OakEmu-RL v0.1.0

Appearance

LayerActivationStatistics ​

python
LayerActivationStatistics(name: str, producer: str, units: int, dead: int, near_dead: int | None, effective_rank: float, mean_absolute: float) -> None

How one activation layer behaves over a batch of real observations.

Source

Properties ​

dead_fraction ​

python
dead_fraction: float

Dead units as a fraction of all units.

Source

near_dead_fraction ​

python
near_dead_fraction: float | None

Near-dead units as a fraction of all units, or None where near-dead has no meaning.

Source

rank_ratio ​

python
rank_ratio: float

Effective rank as a fraction of the units, that is how much of the layer's width is actually in use.

Source

Values ​

dead ​

python
dead: int

Units whose output never changed across the batch. A ReLU stuck at zero, or a saturated tanh, counts here.

Source

effective_rank ​

python
effective_rank: float

Effective rank of the activations (see mrkwatkins.oakemu_rl.framework.diagnostics.rank.effective_rank).

Source

mean_absolute ​

python
mean_absolute: float

Mean absolute activation, for scale.

Source

name ​

python
name: str

The activation module's path within the model, for example conv.1.

Source

near_dead ​

python
near_dead: int | None

Units active (output above zero) on fewer than the threshold fraction of inputs. None for a bounded activation, where it has no meaning.

Source

producer ​

python
producer: str

The layer feeding the activation, for example Conv2d(8, 32, kernel_size=(5, 5), stride=(2, 2)).

Source

units ​

python
units: int

Units measured. Channels for a convolutional output, features for a linear one.

Source