Metrics and their meaning¶
During the search phase, AutoAlbument outputs four metrics:
Average Parameter Change (at the end of an epoch).
a_loss is a loss for the policy network (or Generator in terms of GAN), which applies augmentations to input images.
d_loss is a loss for the Discriminator, the network that tries to guess whether the input image is an augmented or non-augmented one.
loss is a task-specific loss (
CrossEntropyLoss for classification,
BCEWithLogitsLoss for semantic segmentation) that acts as a regularizer and prevents the policy network from applying such augmentations that will make an object with class A looks like an object with class B.
Average Parameter Change¶
Average Parameter Change is a difference between magnitudes of augmentation parameters multiplied by their probabilities at the end of an epoch and the same parameters at the beginning of the epoch. The metric is calculated using the following formula:
mare magnitude values for the i-th augmentation at the end and the beginning of the epoch, respectively.
pare probability values for the i-th augmentation at the end and the beginning of the epoch, respectively.
The intuition behind this metric is that at the beginning, augmentation parameters are initialized at random, so they are now optimal and prone to heavy change at each epoch. After some time, these parameters should begin to converge, and they should change less at each epoch.
Examples for metric values¶
Below are TensorBoard logs for AutoAlbument on different datasets. The search was performed using AutoAlbument configs from the examples directory.
As you see, in all these charts,
loss is slightly decreasing at each epoch, and
d_loss could either decrease or increase.
Average Parameter Change is usually large at first epochs, but then it starts to decrease. As a rule of thumb, to decide whether you should stop AutoAlbument search and use the resulting policy, you should check that
Average Parameter Change is stopped decreasing and started to oscillate, wait for a few more epochs, and use the found policy from that epoch.
In autoalbument-benchmaks, we use AutoAlbument policies produced by the last epoch on these charts.