NVidia's pytorch-based object_detection implementation (#240)

* Delete Caffe2 object_detection

* Added new pytorch-based object_detection

* object_detection: removed unused configs; deleted misleading code

* object_detection Dockerfile now based on public image and specifies exact library versions

.gitmodules

@@ -1,9 +1,3 @@
-[submodule "object_detection/caffe2"]
-	path = object_detection/caffe2/caffe2
-	url = https://github.com/pytorch/pytorch.git
-[submodule "object_detection/detectron"]
-	path = object_detection/caffe2/detectron
-	url = https://github.com/ddkang/Detectron.git
 [submodule "community"]
 	path = community
 	url = https://github.com/mlperf/community.git

object_detection/Dockerfile

@@ -0,0 +1,72 @@
+FROM pytorch/pytorch:1.0.1-cuda10.0-cudnn7-devel
+
+RUN echo 'debconf debconf/frontend select Noninteractive' | debconf-set-selections
+
+# install basics
+RUN apt-get update -y \
+ && apt-get install -y apt-utils=1.2.29ubuntu0.1 \
+                       libglib2.0-0=2.48.2-0ubuntu4.1 \
+                       libsm6=2:1.2.2-1 \
+                       libxext6=2:1.3.3-1 \
+                       libxrender-dev=1:0.9.9-0ubuntu1
+
+RUN pip install ninja==1.8.2.post2 \
+                yacs==0.1.5 \
+                cython==0.29.5 \
+                matplotlib==3.0.2 \
+                opencv-python==4.0.0.21 \
+                mlperf_compliance==0.0.10 \
+                torchvision==0.2.2
+
+# install pycocotools
+RUN git clone https://github.com/cocodataset/cocoapi.git \
+ && cd cocoapi/PythonAPI \
+ && git reset --hard ed842bffd41f6ff38707c4f0968d2cfd91088688 \
+ && python setup.py build_ext install
+
+# For information purposes only, these are the versions of the packages which we've successfully used:
+# $ pip list
+# Package              Version           Location
+# -------------------- ----------------- -------------------------------------------------
+# backcall             0.1.0
+# certifi              2018.11.29
+# cffi                 1.11.5
+# cycler               0.10.0
+# Cython               0.29.5
+# decorator            4.3.2
+# fairseq              0.6.0             /scratch/fairseq
+# ipython              7.2.0
+# ipython-genutils     0.2.0
+# jedi                 0.13.2
+# kiwisolver           1.0.1
+# maskrcnn-benchmark   0.1               /scratch/mlperf/training/object_detection/pytorch
+# matplotlib           3.0.2
+# mkl-fft              1.0.10
+# mkl-random           1.0.2
+# mlperf-compliance    0.0.10
+# ninja                1.8.2.post2
+# numpy                1.16.1
+# opencv-python        4.0.0.21
+# parso                0.3.2
+# pexpect              4.6.0
+# pickleshare          0.7.5
+# Pillow               5.4.1
+# pip                  19.0.1
+# prompt-toolkit       2.0.8
+# ptyprocess           0.6.0
+# pycocotools          2.0
+# pycparser            2.19
+# Pygments             2.3.1
+# pyparsing            2.3.1
+# python-dateutil      2.8.0
+# pytorch-quantization 0.2.1
+# PyYAML               3.13
+# setuptools           40.8.0
+# six                  1.12.0
+# torch                1.0.0.dev20190225
+# torchvision          0.2.1
+# tqdm                 4.31.1
+# traitlets            4.3.2
+# wcwidth              0.1.7
+# wheel                0.32.3
+# yacs                 0.1.5

