def weighted(y_true, y_pred, weights, mask=None):
"""Wrapper function.
Arguments:
y_true: `y_true` argument of `fn`.
y_pred: `y_pred` argument of `fn`.
weights: Weights tensor.
mask: Mask tensor.
Returns:
Scalar tensor.
"""
# score_array has ndim >= 2
score_array = fn(y_true, y_pred)
if mask is not None:
mask = math_ops.cast(mask, y_pred.dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = squeeze_or_expand_dimensions(
mask, None, weights)
weights *= mask
# Apply sample weighting.
if weights is not None:
# Update dimensions of weights to match with values if possible.
score_array, _, weights = squeeze_or_expand_dimensions(
score_array, None, weights)
try:
# Broadcast weights if possible.
weights = weights_broadcast_ops.broadcast_weights(
weights, score_array)
except ValueError:
# Reduce values to same ndim as weight array.
ndim = K.ndim(score_array)
weight_ndim = K.ndim(weights)
score_array = K.mean(score_array,
axis=list(range(weight_ndim, ndim)))
score_array = math_ops.multiply(score_array, weights)
score_array = math_ops.reduce_sum(score_array)
weights = math_ops.reduce_sum(weights)
score_array = math_ops.div_no_nan(score_array, weights)
return K.mean(score_array)
def weighted(y_true, y_pred, weights, mask=None):
"""Wrapper function.
Arguments:
y_true: `y_true` argument of `fn`.
y_pred: `y_pred` argument of `fn`.
weights: Weights tensor.
mask: Mask tensor.
Returns:
Scalar tensor.
"""
# score_array has ndim >= 2
score_array = fn(y_true, y_pred)
if mask is not None:
mask = math_ops.cast(mask, y_pred.dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = squeeze_or_expand_dimensions(mask, None, weights)
weights *= mask
# Apply sample weighting.
if weights is not None:
# Update dimensions of weights to match with values if possible.
score_array, _, weights = squeeze_or_expand_dimensions(
score_array, None, weights)
try:
# Broadcast weights if possible.
weights = weights_broadcast_ops.broadcast_weights(weights, score_array)
except ValueError:
# Reduce values to same ndim as weight array.
ndim = K.ndim(score_array)
weight_ndim = K.ndim(weights)
score_array = K.mean(score_array, axis=list(range(weight_ndim, ndim)))
score_array = math_ops.multiply(score_array, weights)
score_array = math_ops.reduce_sum(score_array)
weights = math_ops.reduce_sum(weights)
score_array = math_ops.div_no_nan(score_array, weights)
return K.mean(score_array)
def apply_mask(y_p, sw, mask):
"""Applies any mask on predictions to sample weights."""
if mask is not None:
mask = math_ops.cast(mask, y_p.dtype)
if sw is not None:
mask, _, sw = (losses_utils.squeeze_or_expand_dimensions(
mask, sample_weight=sw))
sw *= mask
else:
sw = mask
return sw
def update_state(self, y_true, y_pred, sample_weight=None):
if sample_weight is not None:
raise NotImplementedError()
y_pred, y_true, sample_weight = squeeze_or_expand_dimensions(
y_pred, y_true, sample_weight
)
y_true = tf.cast(y_true, tf.int64)
y_pred = tf.equal(tf.argmax(y_pred, axis=-1), self._index)
y_true = tf.equal(y_true, self._index)
return super(IndividualIntersectionOverUnion, self).update_state(
y_true, y_pred, sample_weight
)
def update_state(self, y_true, y_pred, sample_weight=None):
y_pred, y_true, sample_weight = squeeze_or_expand_dimensions(
y_pred, y_true, sample_weight
)
y_pred = tf.argmax(y_pred, axis=-1)
updates = []
for i, iou in enumerate(self.ious):
updates.append(
iou.update_state(
tf.equal(y_true, i), tf.equal(y_pred, i), sample_weight
)
)
return tf.group(tf.nest.flatten(updates))
def call_metric_function(metric_fn, y_true, y_pred, weights=None, mask=None):
"""Invokes metric function and returns the metric result tensor."""
if mask is None:
return metric_fn(y_true, y_pred, sample_weight=weights)
mask = math_ops.cast(mask, y_pred.dtype)
if weights is None:
# Use mask as sample weight.
return metric_fn(y_true, y_pred, sample_weight=mask)
# Update dimensions of weights to match with mask.
mask, _, weights = squeeze_or_expand_dimensions(mask, None, weights)
weights *= mask
return metric_fn(y_true, y_pred, sample_weight=weights)
def call_metric_function(metric_fn, y_true, y_pred, weights=None, mask=None):
"""Invokes metric function and returns the metric result tensor."""
if mask is None:
return metric_fn(y_true, y_pred, sample_weight=weights)
mask = math_ops.cast(mask, y_pred.dtype)
if weights is None:
# Use mask as sample weight.
return metric_fn(y_true, y_pred, sample_weight=mask)
# Update dimensions of weights to match with mask.
mask, _, weights = squeeze_or_expand_dimensions(mask, None, weights)
weights *= mask
return metric_fn(y_true, y_pred, sample_weight=weights)
def update_state(self, y_true, y_pred, sample_weight=None):
y_pred, y_true, sample_weight = squeeze_or_expand_dimensions(
y_pred, y_true, sample_weight
)
if self.threshold is None:
assert y_pred.dtype.is_bool
else:
assert y_pred.dtype.is_floating
y_pred = tf.greater(y_pred, self.threshold)
y_true = tf.cast(y_true, tf.bool)
intersection = tf.cast(tf.logical_and(y_true, y_pred), self._dtype)
union = tf.cast(tf.logical_or(y_true, y_pred), self._dtype)
if sample_weight is not None and not self.ignore_weights:
intersection = intersection * sample_weight
union = union * sample_weight
intersection_update = self.intersection.assign_add(tf.reduce_sum(intersection))
union_update = self.union.assign_add(tf.reduce_sum(union))
return tf.group([intersection_update, union_update])
def _model_loss(model,
inputs,
targets,
output_loss_metrics=None,
sample_weights=None,
training=False):
"""Calculates the loss for a given model.
