How can I use multi-GPU to train UNMT about xlm HOT 6 CLOSED

What do you mean by the training time is the same? Is the perplexity the same at the end of a few epochs? Or do you look at the number of words per second? The number of words per second in the log is given per GPU, so this will be the same. But the loss / perplexity should decrease much faster.

yes, i made a mistake. you are right, multi-gpu training get better valid ppl and acc.
pretraining on 1 GPU:

on 4 GPU

from xlm.

another question, in the UNMT model, only one encoder and one decoder? Thanks.

from xlm.

You should not handle the --local_rank yourself. You can use the following command to train with multi-GPU: https://github.com/facebookresearch/XLM#how-can-i-run-experiments-on-multiple-gpus

export NGPU=8; python -m torch.distributed.launch --nproc_per_node=$NGPU train.py ARGUMENTS

And no, there are 2 separate models for UNMT, one encoder and one decoder, but they are initialized with the same weights (apart from the parameters of the source attention in the decoder that remain randomly initialized).

from xlm.

You should not handle the --local_rank yourself. You can use the following command to train with multi-GPU: https://github.com/facebookresearch/XLM#how-can-i-run-experiments-on-multiple-gpus

export NGPU=8; python -m torch.distributed.launch --nproc_per_node=$NGPU train.py ARGUMENTS

And no, there are 2 separate models for UNMT, one encoder and one decoder, but they are initialized with the same weights (apart from the parameters of the source attention in the decoder that remain randomly initialized).

I using multi-GPU to pre-training the model like export NGPU=8; python -m torch.distributed.launch --nproc_per_node=$NGPU train.py ARGUMENTS, it just run the same job on 8 GPUs, the training time is the same as training on 1 GPU, it doesn't fast the pre-training process.
How to set the params and I can fast the training process on multi-GPU?

from xlm.

What do you mean by the training time is the same? Is the perplexity the same at the end of a few epochs? Or do you look at the number of words per second? The number of words per second in the log is given per GPU, so this will be the same. But the loss / perplexity should decrease much faster.

from xlm.

Looks good :)

from xlm.