Adds the container package for SSD-ResNet34 inference and training (P…

…yTorch SPR) (#96)

* Add specs, docs, and dockerfiles for PyTorch SPR SSD-ResNet34 inference and training

* Update spec for train scripts

* Fix PRETRAINED_MODEL volume in run.sh

* Update training partial

* Copy in requirements

* remove numpy since we already have it, and add shm-size

* Addin missing new line

* Updates based on review feedback

dockerfiles/pytorch/pytorch-spr-ssd-resnet34-inference.Dockerfile

@@ -0,0 +1,79 @@
+# Copyright (c) 2020-2021 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+#
+# THIS IS A GENERATED DOCKERFILE.
+#
+# This file was assembled from multiple pieces, whose use is documented
+# throughout. Please refer to the TensorFlow dockerfiles documentation
+# for more information.
+
+ARG PYTORCH_IMAGE="model-zoo"
+ARG PYTORCH_TAG="pytorch-ipex-spr"
+
+FROM ${PYTORCH_IMAGE}:${PYTORCH_TAG} AS intel-optimized-pytorch
+
+# Build Torch Vision
+ARG TORCHVISION_VERSION=v0.8.0
+
+RUN source ~/anaconda3/bin/activate pytorch && \
+    git clone https://github.com/pytorch/vision && \
+    cd vision && \
+    git checkout ${TORCHVISION_VERSION} && \
+    python setup.py install
+
+RUN source ~/anaconda3/bin/activate pytorch && \
+    pip install matplotlib Pillow pycocotools && \
+    pip install yacs opencv-python cityscapesscripts transformers && \
+    conda install -y libopenblas psutil && \
+    cd /workspace/installs && \
+    wget https://github.com/gperftools/gperftools/releases/download/gperftools-2.7.90/gperftools-2.7.90.tar.gz && \
+    tar -xzf gperftools-2.7.90.tar.gz && \
+    cd gperftools-2.7.90 && \
+    ./configure --prefix=$HOME/.local && \
+    make && \
+    make install && \
+    rm -rf /workspace/installs/
+
+ARG PACKAGE_DIR=model_packages
+
+ARG PACKAGE_NAME="pytorch-spr-ssd-resnet34-inference"
+
+ARG MODEL_WORKSPACE
+
+# ${MODEL_WORKSPACE} and below needs to be owned by root:root rather than the current UID:GID
+# this allows the default user (root) to work in k8s single-node, multi-node
+RUN umask 002 && mkdir -p ${MODEL_WORKSPACE} && chgrp root ${MODEL_WORKSPACE} && chmod g+s+w,o+s+r ${MODEL_WORKSPACE}
+
+ADD --chown=0:0 ${PACKAGE_DIR}/${PACKAGE_NAME}.tar.gz ${MODEL_WORKSPACE}
+
+RUN chown -R root ${MODEL_WORKSPACE}/${PACKAGE_NAME} && chgrp -R root ${MODEL_WORKSPACE}/${PACKAGE_NAME} && chmod -R g+s+w ${MODEL_WORKSPACE}/${PACKAGE_NAME} && find ${MODEL_WORKSPACE}/${PACKAGE_NAME} -type d | xargs chmod o+r+x 
+
+WORKDIR ${MODEL_WORKSPACE}/${PACKAGE_NAME}
+
+FROM intel-optimized-pytorch AS release
+COPY --from=intel-optimized-pytorch /root/anaconda3 /root/anaconda3
+COPY --from=intel-optimized-pytorch /workspace/lib/ /workspace/lib/
+COPY --from=intel-optimized-pytorch /root/.local/ /root/.local/
+
+ENV DNNL_MAX_CPU_ISA="AVX512_CORE_AMX"
+
+ENV PATH="~/anaconda3/bin:${PATH}"
+ENV LD_PRELOAD="/workspace/lib/jemalloc/lib/libjemalloc.so:$LD_PRELOAD"
+ENV MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000"
+ENV BASH_ENV=/root/.bash_profile
+WORKDIR /workspace/
+RUN yum install -y numactl mesa-libGL && \
+    yum clean all && \
+    echo "source activate pytorch" >> /root/.bash_profile

