B ӻdv @sdZddlZddlmZddlmZddlmZddlm Z ddl m Z dd d Z dd d Z GdddeZddZddZddZdS)zTraining-related utilities.N) tensor_shape) tensor_util) generic_utils) array_ops)nestTcsd}t|tsd}|g}tdd|Drd|sPfdd|D}dd|D}qpfdd|D}n t|}|r||d }|S) aSlices batches out of provided arrays (workaround for eager tensors). Unfortunately eager tensors don't have the same slicing behavior as Numpy arrays (they follow the same slicing behavior as symbolic TF tensors), hence we cannot use `generic_utils.slice_arrays` directly and we have to implement this workaround based on `concat`. This has a performance cost. Args: arrays: Single array or list of arrays. indices: List of indices in the array that should be included in the output batch. contiguous: Boolean flag indicating whether the indices are contiguous. Returns: Slice of data (either single array or list of arrays). FTcss|]}t|VqdS)N)rZ is_tf_type).0xr _/var/www/html/venv/lib/python3.7/site-packages/tensorflow/python/keras/engine/training_utils.py 0szslice_arrays..csg|]fddDqS)csg|]}||dqS)r )ri)rr r 2sz+slice_arrays...r )r)indices)rr r2sz slice_arrays..cSsg|]}tj|ddqS)r)Zaxis)rconcat)rrr r r r3scs$g|]}|dddqS)rr r )rr)rr r r5sr) isinstancelistanyr slice_arrays)Zarraysr contiguousZconverted_to_listentriesZslicesr )rr rs  rFcCs\|dk otdd|D}|o.tdd|D}|s>d||fS|sL|||fS|rtt|tt|tt|tt||dk rt|t|g}xt|D]\}}|dkr@t||tj} ||} | r| j nt | } |dk o||dk} | r| d| dfn| df} | | r2t | nt | q| |qWt|||fS)aAdds 1.0 as sample weights for the outputs for which there is no weight. Args: outputs: List of model outputs. sample_weights: List of sample weight inputs. sample_weight_modes: List of sample weight modes or None. check_all_flat: Ensure that inputs are not nested structures. This is not a free check, so we may not want to run it eagerly every iteration. Returns: Tuple of sample weights, one sample weight for every output, and booleans describing the raw sample weights. Ncss|]}|dk VqdS)Nr )rwr r r r Nsz0handle_partial_sample_weights..css|]}|dkVqdS)Nr )rrr r r r PsZtemporalrr ) rrZassert_same_structure list_to_tupleflatten enumeraternpZndarrayshaperappendZones)outputsZsample_weightsZsample_weight_modesZcheck_all_flatZany_sample_weightZpartial_sample_weightZnew_sample_weightsr swZas_numpyoutputZ output_shapeZ is_temporalZsw_shaper r r handle_partial_sample_weights>s@      r"c@s(eZdZdZddZddZddZdS) RespectCompiledTrainableStateatSet and restore trainable state if it has changed since compile. The keras API guarantees that the value of each Layer's `trainable` property at `Model.compile` time will be used when training that model. In order to respect this requirement, it may be necessary to set the trainable value of layers to their compile time values before beginning a training endpoint and restore the values before returing from said endpoint. This scope checks if any layer's trainable state has changed since Model compile, and performs this set and un-set bookkeeping. However, the trainable state of a layer changes quite infrequently, if ever, for many kinds of workflows. Moreover, updating every layer in a model is an expensive operation. As a result, we will only explicitly set and unset the trainable state of a model if a trainable value has changed since compile. cCs||_d|_d|_d|_dS)NF)_model_current_trainable_state_compiled_trainable_state_should_set_trainable)selfmodelr r r __init__sz&RespectCompiledTrainableState.__init__cCsf|j|_|jj|_x6|jD](\}}||jkr"||j|kr"d|_Pq"W|jrb|j|jdS)NT)r$Z_get_trainable_stater%r&itemsr'_set_trainable_state)r(layerZ trainabler r r __enter__s   z'RespectCompiledTrainableState.__enter__cCs|jr|j|jdS)NF)r'r$r,r%)r(Ztype_argZ value_argZ traceback_argr r r __exit__sz&RespectCompiledTrainableState.__exit__N)__name__ __module__ __qualname____doc__r*r.r/r r r r r#ysr#cCsJdd}x$||r,|js td|jd}q Wt|ddrF|j|jfSdS)aBRetrieves input shape and input dtype of layer if applicable. Args: layer: Layer (or model) instance. Returns: Tuple (input_shape, input_dtype). Both could be None if the layer does not have a defined input shape. Raises: ValueError: in case an empty Sequential or Functional model is passed. cSst|dr|jp|jjdkS)N_is_graph_networkZ Sequential)hasattrr4 __class__r0)r-r r r _is_graph_modelsz2get_input_shape_and_dtype.._is_graph_modelz)An empty Model cannot be used as a Layer.r_batch_input_shapeN)NN)Zlayers ValueErrorgetattrr8Zdtype)r-r7r r r get_input_shape_and_dtypes   r;cCs(t|\}}|dk r$t|djSdS)zGets the static batch size of a Layer. Args: layer: a `Layer` instance. Returns: The static batch size of a Layer. Nr)r;rZ Dimensionvalue)r-Zbatch_input_shape_r r r get_static_batch_sizes r>cCst|trt|S|S)zADatasets will stack the list of tensor, so switch them to tuples.)rrtuple)Z maybe_listr r r rs r)T)F)r3numpyrZtensorflow.python.frameworkrrZtensorflow.python.keras.utilsrZtensorflow.python.opsrZtensorflow.python.utilrrr"objectr#r;r>rr r r r s      % :0"