distributed_train on 0.8.x

[drexalt]

I am hoping to follow the steps laid out here to perform distributed training on maxxvitv2, but when reproducing the command I get unrecognized argument errors with an ImageFolder.

My data folder containing train and validation is 'data' and is in the same directory as train.py and distributed_train.sh

The relevant portion of my config looks like this

data_dir: 'data'
dataset: ''
dataset_download: false

Running the following command gives The following arguments are required: --dataset/-d
./distributed_train.sh 4 data --config maxxvit.yaml

If I try the following script and add -d, I get unrecognized arguments: --local_rank=1 --config maxxvit.yaml
./distributed_train.sh 4 -d data --config maxxvit.yaml

Could this be related to the positional data argument getting deprecated?

Thanks!!
Jonah

[rwightman]

@jturner116 yeah, probably the deprecation, if you load a config file that has it set it will use the value in the config (since it is prioritized over the positional) ..

by -d did you mean --data-dir? don't think there is a -d , specifying --data-dir data should override the config value

[rwightman]

I'll look at having the positional override the config... guess it's a bit non-obvious :)

[drexalt]

Thanks for the reply, you are quite right! Working great now :D
Thank you sir!
Jonah