pytorch
/
opacus
/
14142042872
/
2
80%
main: 80%
main: 80%
|
Ran
|
Files
120
|
Run time
2s
|
Badge
Embed ▾
README BADGES
|
coverage: 85.496% (-0.2%) from 85.71%
push
github
facebook-github-botHow to initialize dataloader with DPDDP when poisson sampling is set to False (#745) Summary: Pull Request resolved: https://github.com/pytorch/opacus/pull/745 https://github.com/pytorch/opacus/issues/444 When poisson sampling is enabled, it acts as a distributed sampler by default i.e., splits the dataset across the devices equally and then each device selects each sample from its subset independently with probability sample_rate. However, when poisson sampling is disabled, the user needs to explicitly include the distributed sampler in the dataloader and adjust the batch size based on the world size as shown below. ``` dataloader = DataLoader(dataset, batch_size=batch_size//world_size, sampler=DistributedSampler(dataset)) ``` This is expected behavior and therefore requires no fix. Updating the document to include this information. Reviewed By: HuanyuZhang Differential Revision: D71654351 fbshipit-source-id: d7ce3aeb7
5211 of 6095 relevant lines covered (85.5%)
0.85 hits per line