Args:
model: The model on which metrics are being calculated.
inputs: Either a dictionary of inputs to the model or a list of input
arrays.
targets: List of target arrays.
output_loss_metrics: List of metrics that are used to aggregated output
loss values.
sample_weights: Optional list of sample weight arrays.
training: Whether the model should be run in inference or training mode.
Returns:
Returns the model output, total loss, loss value calculated using the
specified loss function and masks for each output. The total loss includes
regularization losses and applies masking and sample weighting
to the loss value.
"""
# TODO(psv): Dedup code here with graph mode prepare_total_loss() fn.
# Used to keep track of the total loss value (stateless).
# eg., total_loss = loss_weight_1 * output_1_loss_fn(...) +
# loss_weight_2 * output_2_loss_fn(...) +
# layer losses.
total_loss = 0
kwargs = {}
if model._expects_training_arg:
kwargs['training'] = training
if len(inputs) == 1 and not isinstance(inputs, dict):
inputs = inputs[0]
# Allow mixed `NumPy` and `EagerTensor` input here.
if any(
isinstance(input_t, (np.ndarray, float, int))
for input_t in nest.flatten(inputs)):
inputs = nest.map_structure(ops.convert_to_tensor_v2_with_dispatch,
inputs)
outs = model(inputs, **kwargs)
outs = nest.flatten(outs)
if targets:
targets = training_utils_v1.cast_if_floating_dtype_and_mismatch(
targets, outs)
# TODO(sallymatson/psv): check if we should do same mismatch fix for weights
if sample_weights:
sample_weights = [
training_utils_v1.cast_if_floating_dtype(
ops.convert_to_tensor_v2_with_dispatch(val))
if val is not None else None for val in sample_weights
]
masks = [getattr(t, '_keras_mask', None) for t in outs]
targets = nest.flatten(targets)
# Used to keep track of individual output losses.
output_losses = []
with backend.name_scope('loss'):
loss_fns = [
loss_fn for loss_fn in model.loss_functions if loss_fn is not None
]
custom_losses = model.losses # Regularization losses
if not loss_fns and not custom_losses:
if training:
raise ValueError('The model cannot be trained '
'because it has no loss to optimize.')
else:
raise ValueError('The model cannot be evaluated '
'because it has no loss to compute.')
for i, loss_fn in enumerate(loss_fns):
weights = sample_weights[i] if sample_weights else None
mask = masks[i]
with backend.name_scope(model.output_names[i] + '_loss'):
if mask is not None:
mask = math_ops.cast(mask, outs[i].dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
weights = math_ops.cast(weights, outs[i].dtype)
mask, _, weights = (
losses_utils.squeeze_or_expand_dimensions(
mask, sample_weight=weights))
weights *= mask
if hasattr(loss_fn, 'reduction'):
per_sample_losses = loss_fn.call(targets[i], outs[i])
weighted_losses = losses_utils.compute_weighted_loss(
per_sample_losses,
sample_weight=weights,
reduction=losses_utils.ReductionV2.NONE)
loss_reduction = loss_fn.reduction
# `AUTO` loss reduction defaults to `SUM_OVER_BATCH_SIZE` for all
# compile use cases.
if loss_reduction == losses_utils.ReductionV2.AUTO:
loss_reduction = losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE
# Compute the stateless loss value.
output_loss = losses_utils.reduce_weighted_loss(
weighted_losses, reduction=loss_reduction)
else:
# Compute the stateless loss value for a custom loss class.
# Here we assume that the class takes care of loss reduction
# because if this class returns a vector value we cannot
# differentiate between use case where a custom optimizer
# expects a vector loss value vs unreduced per-sample loss value.
output_loss = loss_fn(targets[i],
outs[i],
sample_weight=weights)
loss_reduction = losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE
# If the number of outputs is 1 then we don't append the loss metric
# associated with each model output. When there are multiple outputs
# associated with a model, each output's loss is calculated and returned
# as part of the loss_metrics.
if len(model.outputs) > 1:
# Keep track of the stateful output loss result.
output_losses.append(output_loss_metrics[i](output_loss))
# Scale output loss for distribution. For custom losses we assume
# reduction was mean.
if loss_reduction == losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE:
output_loss = losses_utils.scale_loss_for_distribution(
output_loss)
total_loss += model._loss_weights_list[i] * output_loss
# Add regularization losses
if custom_losses:
total_loss += losses_utils.scale_loss_for_distribution(
math_ops.add_n(custom_losses))
return outs, total_loss, output_losses, masks
def update_confusion_matrix_variables(variables_to_update,
y_true,
y_pred,
thresholds,
top_k=None,
class_id=None,
sample_weight=None,
multi_label=False,
label_weights=None,
thresholds_distributed_evenly=False):
"""Returns op to update the given confusion matrix variables.
For every pair of values in y_true and y_pred:
true_positive: y_true == True and y_pred > thresholds
false_negatives: y_true == True and y_pred <= thresholds
true_negatives: y_true == False and y_pred <= thresholds
false_positive: y_true == False and y_pred > thresholds
The results will be weighted and added together. When multiple thresholds are
provided, we will repeat the same for every threshold.
