Remove old timing calls from selective backprop #221
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
For the timing abstraction (#146), the `DataloaderSpec` needed two addition functions -- `get_num_samples_in_batch` and `get_num_tokens_in_batch`. It was getting messy to pass around function pointers in a named tuple, so instead converted `DataloaderSpec` from a NamedTuple into a regular class called `DataSpec`. Custom datasets can inherit the base `DataSpec` class and override functionality as needed. Moved the `DataSpec` class to `composer.core`, as the `DataSpec` is now bound directly to the state. #120 will also need this change. Renamed `train_dataloader` and `eval_dataloader` in the trainer and state to `train_data` and `eval_data`, since it encompasses more than the dataloader. This PR implements part 3 and 4 of the timing abstraction (#146). The implementation differs from the GH issue by adding `num_tokens`, `num_samples`, `get_batch_size`, and `get_num_tokens` to the new `DataSpec` rather than the pytorch dataset class.
- Updated the `TrainerHparams`, `Trainer`, and `State`, to take max_duration instead of max_epochs. However, for the time being, it is required that `max_duration` be in epoch units. This will be required until `state.max_epochs` is no longer used by anything. - Updated the state to store an instance of `Timer`. Added helper getters to the state for backwords compatibility with the previous API. Added deprecation warnings - Updated the Timer to allow comparisions directly between the timer object and instances of Time. These statements are now equivalent `(timer <=other) == (timer.get(other.unit) <= other)` - Updated the Timer to track the number of batches, samples, and tokens in the current epoch. Tracking batches is required for `state.batch_idx`; and it'd be good to provide equivalent functions for samples and tokens - Updated the Time to allow for integer comparisions. The same units will be assumed. These statements are now equivalent `(time <= 10) == (time <= Time(10, time.unit)` - Updated the rest of the codebase -- mainly tests and hparam files -- to use these new APIs - Removed the old Time parsing logic from the composer.optim.scheduler
This was referenced Jan 13, 2022
This was referenced Jan 13, 2022
ravi-mosaicml
changed the title
Landanjs
approved these changes
Jan 14, 2022
coryMosaicML
pushed a commit
to coryMosaicML/composer
that referenced
this pull request
Feb 23, 2022
Updated the selective backprop API to use the timing abstraction, as detailed in mosaicml#146
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Updated the selective backprop API to use the timing abstraction, as detailed in #146