How to run wikitext-103 dataset using pixelated butterfly?

Open hulihan-start opened this issue 1 year ago • 0 comments

This is an interesting work!

I would like to run the experiment on GPT-2 on WikiText-103 dataset. The dataset could not be accessed from the original url. I downloaded the dataset from a 3rd party but still cannot change the data_dir by using 'python run.py experiment=wt103/gpt2.yaml datamodule.data_dir=/home/dataset/wikitext-103/'.

I am sure the environment is the same as the docker file. Anyone can help me to solve this issue?

Also, when I used 'python run.py experiment=example_full.yaml' to test the example, the program raise an exception: ' Error executing job with overrides: ['experiment=example_full.yaml'] Error in call to target 'src.tasks.seq.SequenceModel': InstantiationException('Error in call to target 'src.models.mnist_model.MNISTLitModel':\nTypeError("init() got an unexpected keyword argument 'num_classes'")\nfull_key: cfg.model') full_key: task'

After I commented line 45-46 in src/tasks/seq.py, the code can train successfully, but the test phase is still abort because of 'pytorch_lightning.utilities.exceptions.MisconfigurationException: No test_dataloader() method defined to run Trainer.test'

Do you have any good suggestions for the users to setup a custom experiment? For example, if I want to train gpt2-fbbfly.yaml on wt103 dataset, what should I type in the command?

Thanks!!!

Oct 21 '24 03:10 hulihan-start