composer.algorithms.selective_backprop.selective_backprop#
Core SelectiveBackprop class and functions.
Functions
Prunes minibatches as a subroutine of |
|
Decides if selective backprop should be run based on time in training. |
Classes
Selectively backpropagate gradients from a subset of each batch. |
- class composer.algorithms.selective_backprop.selective_backprop.SelectiveBackprop(start=0.5, end=0.9, keep=0.5, scale_factor=1.0, interrupt=2, input_key=0, target_key=1)[source]#
Bases:
composer.core.algorithm.Algorithm
Selectively backpropagate gradients from a subset of each batch.
Based on (Jiang et al, 2019), Selective Backprop (SB) prunes minibatches according to the difficulty of the individual training examples, and only computes weight gradients over the pruned subset, reducing iteration time, and speeding up training.
The fraction of the minibatch that is kept for gradient computation is specified by the argument
0 <= keep <= 1
.To speed up SBโs selection forward pass, the argument
scale_factor
can be used to spatially downsample input image tensors. The full-sized inputs will still be used for the weight gradient computation.To preserve convergence, SB can be interrupted with vanilla minibatch gradient steps every
interrupt
steps. Wheninterrupt=0
, SB will be used at every step during the SB interval. Wheninterrupt=2
, SB will alternate with vanilla minibatch steps.- Parameters
start (float, optional) โ SB interval start as fraction of training duration. Default:
0.5
.end (float, optional) โ SB interval end as fraction of training duration. Default:
0.9
.keep (float, optional) โ fraction of minibatch to select and keep for gradient computation. Default:
0.5
.scale_factor (float, optional) โ scale for downsampling input for selection forward pass. Default:
1.
.interrupt (int, optional) โ interrupt SB with a vanilla minibatch step every
interrupt
batches. Default:2
.input_key (str | int | Tuple[Callable, Callable] | Any, optional) โ A key that indexes to the input from the batch. Can also be a pair of get and set functions, where the getter is assumed to be first in the pair. The default is 0, which corresponds to any sequence, where the first element is the input. Default:
0
.target_key (str | int | Tuple[Callable, Callable] | Any, optional) โ A key that indexes to the target from the batch. Can also be a pair of get and set functions, where the getter is assumed to be first in the pair. The default is 1, which corresponds to any sequence, where the second element is the target. Default:
1
.
Example
from composer.algorithms import SelectiveBackprop algorithm = SelectiveBackprop(start=0.5, end=0.9, keep=0.5) trainer = Trainer( model=model, train_dataloader=train_dataloader, eval_dataloader=eval_dataloader, max_duration="1ep", algorithms=[algorithm], optimizers=[optimizer] )
- composer.algorithms.selective_backprop.selective_backprop.select_using_loss(input, target, model, loss_fun, keep=0.5, scale_factor=1)[source]#
Prunes minibatches as a subroutine of
SelectiveBackprop
. Computes the loss function on the provided training examples and runs minibatches according to the difficulty. The fraction of the minibatch that is kept for gradient computation is specified by the argument0 <= keep <= 1
.To speed up SBโs selection forward pass, the argument
scale_factor
can be used to spatially downsample input tensors. The full-sized inputs will still be used for the weight gradient computation.- Parameters
input (Tensor) โ Input tensor to prune.
target (Tensor) โ Output tensor to prune.
model (Callable) โ Model with which to predict outputs.
loss_fun (Callable) โ Loss function of the form
loss(outputs, targets, reduction='none')
. The function must take the keyword argumentreduction='none'
to ensure that per-sample losses are returned.keep (float, optional) โ Fraction of examples in the batch to keep. Default:
0.5
.scale_factor (float, optional) โ Multiplier between 0 and 1 for spatial size. Downsampling requires the input tensor to be at least 3D. Default:
1
.
- Returns
(torch.Tensor, torch.Tensor) โ The pruned batch of inputs and targets
- Raises
ValueError โ If
scale_factor > 1
.TypeError โ If
loss_fun > 1
has the wrong signature or is not callable.
Note
This function runs an extra forward pass through the model on the batch of data. If you are using a non-default precision, ensure that this forward pass runs in your desired precision. For example:
import torch from composer.algorithms.selective_backprop import select_using_loss with torch.cuda.amp.autocast(True): X_new, y_new = select_using_loss( X_sb, y_sb, lin_model, loss_fun, keep=0.5, scale_factor=1 )
- composer.algorithms.selective_backprop.selective_backprop.should_selective_backprop(current_duration, batch_idx, start=0.5, end=0.9, interrupt=2)[source]#
Decides if selective backprop should be run based on time in training.
Returns true if the
current_duration
is betweenstart
andend
. It is recommended that SB be applied during the later stages of a training run, once the model has already โlearnedโ easy examples.To preserve convergence, SB can be interrupted with vanilla minibatch gradient steps every
interrupt
steps. Wheninterrupt=0
, SB will be used at every step during the SB interval. Wheninterrupt=2
, SB will alternate with vanilla minibatch steps.- Parameters
current_duration (float) โ The elapsed training duration. Must be within
[0.0, 1.0)
.batch_idx (int) โ The current batch within the epoch.
start (float, optional) โ The duration at which selective backprop should be enabled, as a percentage. Default:
0.5
.end (float, optional) โ The duration at which selective backprop should be disabled. Default:
0.9
.interrupt (int, optional) โ The number of batches between vanilla minibatch gradient updates. Default:
2
.
- Returns
bool โ If selective backprop should be performed on this batch.