junn.common package

Submodules

junn.common.callbacks module

junn.common.distributed module

Helper functions for distributed training.

junn.common.distributed.barrier(name)

Create a barrier which will synchronize all processes.

Can be used if the worker processes need to wait for the main process.

Parameters

name –

Returns

junn.common.distributed.get_callbacks()

Get Keras callbacks.

If Horovod is present, this will be the BroadCastGlobalVariables callback.

Returns

junn.common.distributed.init(device_pinning=True)

Initialize Horovod if present.

Parameters

device_pinning – Whether GPUs should be pinned.

Returns

junn.common.distributed.is_rank_zero()

Return whether the current process is rank zero, i.e. the main process.

Returns

junn.common.distributed.is_running_distributed()

Return True if JUNN is running distributed and Horovod is initialized.

Returns

junn.common.distributed.local_rank()

Return the local rank (on the physical machine).

Returns

junn.common.distributed.pin_devices()

Pin each GPU to a worker.

Returns

junn.common.distributed.rank()

Return the global rank.

This is the rank within all running instances on possibly multiple machines.

Returns

junn.common.distributed.size()

Return the count of running instances.

Returns

junn.common.distributed.wrap_optimizer(optimizer)

Wrap the Keras optimizer.

If Horovod is present, with DistributedOptimizer.

Parameters

optimizer –

Returns

junn.common.losses module

Additional losses suitable for segmentation tasks.

junn.common.losses.accuracy(y_true, y_pred)

Calculate the accuracy.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

accuracy

junn.common.losses.dice_index(y_true, y_pred)

Calculate the Dice index.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

the Dice index

junn.common.losses.dice_index_direct(y_true, y_pred)

Directly calculate the Dice index.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

Dice index

junn.common.losses.dice_index_weighted(y_true, y_pred, y_weight)

Calculate the Dice index with weighting applied.

Parameters

  • y_true – ground truth

  • y_pred – prediction

  • y_weight – weights

Returns

weighted Dice index

junn.common.losses.dice_loss(y_true, y_pred)

Calculate the Dice loss, i.e. the negative Dice index.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

Dice loss

junn.common.losses.dice_loss_unclipped(y_true, y_pred)

Calculate the Dice loss, i.e. the negative Dice index without clipping.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

Dice loss

junn.common.losses.dice_loss_weighted(y_true, y_pred, y_weight)

Calculate the weighted Dice loss, i.e. negative Dice index.

Parameters

  • y_true – ground truth

  • y_pred – prediction

  • y_weight – weights

Returns

weighted Dice loss

junn.common.losses.epsilon()

Epsilon of the current compute hardware.

Returns

epsilon as a floating point type

junn.common.losses.f_score(y_true, y_pred)

Calculate the F score.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

F-score

junn.common.losses.false_negative(y_true, y_pred)
junn.common.losses.false_positive(y_true, y_pred)
junn.common.losses.flatten_and_clip(values)

Flatten and clip a tensor.

Parameters

values – values

Returns

flattened and clipped to 0-1 values

junn.common.losses.generate_tversky_loss(alpha=0.5, beta=0.5)

Generate a Tversky loss function.

Parameters

  • alpha – alpha

  • beta – beta

Returns

a Tversky loss function with fixed parameters

junn.common.losses.jaccard_index(y_true, y_pred)

Calculate the Jaccard index.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

the Jaccard index

junn.common.losses.mixin_flatten_and_clip(what)

Mixin to assure a function gets a flattened and clipped tensor.

Parameters

what – function to be modified

Returns

modified function

junn.common.losses.nan_loss(y_true, y_pred)

Return always NaN for debugging and testing purposes.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

all NaN

junn.common.losses.precision(y_true, y_pred)

Calculate the precision.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

precision

junn.common.losses.recall(y_true, y_pred)

Calculate the recall.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

recall

junn.common.losses.tp_tn_fp_fn_precision_recall(y_true, y_pred)

Calculate a set of simple metrics in one go.

True positives, true negatives, false positives, false negatives, precision as well as recall.

Parameters

  • y_true – ground truth

  • y_pred – prediction

Returns

tuple of metrics (tp, tn, fp, fn, precision, recall)

junn.common.losses.true_negative(y_true, y_pred)
junn.common.losses.true_positive(y_true, y_pred)
junn.common.losses.tversky_index(y_true, y_pred, alpha=0.5, beta=0.5)

Calculate the Tversky index.

Parameters

  • y_true – ground truth

  • y_pred – prediction

  • alpha – alpha

  • beta – beta

Returns

the Tversky index

junn.common.losses.weighted_loss(y_true, y_pred, y_weight)

Weighted loss function.

Parameters

  • y_true – ground truth

  • y_pred – prediction

  • y_weight – weights

Returns

weighted loss