Documentation

[gorold]

This branch/PR compiles the efforts to enhance the documentation of the uni2ts library.

[liu-jc]

Are we good to merge this? We can create a new one when we have more documentation?

[HALF111]

Hello, I'm sorry to bother you. When reading the code, I think the following parts may be relatively difficult to understand, and it is possible to add some comments in the code or additional documentation to further enhance the understanding of the project.

1. Regarding Data Formats:

   The definitions for wide, long, and long-wide data formats appear unclear, leaving uncertainty about their distinct characteristics in https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/data/builder/simple.py .

2. 'offset' and 'date_offset' Clarification:

   It would be helpful to have a brief explanation for the terms 'offset' and 'date_offset' within the function definitions in the same file.

3. Understanding Column Types and getitem Methods:

   It would be helpful to state the distinction between 'seqs' and 'non-seqs' columns, along with a detailed explanation of the methods '_getitem_int', '_getitem_iterable', and '_getitem_slice' in https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/data/indexer/hf_dataset_indexer.py .

4. MultiSampleTimeSeriesDataset Functionality:

   Clarification on the purpose of 'MultiSampleTimeSeriesDataset' and how it differs from 'TimeSeriesDataset' in https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/data/dataset.py would enhance the understanding of dataset handling in the project.

5. Sequence Packing Explanation:

   It would be better if documentation for the 'PackCollate' function as well as how Sequence Packing works in https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/data/loader.py can be provided, particularly how it uses queues to pack data and the process of forming batches, seems not that clear.

6. DistrParamProj and DistributionOutput Classes:

   The definition and operationals of the DistrParamProj class, along with the DistributionOutput class, which involve nested functions and lambdas, are not clearly outlined. An explanation of their roles within the model definition context would be better. (https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/distribution/_base.py and https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/model/moirai/module.py)

7. Explanation of PackedLoss Functions, etc:

   It would be helpful to provide a comprehensive explanation of the computation and application of 'PackedLoss', 'PackedPointLoss', and 'PackedDistributionLoss' defined in https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/loss/packed/_base.py .

8. Explanation of train_transform_map in Model Definition:

   A brief explanation of specific transformations applied within 'train_transform_map' in https://github.com/SalesforceAIResearch/uni2ts/tree/main/src/uni2ts/model/moirai/pretrain.py would be invaluable, which can help understand how each transform contributes to the processing of training data in the 'MoiraiPretrain` model.

Thank you again for your great work, and hope that you can find these suggestions/information useful!

[chenghaoliu89]

@gorold could you add clarification for load_dataset and _get_transform from lotsa_v1/_base.py?

[chenghaoliu89]

better to add code comment about batch, packed_batch, merged_batch in line 109-116 from loader.py, and show their format, shape, the label of padding used for bin_package etc.

[gorold]

better to add code comment about batch, packed_batch, merged_batch in line 109-116 from loader.py, and show their format, shape, the label of padding used for bin_package etc.

You can see the typehints for format and shape of arrays and tensors, but yeah, I'll add the docstrings for loader.py

[chenghaoliu89]

This is very helpful, could you also explain the what it means when the variate_id, time_id value is 1 or 0 in transform?