For estimation of these metrics over a stream of data, the function creates an
`update_op` operation that updates the given variables.
If `sample_weight` is `None`, weights default to 1.
Use weights of 0 to mask values.
Args:
variables_to_update: Dictionary with 'tp', 'fn', 'tn', 'fp' as valid keys
and corresponding variables to update as values.
y_true: A `Tensor` whose shape matches `y_pred`. Will be cast to `bool`.
y_pred: A floating point `Tensor` of arbitrary shape and whose values are in
the range `[0, 1]`.
thresholds: A float value, float tensor, python list, or tuple of float
thresholds in `[0, 1]`, or NEG_INF (used when top_k is set).
top_k: Optional int, indicates that the positive labels should be limited to
the top k predictions.
class_id: Optional int, limits the prediction and labels to the class
specified by this argument.
sample_weight: Optional `Tensor` whose rank is either 0, or the same rank as
`y_true`, and must be broadcastable to `y_true` (i.e., all dimensions must
be either `1`, or the same as the corresponding `y_true` dimension).
multi_label: Optional boolean indicating whether multidimensional
prediction/labels should be treated as multilabel responses, or flattened
into a single label. When True, the valus of `variables_to_update` must
have a second dimension equal to the number of labels in y_true and
y_pred, and those tensors must not be RaggedTensors.
label_weights: (optional) tensor of non-negative weights for multilabel
data. The weights are applied when calculating TP, FP, FN, and TN without
explicit multilabel handling (i.e. when the data is to be flattened).
thresholds_distributed_evenly: Boolean, whether the thresholds are evenly
distributed within the list. An optimized method will be used if this is
the case. See _update_confusion_matrix_variables_optimized() for more
details.
Returns:
Update op.
Raises:
ValueError: If `y_pred` and `y_true` have mismatched shapes, or if
`sample_weight` is not `None` and its shape doesn't match `y_pred`, or if
`variables_to_update` contains invalid keys.
"""
if multi_label and label_weights is not None:
raise ValueError('`label_weights` for multilabel data should be handled '
'outside of `update_confusion_matrix_variables` when '
'`multi_label` is True.')
if variables_to_update is None:
return
if not any(
key for key in variables_to_update if key in list(ConfusionMatrix)):
raise ValueError(
'Please provide at least one valid confusion matrix '
'variable to update. Valid variable key options are: "{}". '
'Received: "{}"'.format(
list(ConfusionMatrix), variables_to_update.keys()))
variable_dtype = list(variables_to_update.values())[0].dtype
y_true = math_ops.cast(y_true, dtype=variable_dtype)
y_pred = math_ops.cast(y_pred, dtype=variable_dtype)
if thresholds_distributed_evenly:
# Check whether the thresholds has any leading or tailing epsilon added
# for floating point imprecision. The leading and tailing threshold will be
# handled bit differently as the corner case.
# At this point, thresholds should be a list/array with more than 2 items,
# and ranged between [0, 1]. See is_evenly_distributed_thresholds() for more
# details.
thresholds_with_epsilon = thresholds[0] < 0.0 or thresholds[-1] > 1.0
thresholds = ops.convert_to_tensor_v2_with_dispatch(
thresholds, dtype=variable_dtype)
num_thresholds = thresholds.shape.as_list()[0]
if multi_label:
one_thresh = math_ops.equal(
math_ops.cast(1, dtype=dtypes.int32),
array_ops.rank(thresholds),
name='one_set_of_thresholds_cond')
else:
[y_pred,
y_true], _ = ragged_assert_compatible_and_get_flat_values([y_pred, y_true],
sample_weight)
one_thresh = math_ops.cast(True, dtype=dtypes.bool)
invalid_keys = [
key for key in variables_to_update if key not in list(ConfusionMatrix)
]
if invalid_keys:
raise ValueError(
'Invalid keys: {}. Valid variable key options are: "{}"'.format(
invalid_keys, list(ConfusionMatrix)))
with ops.control_dependencies([
check_ops.assert_greater_equal(
y_pred,
math_ops.cast(0.0, dtype=y_pred.dtype),
message='predictions must be >= 0'),
check_ops.assert_less_equal(
y_pred,
math_ops.cast(1.0, dtype=y_pred.dtype),
message='predictions must be <= 1')
]):
if sample_weight is None:
y_pred, y_true = losses_utils.squeeze_or_expand_dimensions(
y_pred, y_true)
else:
sample_weight = math_ops.cast(sample_weight, dtype=variable_dtype)
y_pred, y_true, sample_weight = (
losses_utils.squeeze_or_expand_dimensions(
y_pred, y_true, sample_weight=sample_weight))
y_pred.shape.assert_is_compatible_with(y_true.shape)
if top_k is not None:
y_pred = _filter_top_k(y_pred, top_k)
if class_id is not None:
y_true = y_true[..., class_id]
y_pred = y_pred[..., class_id]
if thresholds_distributed_evenly and compat.forward_compatible(2021, 6, 8):
# The new approach will take effect after 2021/6/8, to give enough time
# for Brella release to pick up the new op tf.math.cumsum with float32.
return _update_confusion_matrix_variables_optimized(
variables_to_update, y_true, y_pred, thresholds,
multi_label=multi_label, sample_weights=sample_weight,
label_weights=label_weights,
thresholds_with_epsilon=thresholds_with_epsilon)
pred_shape = array_ops.shape(y_pred)
num_predictions = pred_shape[0]
if y_pred.shape.ndims == 1:
num_labels = 1
else:
num_labels = gen_math_ops.Prod(input=pred_shape[1:], axis=0)
thresh_label_tile = array_ops.where_v2(one_thresh, num_labels,
array_ops.ones([], dtype=dtypes.int32))
