[AutoMM] Refactor presets #2749
Conversation
|
Job PR-2749-fe3e29c is done. |
zhiqiangdon
requested review from
sxjscience,
bryanyzhu,
cheungdaven and
FANGAreNotGnu
| hyperparameters.update( | ||
| { | ||
| "model.hf_text.checkpoint_name": "google/electra-small-discriminator", | ||
| "model.timm_image.checkpoint_name": "swin_small_patch4_window7_224", |
There was a problem hiding this comment.
For the quick build setting of image classification, we recommend to use mobilenetv3_large_100 before, we might want to stay consistent.
There was a problem hiding this comment.
Changed to mobilenetv3_large_100.
|
|
||
|
|
||
| @automm_presets.register() | ||
| def zero_shot_image_classification(): | ||
| return { | ||
| def zero_shot_image_classification(presets: str = DEFAULT, hpo: bool = False): |
There was a problem hiding this comment.
Will this be too complicated? Can we just register presets a few strings?
The current design is that you need to provide a presets flag in automm_presets.register().
There was a problem hiding this comment.
Also, consider to add docstrings to explain the design. For example, I can understand the current presets are registered per problem_type.
For example, for image_classification, we can have its own presets, like DEFAULT, BEST_QUALITY, etc.
There was a problem hiding this comment.
The string design is not easy to maintain and scale to more problem types and HPO presets. There are much redundancy in the strings (one problem type name may repeat three times) and hyperparameters (different presets of the same problem type may share common ones).
There was a problem hiding this comment.
Docstrings added.
| from .registry import Registry | ||
|
|
||
| automm_presets = Registry("automm_presets") | ||
| matcher_presets = Registry("matcher_presets") | ||
|
|
||
|
|
||
| @automm_presets.register() | ||
| def high_quality_fast_inference(): | ||
| return { | ||
| def default(presets: str = DEFAULT, hpo: bool = False): |
There was a problem hiding this comment.
Why do we need hpo flag?
There was a problem hiding this comment.
This is to support the HPO presets, i.e., returning not only hyperparameters but also hyperparameter_tune_kwargs.
|
Job PR-2749-876c5c2 is done. |
|
Job PR-2749-c9cd115 is done. |
|
Job PR-2749-8940394 is done. |
Issue #, if available:
Description of changes:
best_quality,high_quality, andmedium_qualityas presets.By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.