B ӻd;@sdZddlmZddlmZddlmZddlmZddl m Z ddl m Z ddl m Z dd l mZdd lmZed gd Gd ddeZd!ddZd"ddZddZddZejfddZedgd dejdfddZddZdd ZdS)#z$Utilities related to loss functions.)distribution_strategy_context)ops)backend) keras_tensor) array_ops)control_flow_ops)math_ops) ragged_tensor) keras_exportzkeras.losses.Reduction)Zv1c@s8eZdZdZdZdZdZdZeddZ edd Z d S) ReductionV2aiTypes of loss reduction. Contains the following values: * `AUTO`: Indicates that the reduction option will be determined by the usage context. For almost all cases this defaults to `SUM_OVER_BATCH_SIZE`. When used with `tf.distribute.Strategy`, outside of built-in training loops such as `tf.keras` `compile` and `fit`, we expect reduction value to be `SUM` or `NONE`. Using `AUTO` in that case will raise an error. * `NONE`: No **additional** reduction is applied to the output of the wrapped loss function. When non-scalar losses are returned to Keras functions like `fit`/`evaluate`, the unreduced vector loss is passed to the optimizer but the reported loss will be a scalar value. Caution: **Verify the shape of the outputs when using** `Reduction.NONE`. The builtin loss functions wrapped by the loss classes reduce one dimension (`axis=-1`, or `axis` if specified by loss function). `Reduction.NONE` just means that no **additional** reduction is applied by the class wrapper. For categorical losses with an example input shape of `[batch, W, H, n_classes]` the `n_classes` dimension is reduced. For pointwise losses your must include a dummy axis so that `[batch, W, H, 1]` is reduced to `[batch, W, H]`. Without the dummy axis `[batch, W, H]` will be incorrectly reduced to `[batch, W]`. * `SUM`: Scalar sum of weighted losses. * `SUM_OVER_BATCH_SIZE`: Scalar `SUM` divided by number of elements in losses. This reduction type is not supported when used with `tf.distribute.Strategy` outside of built-in training loops like `tf.keras` `compile`/`fit`. You can implement 'SUM_OVER_BATCH_SIZE' using global batch size like: ``` with strategy.scope(): loss_obj = tf.keras.losses.CategoricalCrossentropy( reduction=tf.keras.losses.Reduction.NONE) .... loss = tf.reduce_sum(loss_obj(labels, predictions)) * (1. / global_batch_size) ``` Please see the [custom training guide]( https://www.tensorflow.org/tutorials/distribute/custom_training) for more details on this. autononesumZsum_over_batch_sizecCs|j|j|j|jfS)N)AUTONONESUMSUM_OVER_BATCH_SIZE)clsr\/var/www/html/venv/lib/python3.7/site-packages/tensorflow/python/keras/utils/losses_utils.pyallQszReductionV2.allcCs||krtd|dS)NzInvalid Reduction Key %s.)r ValueError)rkeyrrrvalidateUs zReductionV2.validateN) __name__ __module__ __qualname____doc__rrrr classmethodrrrrrrr s- r Nc s|t|p dbttjs(tttjs>tj}|j}j}|j}|dk r|dk r||}||dkr|j d drt dgn*||dkr|j d drt dgfSt t }|dks|j d dr$tt|d|fddfdd|dks@|j d drjtt|d|fddfd dfSQRXdS) a$Squeeze last dim if ranks differ from expected by exactly 1. In the common case where we expect shapes to match, `expected_rank_diff` defaults to 0, and we squeeze the last dimension of the larger rank if they differ by 1. But, for example, if `labels` contains class IDs and `predictions` contains 1 probability per class, we expect `predictions` to have 1 more dimension than `labels`, so `expected_rank_diff` would be 1. In this case, we'd squeeze `labels` if `rank(predictions) - rank(labels) == 0`, and `predictions` if `rank(predictions) - rank(labels) == 2`. This will use static shape if available. Otherwise, it will add graph operations, which could result in a performance hit. Args: labels: Label values, a `Tensor` whose dimensions match `predictions`. predictions: Predicted values, a `Tensor` of arbitrary dimensions. expected_rank_diff: Expected result of `rank(predictions) - rank(labels)`. name: Name of the op. Returns: Tuple of `labels` and `predictions`, possibly with last dim squeezed. remove_squeezable_dimensionsNcstdgS)Nr!)rsqueezer) predictionsrrz.remove_squeezable_dimensions..csS)Nrr)r#rrr$r%cstdgS)Nr!)rr"r)labelsrrr$r%csS)Nrr)r&rrr$r%)r name_scope isinstancer RaggedTensorr"convert_to_tensor_v2_with_dispatchshapendimsZdimsZis_compatible_withrr"rankrcondrequal) r&r#Zexpected_rank_diffnameZpredictions_shapeZpredictions_rankZ labels_shapeZ labels_rank rank_diffr)r&r#rr[s@         rc sj}|j}dk rj}|j}|dk rX|dk rX||dksH|ddkrt\ndttfddtdtdfdd}ttd|\dkr̈fSj}|j} | dkrfS|dk rB| dk rB| |dkrt dgn|| dkr8t dgfSt} | tfddfd d fd d } tt| dfd d| fS)aSqueeze or expand last dimension if needed. 