Gradient checkpointing is a technique to reduce GPU memory cost.

There exists a PyTorch implementaion in the official repo. However, it is extremely slow with multiple GPUs.

This repo contains a PyTorch implemention that can work on multiple GPUs.

The main functionality is in checkpoint.py

import checkpoint
checkpoint.CheckpointFunction.apply(function, n, *args)

Parameters:

• function – describes what to run in the forward pass of the model or part of the model. It should also know how to handle the inputs passed as the tuple. For example, in LSTM, if user passes (activation, hidden), function should correctly use the first input as activation and the second input as hidden.
• n – number of inputs to the function
• args – tuple containing inputs to the function AND parameters to optimize in the function. Note that the first n elements in this tuple should be ordered inputs to the function. Other elements are considered as parameters.

Returns:

• Output of running function on inputs to the function

Note: We recommend using checkpointing with cp_BatchNorm2d instead of torch.nn.BatchNorm2d, to avoid accumulating the same batch norm statistics more than once.

We provide an example of applying our checkpointing on memory efficient densenet. It only involves changing a few lines in the original implementation. (The original implementation uses PyTorch official checkpointing.)

# bn_function is a function containing conv1, norm1, relu1.
# naive no checkpointing: bottleneck_output = bn_function(*prev_features)
# official implementation: bottleneck_output = cp.checkpoint(bn_function, *prev_features)
args = prev_features + tuple(self.norm1.parameters()) + tuple(self.conv1.parameters())
# The parameters to optimize in the bn_function are tuple(self.norm1.parameters()) + tuple(self.conv1.parameters())
bottleneck_output = cp.CheckpointFunction.apply(bn_function, len(prev_features), *args)

python-fire is not required for checkpointing, but is required for the efficient densenet demo.

pip install fire

• our checkpointing demo:

CUDA_VISIBLE_DEVICES=0,1 python cp_demo.py --efficient True --data cifar --save model --batch_size 256

• the official implementation demo:

CUDA_VISIBLE_DEVICES=0,1 python original_demo.py --efficient True --data cifar --save model --batch_size 256

This code is tested with PyTorch 1.0.0.dev20181102

Speed tested on TITAN X (Pascal)

Part of our code in checkpoint.py and cp_BatchNorm2d.py is from https://github.com/pytorch/pytorch

The efficient densenet demo is taken from https://github.com/gpleiss/efficient_densenet_pytorch