Add feature vector dataloader

[xianyuanliu]

Description

Update action_dann_lightn example with a dataloader of feature vector input. This dataloader is used in the EPIC challenge.
The trainer and models for feature vectors will be updated in the next PR after this one is merged. Coverage in video.py will be improved in the next PR because of most parts are used in training.
Changes are summarized below.

1. Add EPIC100DatasetAccess
2. Add new hyperparameters: CLASS_TYPE, INPUT_TYPE and NUM_SEGMENTS
3. Improve video-related trainers and ClassNetVideo for multi-class output.

Status

Ready

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• In-line docstrings updated.
• Source for documentation at docs manually updated for new API.

[codecov-commenter]

Codecov Report

Merging #291 (2773771) into main (3706d7c) will decrease coverage by 0.39%.
The diff coverage is 80.92%.

@@            Coverage Diff             @@
##             main     #291      +/-   ##
==========================================
- Coverage   92.96%   92.56%   -0.40%     
==========================================
  Files          45       45              
  Lines        4622     4724     +102     
==========================================
+ Hits         4297     4373      +76     
- Misses        325      351      +26     

Impacted Files	Coverage Δ
kale/loaddata/videos.py	73.18% <51.85%> (-13.48%)	⬇️
kale/loaddata/video_access.py	97.05% <94.91%> (-0.86%)	⬇️
kale/embed/video_feature_extractor.py	96.66% <100.00%> (+0.05%)	⬆️
kale/pipeline/video_domain_adapter.py	96.93% <100.00%> (ø)
kale/predict/class_domain_nets.py	91.89% <100.00%> (+1.89%)	⬆️
kale/prepdata/video_transform.py	88.00% <100.00%> (+1.04%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3706d7c...2773771. Read the comment docs.

[shuo-zhou]

Avoid using absolute path in examples

[xianyuanliu]

It may be necessary because the data is stored somewhere else (e.g. /Shared/data), not in pykale.

[shuo-zhou]

Is it possible to use public dataset for demonstration? Do you expect users to change the .yaml configurations and download the data themselves?

[xianyuanliu]

It uses public video datasets (e.g. EPIC KITCHENS), but each is large (>10G). Therefore, users need to download and prepare data following dataset instructions before running this example. Then, of course, they need to change configurations fit for data.

Reasons that this example does not contain video dataset downloading are as follows:

1. Some datasets provide tools or packages to boost downloading speed, but no need to add these into our pykale.
2. Pykale and data may not be in the same hardware due to the datasets' large-scale.

I may update the readme to describe the dataset format. We may add video dataset downloading in the future when we reduce dataset storage by using some dataset tools (e.g. Hub).

[shuo-zhou]

Examples are used for demonstration only. It should not be run on large, and full dataset. Is there any small and public dataset can be used?

[haipinglu]

I agree with Shuo. PyKale examples should be relatively easy to run. Large-scale data experiment can be created on separate repo under PyKale or your personal repo. We can discuss tomorrow.

[shuo-zhou]

What is the reason of renaming num_classes as dict_num_classes? It is not consistent with the variable names in other examples.

[xianyuanliu]

The reason is that num_classes is a dictionary with a multi-task label (verb and noun). Other examples do not consider this. But I think no need to rename num_classes now. I will change it back.

[github-actions]

This pull request has been automatically marked as stale due to lack of activity.