-
AWS p3.2xlarge(Tesla V100 16GB 300W)
- Ubuntu 16.04
- docker-ce
- nvidia-docker
- nvidia/cuda
- Python 3.6.3/OpenCV 3.3.1/Tensorflow r1.4.1 (build from source)
- Tensorflow Object Detection API (branch r1.5)
- nvidia-docker
- nvidia/cuda:9.0-devel-ubuntu16.04
- Python 2.7.12/OpenCV 3.4.2/Tensorflow r1.8.0 (build from source)
- Tensorflow Object Detection API (branch master)
-
PC
- CPU: i7-8700 3.20GHz 6-core 12-threads
- GPU: NVIDIA GTX1060 6GB 120W
- MEMORY: 32GB
- Ubuntu 16.04
- docker-ce
- nvidia-docker
- nvidia/cuda:9.0-devel-ubuntu16.04
- Pyton 2.7.12/OpenCV 3.4.2/Tensorflow 1.8.0 (build from source)
- Tensorflow Object Detection API (branch master)
- Create PascalVOC data from your JPG images.
- Setting Tensorflow Object Detection API.
- Create TF Record data from PascalVOC data.
- Training.
- Freeze Graph.
- Run.
- Install LabelImg on your Ubuntu desktop PC.
- For Windows PC you can use VMware Player.
- Make labels by hand.
- Move files into dirs.
- Make label_map.pbtxt.
- Upload to training machine.
Install LabelImg on your Ubuntu desktop PC.
Install LabelImg.
mkdir ~/github
sudo apt-get install -y pyqt4-dev-tools
sudo apt-get install -y python-pip
sudo pip install --upgrade pip
sudo pip install lxml
cd ~/github
git clone https://github.com/tzutalin/labelImg
cd ~/github/labelImg
make qt4py2
Make all image's label with LabelImg.
cd ~/github/labelImg
./labelImg.py
Divide directory of jpg file and xml file.
mkdir ~/roadsign_data/PascalVOC/JPEGImages
mkdir ~/roadsign_data/PascalVOC/Annotations
# in your data dir
mv *.jpg ~/roadsign_data/PascalVOC/JPEGImages
mv *.xml ~/roadsign_data/PascalVOC/Annotations
Make your label_map file like this.
file:./roadsign_data/roadsign_label_map.pbtxt
item {
id: 1
name: 'stop'
}
item {
id: 2
name: 'speed_10'
}
item {
id: 3
name: 'speed_20'
}
item {
id: 4
name: 'speed_30'
}
Copy the data to training machine.
Example:
scp -r ~/roadsign_data training_machine:~/github/train_ssd_mobilenet/
- git clone Tensorflow Object Detection API.
- Edit exporter.py for Tensorflow r1.4.1. (for tensorflow/models branch r1.5)
- Build protocol buffer.
- Download checkpoint of ssd_mobilenet.
- Make your pipeline config file.
- Install pycocotools (for tensorflow/models branch master)
Branch r1.5.
cd ~/github
git clone https://github.com/tensorflow/models
cd models/
git fetch
git checkout r1.5
You can use master branch, but it occasionally causes an error.
If you want to run on r1.4.1, you need to fix this problem.
ValueError: Protocol message RewriterConfig has no "layout_optimizer" field.
tensorflow/tensorflow#16268
Edit ~/github/models/research/object_detection/exporter.py L:71-72
rewrite_options = rewriter_config_pb2.RewriterConfig()
sudo apt-get install -y protobuf-compiler
cd ~/github/models/research
protoc object_detection/protos/*.proto --python_out=.
For tensorflow/models master branch, protobuf version 3 is required.
If you get an error please install protobuf version 3 first.
Install protobuf 3 on Ubuntu
# Make sure you grab the latest version
curl -OL https://github.com/google/protobuf/releases/download/v3.2.0/protoc-3.2.0-linux-x86_64.zip
# Unzip
unzip protoc-3.2.0-linux-x86_64.zip -d protoc3
# Move protoc to /usr/local/bin/
sudo mv protoc3/bin/* /usr/local/bin/
# Move protoc3/include to /usr/local/include/
sudo mv protoc3/include/* /usr/local/include/
# Optional: change owner
sudo chwon [user] /usr/local/bin/protoc
sudo chwon -R [user] /usr/local/include/google
Download checkpoint from here. https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
cd ~/github/train_ssd_mobilenet/
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2017_11_17.tar.gz
tar xvf ssd_mobilenet_v1_coco_2017_11_17.tar.gz
In tensorflow/models master branch, you can train new ssd models.
