sgdrf.subsample
Classes for subsampling strategies.
Module Contents
- class sgdrf.subsample.ExponentialSubsampler(n: int, device: torch.device, exp: float)
A subsampling strategy that has an exponential decay, with larger probabilities being more recent.
- Parameters:
n (int) – Number of subsamples to draw
device (torch.device) – Pytorch device to use
exp (float) – Exponential parameter
- n
Number of subsamples to draw
- Type:
int
- device
Pytorch device to use
- Type:
torch.device
- exp
Exponential parameter
- Type:
float
- dist(t: int) torch.Tensor
Get distribution for this subsampler, for t observations.
- Parameters:
t (int) – The number of observations to get a distribution for
- Returns:
Tensor of topic probabilities of length t
- Return type:
torch.Tensor
- class sgdrf.subsample.LatestSubsampler(n: int, device: torch.device)
A subsampling strategy that samples the n most recent observations with probability 1.
- Parameters:
n (int) – Number of subsamples to draw
device (torch.device) – Pytorch device to use
- n
Number of subsamples to draw
- Type:
int
- device
Pytorch device to use
- Type:
torch.device
- dist(t: int) torch.Tensor
Get distribution for this subsampler, for t observations.
- Parameters:
t (int) – The number of observations to get a distribution for
- Returns:
Tensor of topic probabilities of length t
- Return type:
torch.Tensor
- class sgdrf.subsample.MixingSubsampler(n: int, device: torch.device, weight: float, s1: Subsampler, s2: Subsampler)
A subsampling strategy that mixes two other subsampling strategies with a given weight.
- Parameters:
n (int) – Number of subsamples to draw
device (torch.device) – Pytorch device to use
weight (float) – Weight associated with subsampler s1 (s2 is associated with 1 - weight)
s1 (Subsampler) – The first subsampler to mix
s2 (Subsampler) – The second subsampler to mix
- n
Number of subsamples to draw
- Type:
int
- device
Pytorch device to use
- Type:
torch.device
- weight
Weight associated with subsampler s1 (s2 is associated with 1 - weight)
- Type:
float
- s1
The first subsampler to mix
- Type:
- s2
The second subsampler to mix
- Type:
- dist(t: int)
Get distribution for this subsampler, for t observations.
- Parameters:
t (int) – The number of observations to get a distribution for
- Returns:
Tensor of topic probabilities of length t
- Return type:
torch.Tensor
- class sgdrf.subsample.Subsampler(n: int, device: torch.device)
An abstract class for SGDRF subsampling strategies.
- Parameters:
n (int) – Number of subsamples to draw
device (torch.device) – Pytorch device to use
- n
Number of subsamples to draw
- Type:
int
- device
Pytorch device to use
- Type:
torch.device
- abstract dist(t: int) torch.Tensor
Get distribution for this subsampler, for t observations.
- Parameters:
t (int) – The number of observations to get a distribution for
- Returns:
Tensor of topic probabilities of length t
- Return type:
torch.Tensor
- subsample(t: int) torch.Tensor
Generate a subsample, for t observations.
- Parameters:
t (int) – The number of observations to get a subsample for
- Returns:
Tensor of subsample
- Return type:
torch.Tensor
- class sgdrf.subsample.UniformSubsampler(n: int, device: torch.device)
A subsampling strategy that has uniform values for all observations.
- Parameters:
n (int) – Number of subsamples to draw
device (torch.device) – Pytorch device to use
- n
Number of subsamples to draw
- Type:
int
- device
Pytorch device to use
- Type:
torch.device
- dist(t: int) torch.Tensor
Get distribution for this subsampler, for t observations.
- Parameters:
t (int) – The number of observations to get a distribution for
- Returns:
Tensor of topic probabilities of length t
- Return type:
torch.Tensor