[MLH Phase 2] Refactor Logging System #284
Conversation
facebook-github-bot
added
the
CLA Signed
grace-omotoso
force-pushed
the
event-writers-master
branch
from
979419f to
bdf5fd1
Compare
akainth015
force-pushed
the
event-writers-master
branch
from
bdf5fd1 to
e694347
Compare
akainth015
force-pushed
the
event-writers-master
branch
from
e694347 to
fef448c
Compare
The TensorboardWriter can also log rich images and histograms
Co-authored-by: Grace Omotoso <[email protected]>
Implement TensorboardWriter
| if key == "ep" | ||
| else f"{stdout_data}{key}: {value}; " | ||
| evt_stg.put_scalars( | ||
| rolling_btime=rolling_btime, rolling_eta=rolling_eta_secs |
There was a problem hiding this comment.
nit: I prefer keeping all times in human readable string. We may want to add a VisslEventStorage#put_string or change #put_scalars to #put_values and make it type agnostic.
| # ----------------------------------------------------------------------------------- # | ||
| TENSORBOARD_SETUP: | ||
| # whether to use tensorboard for the visualization | ||
| VISUALIZATION_SAMPLE_PERIOD: -1 |
There was a problem hiding this comment.
nit: comment what this is.
nit: move comment below USE_TENSORBOARD
|
|
||
|
|
||
| def default_hook_generator(cfg: AttrDict) -> List[ClassyHook]: | ||
| def default_hook_generator(cfg: AttrDict, event_storage) -> List[ClassyHook]: |
There was a problem hiding this comment.
nit: Type hint for event_storage.
| sample = next(task.data_iterator) | ||
|
|
||
| sample = construct_sample_for_model(sample, task) | ||
| vis_period = task.config.HOOKS.TENSORBOARD_SETUP.VISUALIZATION_SAMPLE_PERIOD |
There was a problem hiding this comment.
nit: Can we comment this code block?
| storage = task.event_storage | ||
| storage.put_images() | ||
| name = f"Model input sample: iteration: {task.iteration_num}" | ||
| for idx, vis_img in enumerate(sample["input"]): |
There was a problem hiding this comment.
thought: If batch size is large, this will upload a lot of images -- do we want to offer the option to only sample a small portion of the inputs?
| several backends (json, tensorboard, etc) | ||
| """ | ||
|
|
||
| def __init__(self, start_iter=0): |
There was a problem hiding this comment.
I think we need to add logic that ensures that we set the start_iter correctly upon loading a checkpoint, I don't see this logic anywherere?
| "tensorboard not installed, cannot use SSLTensorboardHook" | ||
| ) | ||
| logging.info("Setting up SSL Tensorboard Hook...") | ||
| self.tb_writer = tb_writer |
There was a problem hiding this comment.
Do we need tb_writer anymore in this class? Can we refactor ActivationStatisticsMonitor to use event_storage?
There was a problem hiding this comment.
confirming: none of these methods write to tensorboard anymore and that it's only handled in TensorBoardWriter#Write.
| evt_stg.put_scalars( | ||
| rolling_btime=rolling_btime, rolling_eta=rolling_eta_secs | ||
| ) | ||
| logging.info(stdout_data.strip()) |
There was a problem hiding this comment.
So I believe we lose the logging to stdout here, which imo we should definitely keep -- super helpful for debugging.
We could create another EventStorageWriter called StdOutWriter and add to event_storage_writers.
| ] | ||
|
|
||
| @property | ||
| def event_storage(self): |
There was a problem hiding this comment.
I think we can just use self.event_storage internally and externally.
| Add a scalar `value` to the `HistoryBuffer` associated with `name`. | ||
| """ | ||
| history = self._history[name] | ||
| value = float(value) |
There was a problem hiding this comment.
question: Do we need this?
|
@iseessel I see some of the comments are unaddressed, do we need to work on these? Lmk if there is anything I could do! |
Summary: Pull Request resolved: fairinternal/ssl_scaling#284 Reviewed By: odelalleau Differential Revision: D42220017 Pulled By: QuentinDuval fbshipit-source-id: 742419aa859fdbe4bc80f1f9e9f4771fee0f41a2
6 participants
The new logging system will be based around an event storage for each
SelfSupervisionTaskand events will be logged to their corresponding output device byVisslEventWriters like theJsonWriterandTensorboardWriter