Fixed issue #135; rename total_batch_size to train_batch_size
#137
Conversation
…ize` Implemented issue mosaicml#135. Also renamed `total_batch_size` to `train_batch_size`. Updated hparams.
|
@anisehsani -- it would be awesome if each evaluator could support subset num batches |
composer/core/state.py
Outdated
| # but the getter will always return a Precision enum | ||
| precision: Union[str, types.Precision] # type: ignore | ||
| _precision: types.Precision = field(init=False) # but store an enum internally | ||
| steps_per_epoch: int = -1 # type: ignore |
There was a problem hiding this comment.
If steps_per_epoch is a property then it shouldn't need to be defined here too right?
There was a problem hiding this comment.
Yea...it's nonstandard.
Here, I defined it as a field so you can set steps_per_epoch as part of the __init__. But it's optional since it has a default value.
__init_ implicitly calls the property setter, which then sets the private variable _steps_per_epoch. The public getter steps_per_epoch always returns an int -- the actual length of the dataloader (if _steps_per_epoch is None), or the artificially reduced length (if the _steps_per_epoch is not None)
There was a problem hiding this comment.
It seems like so far there isn't a reason to set steps_per_epoch to the __init__ when constructing State though so what about removing it for now and then if a need arises it can be added back? Just b/c it's confusing as is
composer/core/state.py
Outdated
| # but the getter will always return a Precision enum | ||
| precision: Union[str, types.Precision] # type: ignore | ||
| _precision: types.Precision = field(init=False) # but store an enum internally | ||
| steps_per_epoch: int = -1 # type: ignore |
There was a problem hiding this comment.
It seems like so far there isn't a reason to set steps_per_epoch to the __init__ when constructing State though so what about removing it for now and then if a need arises it can be added back? Just b/c it's confusing as is
…ize` (mosaicml#137) 1. Remove the `subset_num_batches` from the dataset hparams. Synthetic datasets should instead use the length of the real dataset as the size, or have a configurable size 2. Add `train_subset_num_batches` and `eval_subset_num_batches` to the trainer hparams 3. Add a check in the trainer that ensures that, if this field is set, then `DatasetHparams.shuffle is False`, or otherwise emit a warning that every epoch may be using a different subset of samples 4. Renamed `total_batch_size` to `train_batch_size`. Updated hparams.
Implemented issue #135.
Also renamed
total_batch_sizetotrain_batch_size. Updated hparams.