Resuming Training with New Dataset Fails

[schopra8]

🐛 Bug

If you train a model with a particular dataset for N epochs and then want to continue training with a new dataset, LitData throws an exception.

To Reproduce

Steps to reproduce the behavior:

1. Train a model with dataset-1
2. Cancel training after the first checkpoint is aved
3. Resume training with trainer.fit(model, datamodule=datamodule, ckpt_path=ckpt_path) where the datamodule now points to dataset-2.
4. Capture the following error

[rank7]: Original Traceback (most recent call last):
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 252, in _worker_loop
[rank7]:     fetcher = _DatasetKind.create_fetcher(dataset_kind, dataset, auto_collation, collate_fn, drop_last)
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 79, in create_fetcher
[rank7]:     return _utils.fetch._IterableDatasetFetcher(dataset, auto_collation, collate_fn, drop_last)
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 21, in __init__
[rank7]:     self.dataset_iter = iter(dataset)
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/litdata/streaming/combined.py", line 155, in __iter__
[rank7]:     self._iterator = _CombinedDatasetIterator(
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/litdata/streaming/combined.py", line 203, in __init__
[rank7]:     self._dataset_iters = [iter(dataset) for dataset in datasets]
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/litdata/streaming/combined.py", line 203, in <listcomp>
[rank7]:     self._dataset_iters = [iter(dataset) for dataset in datasets]
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/litdata/streaming/dataset.py", line 219, in __iter__
[rank7]:     self._validate_state_dict()
[rank7]:   File "/home/sahil/.cache/pypoetry/virtualenvs/auw7Hy33-py3.10/lib/python3.10/site-packages/litdata/streaming/dataset.py", line 447, in _validate_state_dict
[rank7]:     raise ValueError(
[rank7]: ValueError: The provided input_dir URL state doesn't match the current one. Found s3://dataset-2 instead of s3://dataset-1.

Code sample

Expected behavior

Training to start with the optimizer states, model weights, etc. but with a net new dataset.

Environment

• PyTorch Version (e.g., 1.0): 2.3.1
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source):
• Python version: 3.10
• CUDA/cuDNN version: 12.1
• GPU models and configuration: 2x8H100
• Any other relevant information:

Additional context

[tchaton]

Hey @schopra8,

Did you consume exactly N epochs of the first dataset ?

As a temporary hack, did you try dropping the dataloader state from the checkpoint before reloading it. This might unblock you.

We should enable dropping the state entirely when the epoch is terminated to enable reloading with different parameters. Maybe we can do it there: https://github.com/Lightning-AI/litdata/blob/main/src/litdata/streaming/dataloader.py#L468.

Would you be interested in trying to contribute a fix ?

[schopra8]

In reality, I'd use this after exactly N epochs for the first dataset was consumed. In this dummy example, I manually killed training on the first dataset after K steps (less than 1 epoch) and then tried changing the datasets. Would that require a different change to the codebase?

I'll try the hack today -- but definitely down to contribute a fix.

[tchaton]

Yes, we need to hack around and find the right fix. Feel free to make a draft PR and we can help you land a reliable fix.

[tchaton]

Hey @schopra8. Any updates ?

[schopra8]

@tchaton No updates on my end yet -- got busy with another modeling task. I'll be taking a crack it in the next few days

[bhimrazy]

Hi @schopra8,

We’ve released an update with bug fixes, including the one related to this issue. Currently, it only supports StreamingDataset. We will be adding fixes for CombinedStreamingDataset soon as well.

Please feel free to try it out and let us know how it goes. Thanks! 😊

[pratyushmaini]

Hi @bhimrazy @schopra8 is there an update on support for CombinedStreamingDataset. This would be really helpful!

[schopra8]

@pratyushmaini No updates on my end

[bhimrazy]

Hi @bhimrazy @schopra8 is there an update on support for CombinedStreamingDataset. This would be really helpful!

Hi @pratyushmaini, I have an open PR addressing the issues with CombinedStreamingDataset. Feel free to take a look when you get a chance and share any feedback.
I’ll be resuming work on it soon—there’s one case I haven't been able to resolve yet.

Also, if you could elaborate on the issues you're facing, that would be really helpful.