# Reshape predictions and labels, adding a dim for thresholding.
if multi_label:
predictions_extra_dim = array_ops.expand_dims(y_pred, 0)
labels_extra_dim = array_ops.expand_dims(
math_ops.cast(y_true, dtype=dtypes.bool), 0)
else:
# Flatten predictions and labels when not multilabel.
predictions_extra_dim = array_ops.reshape(y_pred, [1, -1])
labels_extra_dim = array_ops.reshape(
math_ops.cast(y_true, dtype=dtypes.bool), [1, -1])
# Tile the thresholds for every prediction.
if multi_label:
thresh_pretile_shape = [num_thresholds, 1, -1]
thresh_tiles = [1, num_predictions, thresh_label_tile]
data_tiles = [num_thresholds, 1, 1]
else:
thresh_pretile_shape = [num_thresholds, -1]
thresh_tiles = [1, num_predictions * num_labels]
data_tiles = [num_thresholds, 1]
thresh_tiled = array_ops.tile(
array_ops.reshape(thresholds, thresh_pretile_shape),
array_ops.stack(thresh_tiles))
# Tile the predictions for every threshold.
preds_tiled = array_ops.tile(predictions_extra_dim, data_tiles)
# Compare predictions and threshold.
pred_is_pos = math_ops.greater(preds_tiled, thresh_tiled)
# Tile labels by number of thresholds
label_is_pos = array_ops.tile(labels_extra_dim, data_tiles)
if sample_weight is not None:
sample_weight = weights_broadcast_ops.broadcast_weights(
math_ops.cast(sample_weight, dtype=variable_dtype), y_pred)
weights_tiled = array_ops.tile(
array_ops.reshape(sample_weight, thresh_tiles), data_tiles)
else:
weights_tiled = None
if label_weights is not None and not multi_label:
label_weights = array_ops.expand_dims(label_weights, 0)
label_weights = weights_broadcast_ops.broadcast_weights(label_weights,
y_pred)
label_weights_tiled = array_ops.tile(
array_ops.reshape(label_weights, thresh_tiles), data_tiles)
if weights_tiled is None:
weights_tiled = label_weights_tiled
else:
weights_tiled = math_ops.multiply(weights_tiled, label_weights_tiled)
update_ops = []
def weighted_assign_add(label, pred, weights, var):
label_and_pred = math_ops.cast(
math_ops.logical_and(label, pred), dtype=var.dtype)
if weights is not None:
label_and_pred *= math_ops.cast(weights, dtype=var.dtype)
return var.assign_add(math_ops.reduce_sum(label_and_pred, 1))
loop_vars = {
ConfusionMatrix.TRUE_POSITIVES: (label_is_pos, pred_is_pos),
}
update_tn = ConfusionMatrix.TRUE_NEGATIVES in variables_to_update
update_fp = ConfusionMatrix.FALSE_POSITIVES in variables_to_update
update_fn = ConfusionMatrix.FALSE_NEGATIVES in variables_to_update
if update_fn or update_tn:
pred_is_neg = math_ops.logical_not(pred_is_pos)
loop_vars[ConfusionMatrix.FALSE_NEGATIVES] = (label_is_pos, pred_is_neg)
if update_fp or update_tn:
label_is_neg = math_ops.logical_not(label_is_pos)
loop_vars[ConfusionMatrix.FALSE_POSITIVES] = (label_is_neg, pred_is_pos)
if update_tn:
loop_vars[ConfusionMatrix.TRUE_NEGATIVES] = (label_is_neg, pred_is_neg)
for matrix_cond, (label, pred) in loop_vars.items():
if matrix_cond in variables_to_update:
update_ops.append(
weighted_assign_add(label, pred, weights_tiled,
variables_to_update[matrix_cond]))
return control_flow_ops.group(update_ops)
def _model_loss(model,
inputs,
targets,
output_loss_metrics=None,
sample_weights=None,
training=False):
"""Calculates the loss for a given model.
Arguments:
model: The model on which metrics are being calculated.
inputs: Either a dictionary of inputs to the model or a list of input
arrays.
targets: List of target arrays.
output_loss_metrics: List of metrics that are used to aggregated output
loss values.
sample_weights: Optional list of sample weight arrays.
training: Whether the model should be run in inference or training mode.
Returns:
Returns the model output, total loss, loss value calculated using the
specified loss function and masks for each output. The total loss includes
regularization losses and applies masking and sample weighting
to the loss value.
"""
total_loss = 0
kwargs = {}
if model._expects_training_arg:
kwargs['training'] = training
if len(inputs) == 1 and not isinstance(inputs, dict):
inputs = inputs[0]
if model._compute_output_and_mask_jointly:
outs, masks = model._call_and_compute_mask(inputs, **kwargs)
masks = generic_utils.to_list(masks)
else:
outs = model.call(inputs, **kwargs)
masks = None
outs = generic_utils.to_list(outs)
if masks is None:
masks = [None for _ in outs]
targets = generic_utils.to_list(targets)
loss_metrics = []
aggregated_loss_metrics = []
with backend.name_scope('loss'):
for i, loss_fn in enumerate(model.loss_functions):
if sample_weights:
weights = sample_weights[i]
else:
weights = None
mask = masks[i]
with backend.name_scope(model.output_names[i] + '_loss'):
if isinstance(loss_fn, losses_module.Loss):
if mask is not None:
mask = math_ops.cast(mask, outs[i].dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = squeeze_or_expand_dimensions(
mask, None, weights)
weights *= mask
output_loss = loss_fn(targets[i],
outs[i],
sample_weight=weights)
else:
weighted_masked_fn = training_utils.weighted_masked_objective(
loss_fn)
output_loss = weighted_masked_fn(targets[i],
outs[i],
weights,
mask=mask)
