Add object detection sample

PyTorch/.gitignore

@@ -6,4 +6,7 @@ traces/
 *.xlsx
 data/cifar-10-python
 checkpoints
-coco128
+coco128
+data/PennFudanPed
+data/PennFudanPed.zip
+objectDetection/PennFudanPed

PyTorch/README.md

@@ -35,6 +35,7 @@ The following sample models are included in this repo to help you get started. T
 
 * [squeezenet - a small image classification model](./squeezenet)
 * [resnet50 - an image classification model](./resnet50)
+* [maskrcnn - an object detection model](./objectDetection/maskrcnn/)
 * *more coming soon*
 
 ## External Links

PyTorch/data/dataset.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python
+# Copyright (c) Microsoft Corporation. All rights reserved.
+
+import argparse
+import pathlib
+import os
+from torchvision import datasets
+import wget
+import zipfile
+
+def get_current_dir():
+    return str(pathlib.Path(__file__).parent.resolve())
+
+def download_cifar_dataset():
+    path = os.path.join(get_current_dir(), 'cifar-10-python')
+    datasets.CIFAR10(root=path, download=True)
+
+def download_pennfudanped_dataset():
+    path = get_current_dir()
+    if (os.path.exists(os.path.join(path, 'PennFudanPed'))):
+        print ("PennFundaPed dataset already downloaded and verified")
+        return
+
+    url='https://www.cis.upenn.edu/~jshi/ped_html/PennFudanPed.zip'
+    print("Downloading PennFundaPed dataset\n")
+    dataset_path = wget.download(url, out=path)
+    try:
+        with zipfile.ZipFile(os.path.join(path, dataset_path)) as z:
+            z.extractall(path=path)
+            print("\nExtracted PennFundaPed dataset")
+    except:
+        print("Invalid file")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(__doc__)
+    parser.add_argument("--dataset", help="datasets: cifar10 or pennfudanped.", default="all")
+    args = parser.parse_args()
+
+    if args.dataset.lower() == 'all':
+        download_cifar_dataset()
+        download_pennfudanped_dataset()
+    elif args.dataset.lower() == 'cifar10':
+        download_cifar_dataset()
+    elif args.dataset.lower() == 'pennfudanped':
+        download_pennfudanped_dataset()
+    else:
+        raise Exception(f"Model {args.dataset} is not supported yet!")

PyTorch/objectDetection/maskrcnn/README.md

@@ -0,0 +1,51 @@
+# maskrcnn Model <!-- omit in toc -->
+
+Sample scripts for training the [Mask R-CNN](https://arxiv.org/abs/1703.06870) model in the [Penn-Fudan Database for Pedestrian Detection and Segmentation](https://www.cis.upenn.edu/~jshi/ped_html/) using PyTorch on DirectML 
+
+These scripts are collected from the tutorial [here](https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html)
+
+- [Setup](#setup)
+- [Prepare Data](#prepare-data)
+- [Training](#training)
+
+## Setup
+Install the following prerequisites:
+```
+pip install -r pytorch\objectDetection\maskrcnn\requirements.txt 
+```
+
+## Prepare Data
+
+After installing the PyTorch on DirectML package (see [GPU accelerated ML training](..\readme.md)), open a console to the `root` directory and run the setup script to download and convert data:
+
+```
+python pytorch\data\dataset.py
+```
+
+Running `dataset.py` should take at least a minute or so, since it downloads the CIFAR-10 dataset. The output of running it should look similar to the following:
+
+```
+>python pytorch\data\dataset.py
+Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to E:\work\dml\PyTorch\data\cifar-10-python\cifar-10-python.tar.gz
+Failed download. Trying https -> http instead. Downloading http://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to E:\work\dml\PyTorch\data\cifar-10-python\cifar-10-python.tar.gz
+170499072it [00:32, 5250164.09it/s]
+Extracting E:\work\dml\PyTorch\data\cifar-10-python\cifar-10-python.tar.gz to E:\work\dml\PyTorch\data\cifar-10-python
+```
+
+## Training
+
+A helper script exists to train Mask R-CNN with PennFudanPed data:
+
+```
+cd pytorch\objectdetection\maskrcnn
+python .\maskrcnn.py
+```
+
+The first few lines of output should look similar to the following (exact numbers may change):
+```
+>python .\maskrcnn.py
+python .\maskrcnn.py
+Epoch: [0]  [ 0/60]  eta: 0:38:26  lr: 0.000090  loss: 2.9777 (2.9777)  loss_classifier: 0.7217 (0.7217)  loss_box_reg: 0.0754 (0.0754)  loss_mask: 1.6228 (1.6228)  loss_objectness: 0.4175 (0.4175)  loss_rpn_box_reg: 0.1404 (0.1404)  time: 38.4439  data: 1.0955
+Epoch: [0]  [10/60]  eta: 0:29:44  lr: 0.000936  loss: 2.4268 (2.4919)  loss_classifier: 0.4056 (0.4158)  loss_box_reg: 0.1691 (0.3631)  loss_mask: 1.1679 (1.1600)  loss_objectness: 0.1162 (0.3120)  loss_rpn_box_reg: 0.1257 (0.2410)  time: 35.6972  data: 0.1034
+Epoch: [0]  [20/60]  eta: 0:23:14  lr: 0.001783  loss: 1.2172 (1.6717)  loss_classifier: 0.0669 (0.2410)  loss_box_reg: 0.1331 (0.2466)  loss_mask: 0.5935 (0.8376)  loss_objectness: 0.0565 (0.1873)  loss_rpn_box_reg: 0.0574 (0.1593)  time: 34.6860  data: 0.0042
+```