dockerfiles/pytorch/pytorch-spr-ssd-resnet34-training.Dockerfile

@@ -0,0 +1,88 @@
+# Copyright (c) 2020-2021 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+#
+# THIS IS A GENERATED DOCKERFILE.
+#
+# This file was assembled from multiple pieces, whose use is documented
+# throughout. Please refer to the TensorFlow dockerfiles documentation
+# for more information.
+
+ARG PYTORCH_IMAGE="model-zoo"
+ARG PYTORCH_TAG="pytorch-ipex-spr"
+
+FROM ${PYTORCH_IMAGE}:${PYTORCH_TAG} AS intel-optimized-pytorch
+
+RUN source ~/anaconda3/bin/activate pytorch && \
+    pip install matplotlib Pillow pycocotools && \
+    pip install yacs opencv-python cityscapesscripts transformers && \
+    conda install -y libopenblas psutil && \
+    cd /workspace/installs && \
+    wget https://github.com/gperftools/gperftools/releases/download/gperftools-2.7.90/gperftools-2.7.90.tar.gz && \
+    tar -xzf gperftools-2.7.90.tar.gz && \
+    cd gperftools-2.7.90 && \
+    ./configure --prefix=$HOME/.local && \
+    make && \
+    make install && \
+    rm -rf /workspace/installs/
+
+ARG PACKAGE_DIR=model_packages
+
+ARG PACKAGE_NAME="pytorch-spr-ssd-resnet34-training"
+
+ARG MODEL_WORKSPACE
+
+# ${MODEL_WORKSPACE} and below needs to be owned by root:root rather than the current UID:GID
+# this allows the default user (root) to work in k8s single-node, multi-node
+RUN umask 002 && mkdir -p ${MODEL_WORKSPACE} && chgrp root ${MODEL_WORKSPACE} && chmod g+s+w,o+s+r ${MODEL_WORKSPACE}
+
+ADD --chown=0:0 ${PACKAGE_DIR}/${PACKAGE_NAME}.tar.gz ${MODEL_WORKSPACE}
+
+RUN chown -R root ${MODEL_WORKSPACE}/${PACKAGE_NAME} && chgrp -R root ${MODEL_WORKSPACE}/${PACKAGE_NAME} && chmod -R g+s+w ${MODEL_WORKSPACE}/${PACKAGE_NAME} && find ${MODEL_WORKSPACE}/${PACKAGE_NAME} -type d | xargs chmod o+r+x 
+
+WORKDIR ${MODEL_WORKSPACE}/${PACKAGE_NAME}
+
+RUN source ~/anaconda3/bin/activate pytorch && \
+    pip install --upgrade pip && \
+    pip install --no-cache-dir https://github.com/mlperf/logging/archive/9ea0afa.zip && \
+    pip install --no-cache-dir \
+        Cython==0.28.4 \
+        git+http://github.com/NVIDIA/apex.git@9041a868a1a253172d94b113a963375b9badd030#egg=apex \
+        mlperf-compliance==0.0.10 \
+        cycler==0.10.0 \
+        kiwisolver==1.0.1 \
+        matplotlib==2.2.2 \
+        Pillow==5.2.0 \
+        pyparsing==2.2.0 \
+        python-dateutil==2.7.3 \
+        pytz==2018.5 \
+        six==1.11.0 \
+        torchvision==0.2.1 \
+        pycocotools==2.0.2
+
+FROM intel-optimized-pytorch AS release
+COPY --from=intel-optimized-pytorch /root/anaconda3 /root/anaconda3
+COPY --from=intel-optimized-pytorch /workspace/lib/ /workspace/lib/
+COPY --from=intel-optimized-pytorch /root/.local/ /root/.local/
+
+ENV DNNL_MAX_CPU_ISA="AVX512_CORE_AMX"
+
+ENV PATH="~/anaconda3/bin:${PATH}"
+ENV LD_PRELOAD="/workspace/lib/jemalloc/lib/libjemalloc.so:$LD_PRELOAD"
+ENV MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000"
+ENV BASH_ENV=/root/.bash_profile
+WORKDIR /workspace/
+RUN yum install -y numactl mesa-libGL && \
+    yum clean all && \
+    echo "source activate pytorch" >> /root/.bash_profile

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/container_build.md

@@ -0,0 +1,26 @@
+## Build the container
+
+The <model name> <mode> package has scripts and a Dockerfile that are
+used to build a workload container that runs the model. This container
+uses the PyTorch/IPEX container as it's base, so ensure that you have built
+the `pytorch-ipex-spr.tar.gz` container prior to building this model container.
+
+Use `docker images` to verify that you have the base container built. For example:
+```
+$ docker images | grep pytorch-ipex-spr
+model-zoo         pytorch-ipex-spr         fecc7096a11e        40 minutes ago      8.31GB
+```
+
+To build the <model name> <mode> container, extract the package and
+run the `build.sh` script.
+```
+# Extract the package
+tar -xzf <package name>
+cd <package dir>
+
+# Build the container
+./build.sh
+```
+
+After the build completes, you should have a container called
+`<docker image>` that will be used to run the model.

