This is a re-implementation of Training Deep Nets with Sublinear Memory Cost. You may also want to have a look at the original mxnet implementation and OpenAI's tensorflow implementation.

Different from TensorFlow and mxnet where the computation graph is static and known before actual computing, pytorch's philosophy is define-by-run and the graph details are not known until forward is finished. This implemention only supports Sequential models. By replacing nn.Sequential with memonger.SublinearSequential, the memory required for backward is reduced from O(N) to O(sqrt(N)).

# previous, O(N) memory footprint
import torch.nn as nn
net1 = nn.Sequential(
    nn.Conv2d(3, 16, kernel=3, padding=1),
    nn.BatchNorm2d(16),
    nn.ReLU(),
    nn.Conv2d(16, 16, kernel=3, padding=1),
    nn.BatchNorm2d(16),
    nn.ReLU(),
    nn.Conv2d(16, 16, kernel=3, padding=1),
    nn.BatchNorm2d(16),
    nn.ReLU(),
    ...
)

# optimized, O(sqrt(N)) memory footprint
from memonger import SublinearSequential
net2 = SublinearSequential(
    *list(net1.children())  
)

Since sublinear memory optimization requires re-forwarding, if your model contains layer with non-derministic behavior (e.g, BatchNorm, Dropout), you need to be careful when using the module. I have supported BatchNorm by re-scaling momentum , dropout by memorizing the random number generator (RNG).