$N$N-`/ d#J& * $Z %P