Copy sample config.
cp ~/github/models/research/object_detection/samples/configs/ssd_mobilenet_v1_coco.config ~/github/train_ssd_mobilenet/ssd_mobilenet_v1_roadsign.config
Edit your pipeline config like this.
pipeline config: ssd_mobilenet_v1_roadsign.config
diff -u ~/github/models/research/object_detection/samples/configs/ssd_mobilenet_v1_coco.config ~/train_ssd_mobilenet/ssd_mobilenet_v1_roadsign.config
--- /home/ubuntu/github/models/research/object_detection/samples/configs/ssd_mobilenet_v1_coco.config 2017-12-20 11:46:42.832787513 +0900
+++ /home/ubuntu/github/train_ssd_mobilenet/ssd_mobilenet_v1_roadsign.config 2018-03-19 11:22:10.521440000 +0900
@@ -6,7 +6,7 @@
model {
ssd {
- num_classes: 90
+ num_classes: 4
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
@@ -155,7 +155,7 @@
epsilon: 1.0
}
}
- fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED/model.ckpt"
+ fine_tune_checkpoint: "ssd_mobilenet_v1_coco_2017_11_17/model.ckpt"
from_detection_checkpoint: true
# Note: The below line limits the training process to 200K steps, which we
# empirically found to be sufficient enough to train the pets dataset. This
@@ -174,9 +174,9 @@
train_input_reader: {
tf_record_input_reader {
- input_path: "PATH_TO_BE_CONFIGURED/mscoco_train.record"
+ input_path: "roadsign_data/tfrecords/train.record"
}
- label_map_path: "PATH_TO_BE_CONFIGURED/mscoco_label_map.pbtxt"
+ label_map_path: "roadsign_data/roadsign_label_map.pbtxt"
}
eval_config: {
@@ -188,9 +188,9 @@
eval_input_reader: {
tf_record_input_reader {
- input_path: "PATH_TO_BE_CONFIGURED/mscoco_val.record"
+ input_path: "roadsign_data/tfrecords/val.record"
}
- label_map_path: "PATH_TO_BE_CONFIGURED/mscoco_label_map.pbtxt"
+ label_map_path: "roadsign_data/roadsign_label_map.pbtxt"
shuffle: false
num_readers: 1
num_epochs: 1
cd ~/github
git clone https://github.com/pdollar/coco.git
cd coco/PythonAPI
make
sudo make install
sudo python setup.py install
cd ~/github/models/
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/research/slim:`pwd`/research:
Check config.yml.
sudo pip install lxml pyyaml
cd ~/github/train_ssd_mobilenet
# Please check config.yml
python build1_trainval.py
python build2_tf_record.py
--train_dir: output directory.
--logtostderr: log to stderror.
--pipeline_config_path: model config file.
cd ~/github/train_ssd_mobilenet
# training/continue from checkpoint
python ~/github/models/research/object_detection/train.py --logtostderr --train_dir=./train --pipeline_config_path=./ssd_mobilenet_v1_roadsign.config
In tensorflow/models master branch, you can train with new ssd models.
In this case, I downloaded ssdlite_mobilenet_v2 and edit config like ssd_mobilenet_v1.
In master branch, local training command has changed.
Running Locally
--model_dir: output directory.
--alsologtostderr: log to stderror.
--pipeline_config_path: model config file.
cd ~/github/train_ssd_mobilenet
# training/continue from checkpoint
python ~/github/models/research/object_detection/model_main.py --alsologtostderr --model_dir=train --pipeline_config_path=./ssdlite_mobilenet_v2_roadsign.config
cd ~/github/train_ssd_mobilenet
python ~/github/models/research/object_detection/eval.py --logtostderr \
--checkpoint_dir=./train \
--eval_dir=eval \
--pipeline_config_path=./ssd_mobilenet_v1_roadsign.config
# If you have output dir, please remove it first.
rm -rf ./output/
# Please change to your checkpoint file.: ./train/model.ckpt-11410
python ~/github/models/research/object_detection/export_inference_graph.py --input_type image_tensor --pipeline_config_path=./ssd_mobilenet_v1_roadsign.config --trained_checkpoint_prefix ./train/model.ckpt-11410 --output_directory ./output \
--config_override " \
model{ \
ssd { \
post_processing { \
batch_non_max_suppression { \
score_threshold: 0.5 \
} \
} \
} \
}"
# if you want to strip more, then execute next.
python ./freeze_graph.py
ls -l ./output/
freeze_graph.pb and roadsign_label_map.pbtxt are required for running.
You can use Tensorflow realtime_object_detection.
