Need help with pytorch-checkpoint?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

144 Stars 13 Forks Other 14 Commits 1 Opened issues

Services available


Need anything else?

Contributors list

# 281,414
13 commits


Gradient checkpointing is a technique to reduce GPU memory cost.

Official implementation

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

This implementation

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

Main results

| Method | # GPU | Batch | Memory | Time | |--------|:----:|:-----:|:------:|:-----:| |Naive|2|256| 5.25G | 0.27s | |Official|2|256|2.98G|1.41s| |This repo|2|256|2.97G|0.31s|


The main functionality is in

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


  • 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.

* 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.

DenseNet example

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.) ```python

bn_function is a function containing conv1, norm1, relu1.

naive no checkpointing: bottleneckoutput = bnfunction(*prev_features)

official implementation: bottleneckoutput = cp.checkpoint(bnfunction, *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())

bottleneckoutput = cp.CheckpointFunction.apply(bnfunction, 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 --efficient True --data cifar --save model --batch_size 256
* the official implementation demo:
CUDA_VISIBLE_DEVICES=0,1 python --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)

Full results

| Method | # GPU | Batch | Memory | Time | |--------|:----:|:-----:|:------:|:-----:| |Naive|1|256| 9.93G | 0.42s | |Naive|2|4| 0.65G | 0.10s | |Naive|2|256| 5.25G | 0.27s | |Naive|2|512| 9.93G | 0.50s | |Official|1|256|5.38G|0.52s| |Official|1|512|10.1G|1.00s| |Official|2|4|0.62G|1.40s| |Official|2|256|2.98G|1.41s| |Official|2|512|5.39G|1.53s| |This repo|1|256|5.37G|0.50s| |This repo|1|512|10.1G|0.97s| |This repo|2|4|0.62G|0.13s| |This repo|2|256|2.97G|0.31s| |This repo|2|512|5.37G|0.58s|


Part of our code in and is from

The efficient densenet demo is taken from

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.