def normal_binary_focal_loss_wi(y_true, y_pred):
# taken from https: // focal - loss.readthedocs.io / en / latest / generated / focal_loss.BinaryFocalLoss.html
# also check https://arxiv.org/pdf/1708.02002.pdf
# unpacking labels and weights
target = tf.convert_to_tensor(y_true)
target, weights = tf.split(target, num_or_size_splits=2, axis=-1)
target = tf.dtypes.cast(target, dtype=tf.bool)
weights = tf.keras.backend.cast(weights, tf.float32)
output = tf.convert_to_tensor(y_pred)
output = tf.keras.backend.cast(output, tf.float32)
epsilon_ = tf.convert_to_tensor(epsilon(), tf.float32)
p = output
q = 1 - p
# avoid zeros in p and q --> would cause problems with log(0) later
p = tf.math.maximum(p, epsilon_)
q = tf.math.maximum(q, epsilon_)
# Loss for the positive examples
pos_loss = -(q**gamma) * tf.math.log(p)
# Loss for the negative examples
neg_loss = -(p**gamma) * tf.math.log(q)
# choose either pos_loss or neg_loss, depending on the input label
loss = tf.where(target, pos_loss, neg_loss) * weights
loss = reduce_weighted_loss(loss,
reduction=ReductionV2.SUM_OVER_BATCH_SIZE)
return loss
def _normal_focal_loss(y_true, y_pred):
y_pred = tf.convert_to_tensor(y_pred)
y_pred = tf.dtypes.cast(y_pred, dtype=tf.float32)
y_true = tf.convert_to_tensor(y_true)
y_true = tf.dtypes.cast(y_true, dtype=tf.bool)
epsilon_ = tf.convert_to_tensor(epsilon(), tf.float32)
p = y_pred
q = 1 - p
# avoid zeros in p and q --> would cause problems with log(0) later
p = tf.math.maximum(p, epsilon_)
q = tf.math.maximum(q, epsilon_)
# Loss for the positive examples
pos_loss = -(q**gamma) * tf.math.log(p)
# Loss for the negative examples
neg_loss = -(p**gamma) * tf.math.log(q)
# choose either pos_loss or neg_loss, depending on the input label
loss = tf.where(y_true, pos_loss, neg_loss)
loss = reduce_weighted_loss(loss,
reduction=ReductionV2.SUM_OVER_BATCH_SIZE)
return loss
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 _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 wrapper(*args, **kwargs):
loss_tensor = function(*args, **kwargs)
return reduce_weighted_loss(loss_tensor)
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