object_detection/caffe2/README.md → object_detection/README.md

@@ -1,90 +1,95 @@
-# 1. Problem
-Object detection and segmentation. Metrics are mask and box mAP.
-
-# 2. Directions
-### Steps to configure machine
-Standard script.
-
-### Steps to download and verify data
-init and update the submodules in this directory.
-
-Run the provided shell scripts *in this directory*.
-
-### Steps to run and time
-Build the docker container.
-
-```
-sudo docker build -t detectron .
-```
-
-Run the docker container and mount the data appropriately
-
-```
-sudo nvidia-docker run
--v /mnt/disks/data/coco/:/packages/detectron/lib/datasets/data/coco
--it detectron /bin/bash
-```
-
-(replace /mnt/disks/data/coco/ with the data directory)
-
-Run the command:
-```
-time stdbuf -o 0 \
-  python tools/train_net.py --cfg configs/12_2017_baselines/e2e_mask_rcnn_R-50-FPN_1x.yaml \
-    --box_min_ap 0.377 --mask_min_ap 0.339 \
-    --seed 3 | tee run.log
- ```
-
-# 3. Dataset/Environment
-### Publication/Attribution
-Microsoft COCO: Common Objects in Context
-
-### Data preprocessing
-Only horizontal flips are allowed.
-
-### Training and test data separation
-As provided by MS-COCO (2017 version).
-
-### Training data order
-Randomly.
-
-### Test data order
-Any order.
-
-# 4. Model
-### Publication/Attribution
-He, Kaiming, et al. "Mask r-cnn." Computer Vision (ICCV), 2017 IEEE International Conference on.
-IEEE, 2017.
-
-We use a version of Mask R-CNN with a ResNet50 backbone.
-
-### List of layers 
-Running the timing script will display a list of layers.
-
-### Weight and bias initialization
-The ResNet50 base must be loaded from the provided weights. They may be quantized.
-
-### Loss function
-Multi-task loss (classification, box, mask). Described in the Mask R-CNN paper.
-
-Classification: Smooth L1 loss
-
-Box: Log loss for true class.
-
-Mask: per-pixel sigmoid, average binary cross-entropy loss.
-
-### Optimizer
-Momentum SGD. Weight decay of 0.0001, momentum of 0.9.
-
-# 5. Quality
-### Quality metric
-As Mask R-CNN can provide both boxes and masks, we evaluate on both box and mask mAP.
-
-### Quality target
-Box mAP of 0.377, mask mAP of 0.339
-
-### Evaluation frequency
-Once per epoch, 118k.
-
-### Evaluation thoroughness
-Evaluate over the entire validation set. Use the official COCO API to compute mAP.
+# 1. Problem
+Object detection and segmentation. Metrics are mask and box mAP.
+
+# 2. Directions
+
+### Steps to configure machine
+
+1. Checkout the MLPerf repository
+```
+mkdir -p mlperf
+cd mlperf
+git clone https://github.com/mlperf/training.git
+```
+2. Install CUDA and Docker
+```
+source training/install_cuda_docker.sh
+```
+3. Build the docker image for the object detection task
+```
+cd training/object_detection/
+nvidia-docker build . -t mlperf/object_detection
+```
+
+4. Run docker container and install code
+```
+nvidia-docker run -v .:/workspace -t -i --rm --ipc=host mlperf/object_detection \
+    "cd mlperf/training/object_detection && ./install.sh"
+```
+Now exit the docker container (Ctrl-D) to get back to your host.
+
+### Steps to download data
+```
+# From training/object_detection/
+source download_dataset.sh
+```
+
+### Steps to run benchmark.
+```
+nvidia-docker run -v .:/workspace -t -i --rm --ipc=host mlperf/object_detection \
+    "cd mlperf/training/object_detection && ./run_and_time.sh"
+```
+
+# 3. Dataset/Environment
+### Publication/Attribution
+Microsoft COCO: Common Objects in Context
+
+### Data preprocessing
+Only horizontal flips are allowed.
+
+### Training and test data separation
+As provided by MS-COCO (2017 version).
+
+### Training data order
+Randomly.
+
+### Test data order
+Any order.
+
+# 4. Model
+### Publication/Attribution
+He, Kaiming, et al. "Mask r-cnn." Computer Vision (ICCV), 2017 IEEE International Conference on.
+IEEE, 2017.
+
+We use a version of Mask R-CNN with a ResNet50 backbone.
+
+### List of layers
+Running the timing script will display a list of layers.
+
+### Weight and bias initialization
+The ResNet50 base must be loaded from the provided weights. They may be quantized.
+
+### Loss function
+Multi-task loss (classification, box, mask). Described in the Mask R-CNN paper.
+
+Classification: Smooth L1 loss
+
+Box: Log loss for true class.
+
+Mask: per-pixel sigmoid, average binary cross-entropy loss.
+
+### Optimizer
+Momentum SGD. Weight decay of 0.0001, momentum of 0.9.
+
+# 5. Quality
+### Quality metric
+As Mask R-CNN can provide both boxes and masks, we evaluate on both box and mask mAP.
+
+### Quality target
+Box mAP of 0.377, mask mAP of 0.339
+
+### Evaluation frequency
+Once per epoch, 118k.
+
+### Evaluation thoroughness
+Evaluate over the entire validation set. Use the official COCO API to compute mAP.

object_detection/download_dataset.sh

@@ -1,4 +1,26 @@
-wget https://s3-us-west-2.amazonaws.com/detectron/coco/coco_annotations_minival.tgz
-wget http://images.cocodataset.org/zips/train2014.zip
-wget http://images.cocodataset.org/zips/val2014.zip
-wget http://images.cocodataset.org/annotations/annotations_trainval2014.zip
+#!/bin/bash
+
+# Get COCO 2014 data sets
+mkdir -p pytorch/datasets/coco
+pushd pytorch/datasets/coco
+
+curl -O https://dl.fbaipublicfiles.com/detectron/coco/coco_annotations_minival.tgz
+tar xzf coco_annotations_minival.tgz
+
+curl -O http://images.cocodataset.org/zips/train2014.zip
+unzip train2014.zip
+
+curl -O http://images.cocodataset.org/zips/val2014.zip
+unzip val2014.zip
+
+curl -O http://images.cocodataset.org/annotations/annotations_trainval2014.zip
+unzip annotations_trainval2014.zip
+
+# TBD: MD5 verification
+# $md5sum *.zip *.tgz
+#f4bbac642086de4f52a3fdda2de5fa2c  annotations_trainval2017.zip
+#cced6f7f71b7629ddf16f17bbcfab6b2  train2017.zip
+#442b8da7639aecaf257c1dceb8ba8c80  val2017.zip
+#2d2b9d2283adb5e3b8d25eec88e65064  coco_annotations_minival.tgz
+
+popd