sup3r.utilities.loss_metrics.LowResLoss

class LowResLoss(s_enhance=1, t_enhance=1, t_method='average', tf_loss='MeanSquaredError', ex_loss=None)[source]

Bases: Loss

Content loss that is calculated by coarsening the synthetic and true high-resolution data pairs and then performing the pointwise content loss on the low-resolution fields

Initialize the loss with given weight

Parameters:

s_enhance (int) – factor by which to coarsen spatial dimensions. 1 will keep the spatial axes as high-res
t_enhance (int) – factor by which to coarsen temporal dimension. 1 will keep the temporal axes as high-res
t_method (str) – Accepted options: [subsample, average] Subsample will take every t_enhance-th time step, average will average over t_enhance time steps
tf_loss (str) – The tensorflow loss function to operate on the low-res fields. Must be the name of a loss class that can be retrieved from tf.keras.losses e.g., “MeanSquaredError” or “MeanAbsoluteError”
ex_loss (None | str) – Optional additional loss metric evaluating the spatial or temporal extremes of the high-res data. Can be “SpatialExtremesOnlyLoss” or “TemporalExtremesOnlyLoss” (keys in EX_LOSS_METRICS).

Methods

`call`(y_true, y_pred)	Invokes the Loss instance.
`from_config`(config)	Instantiates a Loss from its config (output of get_config()).
`get_config`()	Returns the config dictionary for a Loss instance.

Attributes

EX_LOSS_METRICS

__call__(x1, x2)[source]

Custom content loss calculated on re-coarsened low-res fields

Parameters:

x1 (tf.tensor) – Synthetic high-res generator output, shape is either of these: (n_obs, spatial_1, spatial_2, features) (n_obs, spatial_1, spatial_2, temporal, features)
x2 (tf.tensor) – True high resolution data, shape is either of these: (n_obs, spatial_1, spatial_2, features) (n_obs, spatial_1, spatial_2, temporal, features)

Returns:

tf.tensor – 0D tensor loss value

abstract call(y_true, y_pred)

Invokes the Loss instance.

Args:

y_true: Ground truth values. shape = [batch_size, d0, .. dN],: except sparse loss functions such as sparse categorical crossentropy where shape = [batch_size, d0, .. dN-1]

y_pred: The predicted values. shape = [batch_size, d0, .. dN]

Returns:

Loss values with the shape [batch_size, d0, .. dN-1].

classmethod from_config(config)

Instantiates a Loss from its config (output of get_config()).

Args:: config: Output of get_config().
Returns:: A Loss instance.

get_config(): Returns the config dictionary for a Loss instance.