GitLab

The option is --nouse_estimator

--ml_perf now just changes the model to make it MLPerf compliant. --output_ml_perf_compliance_logging adds the MLPerf compliance logs.

* first pass at __getattr__ abuse logger

* first pass at adding tags to NCF

* minor formatting updates

* fix tag name

* convert metrics to python floats

* getting closer...

* direct mlperf logs to a file

* small tweaks and add stitching

* update tags

* fix tag and add a sudo call

* tweak format of run.sh

* delint

* use distribution strategies for evaluation

* address PR comments

* delint and fix test

* adjust flag validation for xla

* add prefix to distinguish log stitching

* fix index bug

* fix clear cache for root user

* dockerize cache drop

* TIL some regex magic

* intermediate commit

finish replacing spillover with resampled padding

intermediate commit

* resolve merge conflict

* intermediate commit

* further consolidate the data pipeline

* complete first pass at data pipeline refactor

* remove some leftover code

* fix test

* remove resampling, and move train padding logic into neumf.py

* small tweaks

* fix weight bug

* address PR comments

* fix dict zip. (Reed led me astray)

* delint

* make data test deterministic and delint

* Reed didn't lead me astray. I just can't read.

* more delinting

* even more delinting

* use resampling for last batch padding

* pad last batch with unique data

* Revert "pad last batch with unique data"

This reverts commit cbdf46efcd5c7907038a24105b88d38e7f1d6da2.

* move padded batch to the beginning

* delint

* fix step check for synthetic data

* move flagfile into the cache_dir

* remove duplicate code

* delint

* Use data_dir instead of flags.FLAGS.data_dir in data_preprocessing.py.

* Use data_dir instead of flags.FLAGS.data_dir in data_preprocessing.py.

* Replace multiprocess pool with popen_helper.get_pool() in data_preprocessing.

* Use data_dir instead of flags.FLAGS.data_dir in data_preprocessing.py.

* Use data_dir instead of flags.FLAGS.data_dir in data_preprocessing.py.

Add comments, exit async process after waiting for flagfile for too long and make directory for data_dir in case it does not exist.

* Add --use_synthetic_data option to NCF.

* Add comment to _SYNTHETIC_BATCHES_PER_EPOCH

* Fix test

* Hopefully fix lint issue

* move evaluation from numpy to tensorflow

fix syntax error

don't use sigmoid to convert logits. there is too much precision loss.

WIP: add logit metrics

continue refactor of NCF evaluation

fix syntax error

fix bugs in eval loss calculation

fix eval loss reweighting

remove numpy based metric calculations

fix logging hooks

fix sigmoid to softmax bug

fix comment

catch rare PIPE error and address some PR comments

* fix metric test and address PR comments

* delint and fix python2

* fix test and address PR comments

* extend eval to TPUs

* bug fixes and add seed

* more random corrections

* make cleanup more robust

* return cleanup fn

* delint and address PR comments.

* delint and fix tests

* delinting is never done

* add pipeline hashing

* delint

Sometimes it takes longer than 15 seconds, and even longer than 1 minute, to spawn and create the alive file.

* Fix spurious "did not start correctly" error.

The error "Generation subprocess did not start correctly" would occur if the async process started up after the main process checked for the subproc_alive file.

* Add error message

* Fix convergence issues for MLPerf.

Thank you to @robieta for helping me find these issues, and for providng an algorithm for the `get_hit_rate_and_ndcg_mlperf` function.

This change causes every forked process to set a new seed, so that forked processes do not generate the same set of random numbers. This improves evaluation hit rates.

Additionally, it adds a flag, --ml_perf, that makes further changes so that the evaluation hit rate can match the MLPerf reference implementation.

I ran 4 times with --ml_perf and 4 times without. Without --ml_perf, the highest hit rates achieved by each run were 0.6278, 0.6287, 0.6289, and 0.6241. With --ml_perf, the highest hit rates were 0.6353, 0.6356, 0.6367, and 0.6353.

* fix lint error

* Fix failing test

* Address @robieta's feedback

* Address more feedback

This is done by using a higher Pickle protocol version, which the Python docs describe as being "slightly more efficient". This reduces the file write time at the beginning from 2 1/2 minutes to 5 seconds.

The data_async_generation.py process would print to stderr, but the main process would redirect it's stderr to a pipe. The main process never read from the pipe, so when the pipe was full, data_async_generation.py would stall on a write to stderr. This change makes data_async_generation.py not write to stdout/stderr.

* add indirection file

* remove unused imports

* fix import

* intermediate commit

* ncf now working

* reorder pipeline

* allow batched decode for file backed dataset

* fix bug

* more tweaks

* parallize false negative generation

* shared pool hack

* workers ignore sigint

* intermediate commit

* simplify buffer backed dataset creation to fixed length record approach only. (more cleanup needed)

* more tweaks

* simplify pipeline

* fix misplaced cleanup() calls. (validation works\!)

* more tweaks

* sixify memoryview usage

* more sixification

* fix bug

* add future imports

* break up training input pipeline

* more pipeline tuning

* first pass at moving negative generation to async

* refactor async pipeline to use files instead of ipc

* refactor async pipeline

* move expansion and concatenation from reduce worker to generation workers

* abandon complete async due to interactions with the tensorflow threadpool

* cleanup

* remove performance_comparison.py

* experiment with rough generator + interleave pipeline

* yet more pipeline tuning

* update on-the-fly pipeline

* refactor preprocessing, and move train generation behind a GRPC server

* fix leftover call

* intermediate commit

* intermediate commit

* fix index error in data pipeline, and add logging to train data server

* make sharding more robust to imbalance

* correctly sample with replacement

* file buffers are no longer needed for this branch

* tweak sampling methods

* add README for data pipeline

* fix eval sampling, and vectorize eval metrics

* add spillover and static training batch sizes

* clean up cruft from earlier iterations

* rough delint

* delint 2 / n

* add type annotations

* update run script

* make run.sh a bit nicer

* change embedding initializer to match reference

* rough pass at pure estimator model_fn

* impose static shape hack (revisit later)

* refinements

* fix dir error in run.sh

* add documentation

* add more docs and fix an assert

* old data test is no longer valid. Keeping it around as reference for the new one

* rough draft of data pipeline validation script

* don't rely on shuffle default

* tweaks and documentation

* add separate eval batch size for performance

* initial commit

* terrible hacking

* mini hacks

* missed a bug

* messing about trying to get TPU running

* TFRecords based TPU attempt

* bug fixes

* don't log remotely

* more bug fixes

* TPU tweaks and bug fixes

* more tweaks

* more adjustments

* rework model definition

* tweak data pipeline

* refactor async TFRecords generation

* temp commit to run.sh

* update log behavior

* fix logging bug

* add check for subprocess start to avoid cryptic hangs

* unify deserialize and make it TPU compliant

* delint

* remove gRPC pipeline code

* fix logging bug

* delint and remove old test files

* add unit tests for NCF pipeline

* delint

* clean up run.sh, and add run_tpu.sh

* forgot the most important line

* fix run.sh bugs

* yet more bash debugging

* small tweak to add keras summaries to model_fn

* Clean up sixification issues

* address PR comments

* delinting is never over