mosdragon / faster-rcnn.pytorch

A faster pytorch implementation of faster r-cnn

A Faster Pytorch Implementation of Faster R-CNN

Download Repo:

git clone https://github.com/mosdragon/faster-rcnn.pytorch
cd faster-rcnn.pytorch

Compile Cuda Code

cd lib
bash make.sh  # If it crashes, modify the $CUDA_PATH

Download Pre-trained VGG16 Network

mkdir -p data/pretrained_model/
wget https://filebox.ece.vt.edu/~jw2yang/faster-rcnn/pretrained-base-models/vgg16_caffe.pth -O data/pretrained_model/vgg16_caffe.pth

Setup faster-rcnn to work with SUMO

cd data
mkdir VOCdevkit2007
ln -s /path/to/dataset_sumo VOCdevkit2007/VOC2007

Install Dependencies

python3 -m venv frcnn_venv
source frcnn_venv/bin/activate

pip install -r requirements.txt

Training:

Note: Run this inside tmux or screen.

Use the Virtualenv

source frcnn_venv/bin/activate

Set ENV Vars

export DATASET="pascal_voc"  # We symlinked sumo as pascal earlier.
export BATCH_SIZE=4
export NW=4  # Set this to number of GPUs. 4 works well for Skynet.
export LR="1e-3"
export LR_DECAY="5"
export MAX_EPOCHS="9"  # This should take 48 hours.

We get a model built that we can run anytime after each epoch (about 5 hours each). We can set $MAX_EPOCH to something large above, and use model from Epoch 5 or something and let the network continue training behind the scenes unless we have no more GPUs to use for inference.

Begin Training

echo "Training"
time python trainval_net.py --dataset ${DATASET} \
  --net vgg16 \
  --bs ${BATCH_SIZE} --nw ${NW} \
  --lr ${LR} --lr_decay_step ${LR_DECAY} \
  --epochs ${MAX_EPOCHS}
  --cuda

Once this is running, we can go back and generate the test partitions in the SUMO dataset. Those will be useful to us later.

After Training

As each epoch finishes, we will find a trained model file under models/pascal_voc/vgg16/faster_rcnn_${session}_${epoch}_${step}.pth. This is a model we can then use to train.

Using a Small Sample for Training

Anytime we modify our training data, me must remove the cache this repo creates.

cd data/cache/
# You will see something like pascal_voc2007_train.pkl. Delete that file.

We can just create new temporary train.txt and trainval.txt files.

cd /path/to/dataset_sumo
cd ImageSets/Main

# Backup the full size train/trainval file.
mkdir backup
cp train.txt backup/
cp trainval.txt backup/

# Take first 100 samples from train/trainval files.
head -n 100 backup/train.txt > train.txt
head -n 100 backup/trainval.txt > trainval.txt

Now run the training steps above. Undo these steps and delete cache again when done experimenting with small sample size.