# If the number of outputs is 1 then we don't append the loss metric
# associated with each model output. When there are multiple outputs
# associated with a model, each output's loss is calculated and returned
# as part of the loss_metrics.
if len(model.outputs) > 1:
loss_metrics.append(backend.mean(output_loss))
if output_loss_metrics is not None:
# Keep track of the stateful loss result.
aggregated_loss_metrics.append(
training_utils.call_metric_function(
output_loss_metrics[i],
targets[i],
outs[i],
weights=weights,
mask=mask))
loss_weight = model.loss_weights_list[i]
if total_loss is None:
total_loss = loss_weight * output_loss
else:
total_loss += loss_weight * output_loss
total_loss = backend.mean(total_loss)
# Add regularization losses
custom_losses = model.losses
if custom_losses:
total_loss += math_ops.add_n(custom_losses)
model._clear_losses()
return outs, total_loss, loss_metrics, aggregated_loss_metrics, masks
def _model_loss(model,
inputs,
targets,
output_loss_metrics=None,
sample_weights=None,
training=False):
"""Calculates the loss for a given model.
Arguments:
model: The model on which metrics are being calculated.
inputs: Either a dictionary of inputs to the model or a list of input
arrays.
targets: List of target arrays.
output_loss_metrics: List of metrics that are used to aggregated output
loss values.
sample_weights: Optional list of sample weight arrays.
training: Whether the model should be run in inference or training mode.
Returns:
Returns the model output, total loss, loss value calculated using the
specified loss function and masks for each output. The total loss includes
regularization losses and applies masking and sample weighting
to the loss value.
"""
total_loss = 0
kwargs = {}
if model._expects_training_arg:
kwargs['training'] = training
if len(inputs) == 1 and not isinstance(inputs, dict):
inputs = inputs[0]
if model._compute_output_and_mask_jointly:
outs, masks = model._call_and_compute_mask(inputs, **kwargs)
masks = nest.flatten(masks)
else:
outs = model.call(inputs, **kwargs)
masks = None
outs = nest.flatten(outs)
if masks is None:
masks = [None for _ in outs]
targets = nest.flatten(targets)
loss_metrics = []
aggregated_loss_metrics = []
with backend.name_scope('loss'):
for i, loss_fn in enumerate(model.loss_functions):
if sample_weights:
weights = sample_weights[i]
else:
weights = None
mask = masks[i]
with backend.name_scope(model.output_names[i] + '_loss'):
if isinstance(loss_fn, losses_module.Loss):
if mask is not None:
mask = math_ops.cast(mask, outs[i].dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = squeeze_or_expand_dimensions(
mask, None, weights)
weights *= mask
output_loss = loss_fn(targets[i], outs[i], sample_weight=weights)
else:
weighted_masked_fn = training_utils.weighted_masked_objective(loss_fn)
output_loss = weighted_masked_fn(
targets[i], outs[i], weights, mask=mask)
# If the number of outputs is 1 then we don't append the loss metric
# associated with each model output. When there are multiple outputs
# associated with a model, each output's loss is calculated and returned
# as part of the loss_metrics.
if len(model.outputs) > 1:
loss_metrics.append(backend.mean(output_loss))
if output_loss_metrics is not None:
# Keep track of the stateful loss result.
aggregated_loss_metrics.append(
training_utils.call_metric_function(
output_loss_metrics[i],
targets[i],
outs[i],
weights=weights,
mask=mask))
loss_weight = model.loss_weights_list[i]
if total_loss is None:
total_loss = loss_weight * output_loss
else:
total_loss += loss_weight * output_loss
total_loss = backend.mean(total_loss)
# Add regularization losses
custom_losses = model.losses
if custom_losses:
total_loss += math_ops.add_n(custom_losses)
model._clear_losses()
return outs, total_loss, loss_metrics, aggregated_loss_metrics, masks
def update_confusion_matrix_variables(variables_to_update,
y_true,
y_pred,
thresholds,
top_k=None,
class_id=None,
sample_weight=None):
"""Returns op to update the given confusion matrix variables.
For every pair of values in y_true and y_pred:
true_positive: y_true == True and y_pred > thresholds
false_negatives: y_true == True and y_pred <= thresholds
true_negatives: y_true == False and y_pred <= thresholds
false_positive: y_true == False and y_pred > thresholds
The results will be weighted and added together. When multiple thresholds are
provided, we will repeat the same for every threshold.
For estimation of these metrics over a stream of data, the function creates an
`update_op` operation that updates the given variables.
If `sample_weight` is `None`, weights default to 1.
Use weights of 0 to mask values.
Args:
variables_to_update: Dictionary with 'tp', 'fn', 'tn', 'fp' as valid keys
and corresponding variables to update as values.
y_true: A `Tensor` whose shape matches `y_pred`. Will be cast to `bool`.
y_pred: A floating point `Tensor` of arbitrary shape and whose values are in
the range `[0, 1]`.
thresholds: A float value or a python list or tuple of float thresholds in
`[0, 1]`, or NEG_INF (used when top_k is set).
top_k: Optional int, indicates that the positive labels should be limited to
the top k predictions.
class_id: Optional int, limits the prediction and labels to the class
specified by this argument.
sample_weight: Optional `Tensor` whose rank is either 0, or the same rank as
`y_true`, and must be broadcastable to `y_true` (i.e., all dimensions must
be either `1`, or the same as the corresponding `y_true` dimension).
