losses

Defines the Multinomial NLL loss and the adaptive counts loss.

bpreveal.losses.multinomialNll(trueCounts, logits)

The heart of what makes BPNet great - the loss function for profiles.

Parameters:

• trueCounts (tensorflow.Tensor) – The experimentally-observed counts. Shape (batch-size x output-length x num-tasks)

• logits (tensorflow.Tensor) – The logits that the model is currently emitting. Shape (batch-size x output-length x num-tasks)

Returns:

A scalar representing the profile loss of this batch.

Return type:

float

bpreveal.losses.weightedMse(weightTensor)

Loss for the adaptive counts loss weight.

Given a weight tensor (a tensorflow Variable of shape (1,)) return a loss function that calculates mean square error and multiplies the error by the weight. This is used to implement the automatic counts weight algorithm.

Parameters:

weightTensor (tensorflow.Tensor) – The tensor that will be adjusted by the dynamic counts loss weight algorithm.

Returns:

A loss function.

bpreveal.losses.dummyMse(yTrue, yPred)

Used for loading models. If you’re going to train, you have to create a weightedMse, since this one has no loss weight to it. But if you’re just predicting, you can pass this function to custom_objects when you load a BPReveal model Always returns nan, so that if you do accidentally use it, all of your values will be poisoned.

Parameters:

• yTrue (tensorflow.Tensor)

• yPred (tensorflow.Tensor)

Return type:

float