Param dlrm

.gitignore

@@ -0,0 +1,138 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

Dockerfile

@@ -9,5 +9,7 @@ FROM ${FROM_IMAGE_NAME}
 ADD requirements.txt .
 RUN pip install -r requirements.txt
 
+RUN pip install torch==1.3.1
+
 WORKDIR /code
 ADD . .

README.md

@@ -311,23 +311,24 @@ Benchmarking
       ./bench/dlrm_s_criteo_terabyte.sh ["--test-freq=10240 --memory-map --data-sub-sample-rate=0.875"]
     ```
     - Corresponding pre-trained model is available under [CC-BY-NC license](https://creativecommons.org/licenses/by-nc/2.0/) and can be downloaded here
-    [dlrm_emb64_subsample0.875_maxindrange10M_pretrained.pt](https://dlrm.s3-us-west-1.amazonaws.com/models/tb0875_10M.pt)   
+
+    [dlrm_emb64_subsample0.875_maxindrange10M_pretrained.pt](https://dlrm.s3-us-west-1.amazonaws.com/models/tb0875_10M.pt)
 
 <img src="./terabyte_0875_loss_accuracy_plots.png" width="900" height="320">
 
 *NOTE: Benchmarking scripts accept extra arguments which will be passed along to the model, such as --num-batches=100 to limit the number of data samples*
 
-4) The code supports interface with [MLPerf benchmark](https://mlperf.org). 
+4) The code supports interface with [MLPerf benchmark](https://mlperf.org).
    - Please refer to the following training parameters
    ```
      --mlperf-logging that keeps track of multiple metrics, including area under the curve (AUC)
-   
+
      --mlperf-acc-threshold that allows early stopping based on accuracy metric
-   
+
      --mlperf-auc-threshold that allows early stopping based on AUC metric
-   
+
      --mlperf-bin-loader that enables preprocessing of data into a single binary file
-   
+
      --mlperf-bin-shuffle that controls whether a random shuffle of mini-batches is performed
    ```
    - The MLPerf training model is completely specified and can be trained using the following script
@@ -367,6 +368,8 @@ pydot (*optional*)
 
 torchviz (*optional*)
 
+tqdm
+
 
 License
 -------

README.params.md

@@ -0,0 +1,51 @@
+
+# DLRM Distributed Branch
+
+Extend the PyTorch implementation to run DLRM on multi nodes on distributed platforms.
+The distributed version will be needed when data model becomes large.
+
+It inherents all the parameters from master DLRM implementation. 
+The distributed version add one more parameter:
+
+**--dist-backend**:
+   The backend support for the distributed version. As in torch.distributed package,
+   it can be "nccl", "mpi", and "gloo".
+
+In addition, it introduces the following new parameter::
+**--arch-project-size** : 
+   Reducing the number of interaction features for the dot operation. 
+   A project operation is applied to the dotted features to reduce its dimension size.
+   This is mainly due to the memory concern. It reduces the memory size needed for top MLP. 
+   A side effect is that it may also imrpove the model accuracy.
+
+## Usage
+
+Currently, it is launched with mpirun on multi-nodes. The hostfile need to be created or 
+a host list should be given. The DLRM parameters should be given in the same way as single
+node master branch.
+```bash
+mpirun -np 128 -hostfile hostfile python dlrm_s_pytorch.py ...
+```
+
+## Example
+```bash
+python dlrm_s_pytorch.py 
+   --arch-sparse-feature-size=128 
+   --arch-mlp-bot="2000-1024-1024-128" 
+   --arch-mlp-top="4096-4096-4096-1" 
+   --arch-embedding-size=$large_arch_emb 
+   --data-generation=random 
+   --loss-function=bce 
+   --round-targets=True 
+   --learning-rate=0.1 
+   --mini-batch-size=2048 
+   --print-freq=10240 
+   --print-time 
+   --test-mini-batch-size=16384 
+   --test-num-workers=16
+   --num-indices-per-lookup-fixed=1 
+   --num-indices-per-lookup=100
+   --arch-projection-size 30
+   --use-gpu
+```
+

data_loader_terabyte.py

@@ -325,7 +325,7 @@ def _test_bin():
 
     original_dataset = CriteoDataset(
         dataset='terabyte',
-        max_ind_range=-1,
+        max_ind_range=10 * 1000 * 1000,
         sub_sample_rate=1,
         randomize=True,
         split=args.split,