Returns:
Update op.
Raises:
ValueError: If `y_pred` and `y_true` have mismatched shapes, or if
`sample_weight` is not `None` and its shape doesn't match `y_pred`, or if
`variables_to_update` contains invalid keys.
"""
if variables_to_update is None:
return
y_true = ops.convert_to_tensor(y_true)
y_pred = ops.convert_to_tensor(y_pred)
y_pred.shape.assert_is_compatible_with(y_true.shape)
if not any(
key for key in variables_to_update if key in list(ConfusionMatrix)):
raise ValueError(
'Please provide at least one valid confusion matrix '
'variable to update. Valid variable key options are: "{}". '
'Received: "{}"'.format(
list(ConfusionMatrix), variables_to_update.keys()))
invalid_keys = [
key for key in variables_to_update if key not in list(ConfusionMatrix)
]
if invalid_keys:
raise ValueError(
'Invalid keys: {}. Valid variable key options are: "{}"'.format(
invalid_keys, list(ConfusionMatrix)))
with ops.control_dependencies([
check_ops.assert_greater_equal(
y_pred,
math_ops.cast(0.0, dtype=y_pred.dtype),
message='predictions must be >= 0'),
check_ops.assert_less_equal(
y_pred,
math_ops.cast(1.0, dtype=y_pred.dtype),
message='predictions must be <= 1')
]):
y_pred, y_true, sample_weight = squeeze_or_expand_dimensions(
math_ops.cast(y_pred, dtype=dtypes.float32),
math_ops.cast(y_true, dtype=dtypes.bool), sample_weight)
if top_k is not None:
y_pred = _filter_top_k(y_pred, top_k)
if class_id is not None:
y_true = y_true[..., class_id]
y_pred = y_pred[..., class_id]
thresholds = to_list(thresholds)
num_thresholds = len(thresholds)
num_predictions = array_ops.size(y_pred)
# Reshape predictions and labels.
predictions_2d = array_ops.reshape(y_pred, [1, -1])
labels_2d = array_ops.reshape(
math_ops.cast(y_true, dtype=dtypes.bool), [1, -1])
# Tile the thresholds for every prediction.
thresh_tiled = array_ops.tile(
array_ops.expand_dims(array_ops.constant(thresholds), 1),
array_ops.stack([1, num_predictions]))
# Tile the predictions for every threshold.
preds_tiled = array_ops.tile(predictions_2d, [num_thresholds, 1])
# Compare predictions and threshold.
pred_is_pos = math_ops.greater(preds_tiled, thresh_tiled)
# Tile labels by number of thresholds
label_is_pos = array_ops.tile(labels_2d, [num_thresholds, 1])
if sample_weight is not None:
weights = weights_broadcast_ops.broadcast_weights(
math_ops.cast(sample_weight, dtype=dtypes.float32), y_pred)
weights_tiled = array_ops.tile(
array_ops.reshape(weights, [1, -1]), [num_thresholds, 1])
else:
weights_tiled = None
update_ops = []
def weighted_assign_add(label, pred, weights, var):
label_and_pred = math_ops.cast(
math_ops.logical_and(label, pred), dtype=dtypes.float32)
if weights is not None:
label_and_pred *= weights
return state_ops.assign_add(var, math_ops.reduce_sum(label_and_pred, 1))
loop_vars = {
ConfusionMatrix.TRUE_POSITIVES: (label_is_pos, pred_is_pos),
}
update_tn = ConfusionMatrix.TRUE_NEGATIVES in variables_to_update
update_fp = ConfusionMatrix.FALSE_POSITIVES in variables_to_update
update_fn = ConfusionMatrix.FALSE_NEGATIVES in variables_to_update
if update_fn or update_tn:
pred_is_neg = math_ops.logical_not(pred_is_pos)
loop_vars[ConfusionMatrix.FALSE_NEGATIVES] = (label_is_pos, pred_is_neg)
if update_fp or update_tn:
label_is_neg = math_ops.logical_not(label_is_pos)
loop_vars[ConfusionMatrix.FALSE_POSITIVES] = (label_is_neg, pred_is_pos)
if update_tn:
loop_vars[ConfusionMatrix.TRUE_NEGATIVES] = (label_is_neg, pred_is_neg)
for matrix_cond, (label, pred) in loop_vars.items():
if matrix_cond in variables_to_update:
update_ops.append(
weighted_assign_add(label, pred, weights_tiled,
variables_to_update[matrix_cond]))
return control_flow_ops.group(update_ops)
def _model_loss(model,
inputs,
targets,
output_loss_metrics=None,
sample_weights=None,
training=False):
"""Calculates the loss for a given model.
Arguments:
model: The model on which metrics are being calculated.
inputs: Either a dictionary of inputs to the model or a list of input
arrays.
targets: List of target arrays.
output_loss_metrics: List of metrics that are used to aggregated output
loss values.
sample_weights: Optional list of sample weight arrays.
training: Whether the model should be run in inference or training mode.
Returns:
Returns the model output, total loss, loss value calculated using the
specified loss function and masks for each output. The total loss includes
regularization losses and applies masking and sample weighting
to the loss value.