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/datasets.md

@@ -0,0 +1,33 @@
+## Datasets
+
+### COCO
+
+The [COCO dataset](https://cocodataset.org) is used to run <model name>.
+
+Download and extract the 2017 training/validation images and annotations from the
+[COCO dataset website](https://cocodataset.org/#download) to a `coco` folder
+and unzip the files. After extracting the zip files, your dataset directory
+structure should look something like this:
+```
+coco
+├── annotations
+│   ├── captions_train2017.json
+│   ├── captions_val2017.json
+│   ├── instances_train2017.json
+│   ├── instances_val2017.json
+│   ├── person_keypoints_train2017.json
+│   └── person_keypoints_val2017.json
+├── train2017
+│   ├── 000000454854.jpg
+│   ├── 000000137045.jpg
+│   ├── 000000129582.jpg
+│   └── ...
+└── val2017
+    ├── 000000000139.jpg
+    ├── 000000000285.jpg
+    ├── 000000000632.jpg
+    └── ...
+```
+The parent of the `annotations`, `train2017`, and `val2017` directory (in this example `coco`)
+is the directory that should be used when setting the `DATASET_DIR` environment
+variable for <model name> (for example: `export DATASET_DIR=/home/<user>/coco`).

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/description.md

@@ -0,0 +1,5 @@
+<!-- 10. Description -->
+## Description
+
+This document has instructions for running <model name> <mode> using
+Intel-optimized PyTorch.

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/docker_spr.md

@@ -0,0 +1,33 @@
+## Run the model
+
+Download the pretrained model weights using the script from the MLPerf repo
+and set the `PRETRAINED_MODEL` environment variable to point to the downloaded file:
+```
+wget https://raw.githubusercontent.com/mlcommons/inference/v0.7/others/cloud/single_stage_detector/download_model.sh
+sh download_model.sh
+export PRETRAINED_MODEL=$(pwd)/pretrained/resnet34-ssd1200.pth
+```
+
+After downloading the pretrained model and following the instructions to
+[build the container](#build-the-container) and [prepare the dataset](#datasets),
+use the `run.sh` script from the container package to run <model name> <mode>
+using docker. Set environment variables to specify the dataset directory,
+precision to run, and an output directory for logs. By default, the `run.sh`
+script will run the `inference_realtime.sh` quickstart script. To run a different
+script, specify the name of the script using the `SCRIPT` environment variable.
+```
+# Navigate to the container package directory
+cd <package dir>
+
+# Set the required environment vars
+export DATASET_DIR=<path to the coco dataset>
+export PRETRAINED_MODEL=<path to the resnet34-ssd1200.pth file>
+export PRECISION=<specify the precision to run>
+export OUTPUT_DIR=<directory where log files will be written>
+
+# Run the container with inference_realtime.sh quickstart script
+./run.sh
+
+# To run a difference quickstart script, us the SCRIPT env var
+SCRIPT=accuracy.sh ./run.sh
+```

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/license.md

@@ -0,0 +1,4 @@
+<!--- 80. License -->
+## License
+
+Licenses can be found in the model package, in the `licenses` directory.

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/quickstart.md

@@ -0,0 +1,8 @@
+<!--- 40. Quick Start Scripts -->
+## Quick Start Scripts
+
+| Script name | Description |
+|-------------|-------------|
+| `inference_realtime.sh` | Runs multi instance realtime inference using 4 cores per instance for the specified precision (fp32, int8 or bf16). |
+| `inference_throughput.sh` | Runs multi instance batch inference using 1 instance per socket for the specified precision (fp32, int8 or bf16). |
+| `accuracy.sh` | Measures the inference accuracy (providing a `DATASET_DIR` environment variable is required) for the specified precision (fp32, int8 or bf16). |

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/title.md

@@ -0,0 +1,2 @@
+<!--- 0. Title -->
+# PyTorch <model name> <mode>

quickstart/object_detection/pytorch/ssd-resnet34/inference/cpu/.docs/wrapper_package.md

@@ -0,0 +1,16 @@
+## Model Package
+
+The model package includes the Dockerfile and scripts needed to build and
+run <model name> <mode> in a container.
+```
+<package dir>
+├── README.md
+├── build.sh
+├── licenses
+│   ├── LICENSE
+│   └── third_party
+├── model_packages
+│   └── <package name>
+├── <package dir>.Dockerfile
+└── run.sh
+```