1. Squeezes last dim of `y_pred` or `y_true` if their rank differs by 1 (using `remove_squeezable_dimensions`). 2. Squeezes or expands last dim of `sample_weight` if its rank differs by 1 from the new rank of `y_pred`. If `sample_weight` is scalar, it is kept scalar. This will use static shape if available. Otherwise, it will add graph operations, which could result in a performance hit. Args: y_pred: Predicted values, a `Tensor` of arbitrary dimensions. y_true: Optional label `Tensor` whose dimensions match `y_pred`. sample_weight: Optional weight scalar or `Tensor` whose dimensions match `y_pred`. Returns: Tuple of `y_pred`, `y_true` and `sample_weight`. Each of them possibly has the last dimension squeezed, `sample_weight` could be extended by one dimension. If `sample_weight` is None, (y_pred, y_true) is returned. Nr r!cs tS)N)rr)y_predy_truerrr$sz.squeeze_or_expand_dimensions..cstfddS)NcsfS)Nrr)r2r3rrr$r%z@squeeze_or_expand_dimensions....)rr.r) is_last_dim_1 squeeze_dimsr2r3rrr$srcstdgS)Nr!)rr"r) sample_weightrrr$r%cs*fdd}ttd|fddS)NcstdgS)Nr!)r expand_dimsr)r6rrr$r%zMsqueeze_or_expand_dimensions.._maybe_expand_weights..r!csS)Nrr)r6rrr$r%)rr.rr/)Zexpand_weights)r1r6rr_maybe_expand_weightss z;squeeze_or_expand_dimensions.._maybe_expand_weightscsttdS)Nr )rr.rr/r)r8maybe_squeeze_weightsr1rr_maybe_adjust_weightss z;squeeze_or_expand_dimensions.._maybe_adjust_weightscsS)Nrr)r6rrr$r%) r+r,rrr-rr/rr.r"r7) r2r3r6Z y_pred_shapeZ y_pred_rankZ y_true_shapeZ y_true_rankZmaybe_squeeze_dimsZ weights_shapeZ weights_rankZweights_rank_tensorr:r)r8r4r9r1r6r5r2r3rsqueeze_or_expand_dimensionssH    r;cCst|}tj||ddS)a,Computes a safe mean of the losses. Args: losses: `Tensor` whose elements contain individual loss measurements. num_present: The number of measurable elements in `losses`. Returns: A scalar representing the mean of `losses`. If `num_present` is zero, then zero is returned. value)r0)r reduce_sumZ div_no_nan)lossesZ num_presentZ total_lossrrr _safe_means r?c Cs0td}tjtj||d|jdSQRXdS)z3Computes the number of elements in `losses` tensor.Z num_elements)r0)dtypeN)rr'rcastrsizer@)r>scoperrr _num_elementss rDcCs6|tjkr|}n"t|}|tjkr2t|t|}|S)z2Reduces the individual weighted loss measurements.)r rrr=rr?rD)weighted_losses reductionlossrrrreduce_weighted_losss    rHz/keras.__internal__.losses.compute_weighted_lossc Cst||tjkrtj}|dkr&d}t|p0d|t_t |t j t j fs\t|}|j}t |t j sxt|}t|d}t|d}t|d|\}}}t||}t||}t||}|SQRXdS)aComputes the weighted loss. Args: losses: `Tensor` of shape `[batch_size, d1, ... dN]`. sample_weight: Optional `Tensor` whose rank is either 0, or the same rank as `losses`, or be broadcastable to `losses`. reduction: (Optional) Type of `tf.keras.losses.Reduction` to apply to loss. Default value is `SUM_OVER_BATCH_SIZE`. name: Optional name for the op. Raises: ValueError: If the shape of `sample_weight` is not compatible with `losses`. Returns: Weighted loss `Tensor` of the same type as `losses`. If `reduction` is `NONE`, this has the same shape as `losses`; otherwise, it is scalar. Ng?Z weighted_lossfloat32)r rrrrr'rZget_default_graphZ_last_loss_reductionr(rZ KerasTensorr r)r*r@rrAr;multiplyrH)r>r6rFr0Z input_dtype_rErGrrrcompute_weighted_losss*           rLcCs"tj}|dkr|d|9}|S)zBScales and returns the given loss value by the number of replicas.r g?)rZ get_strategyZnum_replicas_in_sync)Z loss_valueZ num_replicasrrrscale_loss_for_distributionOs  rMcstdxT|D]L}|jjrJdks,|jjjkr4|jn|jhddhkrJd|jjr |Sq Wrpfdd|D}|S)abCast a list of losses to a common dtype. If any loss is floating-point, they will all be casted to the most-precise floating-point loss. Otherwise the losses are not casted. We also skip casting losses if there are any complex losses. Args: losses: A list of losses. Returns: `losses`, but they have been casted to a common dtype. NZbfloat16Zfloat16rIcsg|]}t|qSr)rrA).0rG) highest_floatrr osz/cast_losses_to_common_dtype..)r@Z is_floatingrBZ is_complex)r>rGr)rOrcast_losses_to_common_dtypeXs  rQ)rN)NN)rZtensorflow.python.distributerZtensorflow.python.frameworkrZtensorflow.python.kerasrZtensorflow.python.keras.enginerZtensorflow.python.opsrrrZtensorflow.python.ops.raggedr Z tensorflow.python.util.tf_exportr objectr rr;r?rDrrHrLrMrQrrrrs,         > > Y 7