"""
# Used to keep track of the total loss value (stateless).
# eg., total_loss = loss_weight_1 * output_1_loss_fn(...) +
# loss_weight_2 * output_2_loss_fn(...) +
# layer losses.
total_loss = 0
kwargs = {}
if model._expects_training_arg:
kwargs['training'] = training
if len(inputs) == 1 and not isinstance(inputs, dict):
inputs = inputs[0]
# Allow mixed `NumPy` and `EagerTensor` input here.
if any(
isinstance(input_t, (np.ndarray, float, int))
for input_t in nest.flatten(inputs)):
inputs = nest.map_structure(ops.convert_to_tensor, inputs)
outs = model(inputs, **kwargs)
outs = nest.flatten(outs)
# `None` by default for `EagerTensors`.
masks = [t._keras_mask for t in outs]
targets = nest.flatten(targets)
# Used to keep track of individual output losses (stateless).
output_losses = []
# Used to keep track of individual output losses (stateful).
aggregated_output_losses = []
with backend.name_scope('loss'):
for i, loss_fn in enumerate(model.loss_functions):
weights = sample_weights[i] if sample_weights else None
mask = masks[i]
with backend.name_scope(model.output_names[i] + '_loss'):
if mask is not None:
mask = math_ops.cast(mask, outs[i].dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = (
losses_utils.squeeze_or_expand_dimensions(
mask, None, weights))
weights *= mask
# Reset reduction on the loss so that we can get the per sample loss
# value. We use this to get both the stateless and stateful loss
# values without having to compute the underlying loss function
# twice.
weighted_losses = None
if hasattr(loss_fn, 'reduction'):
current_loss_reduction = loss_fn.reduction
loss_fn.reduction = losses_utils.ReductionV2.NONE
weighted_losses = loss_fn(targets[i],
outs[i],
sample_weight=weights)
loss_fn.reduction = current_loss_reduction
# Compute the stateless loss value.
output_loss = losses_utils.reduce_weighted_loss(
weighted_losses)
else:
# Compute the stateless loss value for a custom loss class.
# Here we assume that the class takes care of loss reduction
# because if this class returns a vector value we cannot
# differentiate between use case where a custom optimizer
# expects a vector loss value vs unreduced per-sample loss value.
output_loss = loss_fn(targets[i],
outs[i],
sample_weight=weights)
# If the number of outputs is 1 then we don't append the loss metric
# associated with each model output. When there are multiple outputs
# associated with a model, each output's loss is calculated and returned
# as part of the loss_metrics.
if len(model.outputs) > 1:
output_losses.append(backend.mean(output_loss))
if output_loss_metrics is not None:
# Compute the stateful loss value.
if weighted_losses is not None:
aggregated_output_loss = output_loss_metrics[i](
weighted_losses)
else:
# Custom loss class.
aggregated_output_loss = training_utils.call_metric_function(
output_loss_metrics[i],
targets[i],
outs[i],
weights=weights)
# Keep track of the stateful output loss result.
aggregated_output_losses.append(aggregated_output_loss)
loss_weight = model.loss_weights_list[i]
if total_loss is None:
total_loss = loss_weight * output_loss
else:
total_loss += loss_weight * output_loss
total_loss = backend.mean(total_loss)
# Add regularization losses
custom_losses = model.losses
if custom_losses:
total_loss += losses_utils.scale_loss_for_distribution(
math_ops.add_n(custom_losses))
model._clear_losses()
return outs, total_loss, output_losses, aggregated_output_losses, masks
def _model_loss(model,
inputs,
targets,
output_loss_metrics=None,
sample_weights=None,
training=False):
"""Calculates the loss for a given model.
Arguments:
model: The model on which metrics are being calculated.
inputs: Either a dictionary of inputs to the model or a list of input
arrays.
targets: List of target arrays.
output_loss_metrics: List of metrics that are used to aggregated output
loss values.
sample_weights: Optional list of sample weight arrays.
training: Whether the model should be run in inference or training mode.
Returns:
Returns the model output, total loss, loss value calculated using the
specified loss function and masks for each output. The total loss includes
regularization losses and applies masking and sample weighting
to the loss value.
"""
# TODO(psv): Dedup code here with graph mode prepare_total_loss() fn.
# Used to keep track of the total loss value (stateless).
# eg., total_loss = loss_weight_1 * output_1_loss_fn(...) +
# loss_weight_2 * output_2_loss_fn(...) +
# layer losses.
total_loss = 0
kwargs = {}
if model._expects_training_arg:
kwargs['training'] = training
if len(inputs) == 1 and not isinstance(inputs, dict):
inputs = inputs[0]
# Allow mixed `NumPy` and `EagerTensor` input here.
if any(
isinstance(input_t, (np.ndarray, float, int))
for input_t in nest.flatten(inputs)):
inputs = nest.map_structure(ops.convert_to_tensor, inputs)
outs = model(inputs, **kwargs)
outs = nest.flatten(outs)
masks = [getattr(t, '_keras_mask', None) for t in outs]
targets = nest.flatten(targets)
# Used to keep track of individual output losses.
output_losses = []
with backend.name_scope('loss'):
loss_fns = [
loss_fn for loss_fn in model.loss_functions if loss_fn is not None
]
for i, loss_fn in enumerate(loss_fns):
weights = sample_weights[i] if sample_weights else None
mask = masks[i]
with backend.name_scope(model.output_names[i] + '_loss'):
if mask is not None:
mask = math_ops.cast(mask, outs[i].dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = (
losses_utils.squeeze_or_expand_dimensions(mask, None, weights))
weights *= mask
weighted_losses = None
if hasattr(loss_fn, 'reduction'):
per_sample_losses = loss_fn.call(targets[i], outs[i])
weighted_losses = losses_utils.compute_weighted_loss(
per_sample_losses,
sample_weight=weights,
reduction=losses_utils.ReductionV2.NONE)
loss_reduction = loss_fn.reduction
# `AUTO` loss reduction defaults to `SUM_OVER_BATCH_SIZE` for all
# compile use cases.
if loss_reduction == losses_utils.ReductionV2.AUTO:
loss_reduction = losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE
# Compute the stateless loss value.
output_loss = losses_utils.reduce_weighted_loss(
weighted_losses, reduction=loss_reduction)
if loss_reduction == losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE:
output_loss = losses_utils.scale_loss_for_distribution(output_loss)
else:
# Compute the stateless loss value for a custom loss class.
# Here we assume that the class takes care of loss reduction
# because if this class returns a vector value we cannot
# differentiate between use case where a custom optimizer
# expects a vector loss value vs unreduced per-sample loss value.
output_loss = loss_fn(targets[i], outs[i], sample_weight=weights)
# For custom losses we assume reduction was mean.
output_loss = losses_utils.scale_loss_for_distribution(output_loss)
# If the number of outputs is 1 then we don't append the loss metric
# associated with each model output. When there are multiple outputs
# associated with a model, each output's loss is calculated and returned
# as part of the loss_metrics.
if len(model.outputs) > 1:
# Keep track of the stateful output loss result.
output_losses.append(output_loss_metrics[i](output_loss))
total_loss += model._loss_weights_list[i] * output_loss
# Add regularization losses
custom_losses = model.losses
if custom_losses:
total_loss += losses_utils.scale_loss_for_distribution(
math_ops.add_n(custom_losses))
return outs, total_loss, output_losses, masks
def _model_loss(model,
inputs,
targets,
output_loss_metrics=None,
sample_weights=None,
training=False):
"""Calculates the loss for a given model.
Arguments:
model: The model on which metrics are being calculated.
inputs: Either a dictionary of inputs to the model or a list of input
arrays.
targets: List of target arrays.
output_loss_metrics: List of metrics that are used to aggregated output
loss values.
sample_weights: Optional list of sample weight arrays.
training: Whether the model should be run in inference or training mode.
Returns:
Returns the model output, total loss, loss value calculated using the
specified loss function and masks for each output. The total loss includes
regularization losses and applies masking and sample weighting
to the loss value.
"""
# Used to keep track of the total loss value (stateless).
# eg., total_loss = loss_weight_1 * output_1_loss_fn(...) +
# loss_weight_2 * output_2_loss_fn(...) +
# layer losses.
total_loss = 0
kwargs = {}
if model._expects_training_arg:
kwargs['training'] = training
if len(inputs) == 1 and not isinstance(inputs, dict):
inputs = inputs[0]
# Allow mixed `NumPy` and `EagerTensor` input here.
if any(
isinstance(input_t, (np.ndarray, float, int))
for input_t in nest.flatten(inputs)):
inputs = nest.map_structure(ops.convert_to_tensor, inputs)
outs = model(inputs, **kwargs)
outs = nest.flatten(outs)
# `None` by default for `EagerTensors`.
masks = [t._keras_mask for t in outs]
targets = nest.flatten(targets)
# Used to keep track of individual output losses (stateless).
output_losses = []
# Used to keep track of individual output losses (stateful).
aggregated_output_losses = []
with backend.name_scope('loss'):
for i, loss_fn in enumerate(model.loss_functions):
weights = sample_weights[i] if sample_weights else None
mask = masks[i]
with backend.name_scope(model.output_names[i] + '_loss'):
if mask is not None:
mask = math_ops.cast(mask, outs[i].dtype)
# Update weights with mask.
if weights is None:
weights = mask
else:
# Update dimensions of weights to match with mask if possible.
mask, _, weights = (
losses_utils.squeeze_or_expand_dimensions(mask, None, weights))
weights *= mask
# Reset reduction on the loss so that we can get the per sample loss
# value. We use this to get both the stateless and stateful loss
# values without having to compute the underlying loss function
# twice.
weighted_losses = None
if hasattr(loss_fn, 'reduction'):
current_loss_reduction = loss_fn.reduction
loss_fn.reduction = losses_utils.ReductionV2.NONE
weighted_losses = loss_fn(targets[i], outs[i], sample_weight=weights)
loss_fn.reduction = current_loss_reduction
# Compute the stateless loss value.
output_loss = losses_utils.reduce_weighted_loss(weighted_losses)
else:
# Compute the stateless loss value for a custom loss class.
# Here we assume that the class takes care of loss reduction
# because if this class returns a vector value we cannot
# differentiate between use case where a custom optimizer
# expects a vector loss value vs unreduced per-sample loss value.
output_loss = loss_fn(targets[i], outs[i], sample_weight=weights)
# If the number of outputs is 1 then we don't append the loss metric
# associated with each model output. When there are multiple outputs
# associated with a model, each output's loss is calculated and returned
# as part of the loss_metrics.
if len(model.outputs) > 1:
output_losses.append(backend.mean(output_loss))
if output_loss_metrics is not None:
# Compute the stateful loss value.
if weighted_losses is not None:
aggregated_output_loss = output_loss_metrics[i](weighted_losses)
else:
# Custom loss class.
aggregated_output_loss = training_utils.call_metric_function(
output_loss_metrics[i], targets[i], outs[i], weights=weights)
# Keep track of the stateful output loss result.
aggregated_output_losses.append(aggregated_output_loss)
loss_weight = model.loss_weights_list[i]
if total_loss is None:
total_loss = loss_weight * output_loss
else:
total_loss += loss_weight * output_loss
total_loss = backend.mean(total_loss)
# Add regularization losses
custom_losses = model.losses
if custom_losses:
total_loss += losses_utils.scale_loss_for_distribution(
math_ops.add_n(custom_losses))
return outs, total_loss, output_losses, aggregated_output_losses, masks
