SSD (Single Shot MultiBox Detector) - Tensorflow 2.0

Preparation

• Download PASCAL VOC dataset (2007 or 2012) and extract at ./data
• Install necessary dependencies:

pip install -r requirements.txt

Training

Arguments for the training script:

>> python train.py --help
usage: train.py [-h] [--data-dir DATA_DIR] [--data-year DATA_YEAR]
                [--arch ARCH] [--batch-size BATCH_SIZE]
                [--num-batches NUM_BATCHES] [--neg-ratio NEG_RATIO]
                [--initial-lr INITIAL_LR] [--momentum MOMENTUM]
                [--weight-decay WEIGHT_DECAY] [--num-epochs NUM_EPOCHS]
                [--checkpoint-dir CHECKPOINT_DIR]
                [--pretrained-type PRETRAINED_TYPE] [--gpu-id GPU_ID]

Arguments explanation:

• --data-dir dataset directory (must specify to VOCdevkit folder)

• --data-year the year of the dataset (2007 or 2012)

• --arch SSD network architecture (ssd300 or ssd512)

• --batch-size training batch size

• --num-batches number of batches to train (-1: train all)

• --neg-ratio ratio used in hard negative mining when computing loss

• --initial-lr initial learning rate

• --momentum momentum value for SGD

• --weight-decay weight decay value for SGD

• --num-epochs number of epochs to train

• --checkpoint-dir checkpoint directory

• --pretrained-type pretrained weight type (base: using pretrained VGG backbone, other options: see testing section)

• --gpu-id GPU ID

• how to train SSD300 using PASCAL VOC2007 for 100 epochs:

python train.py --data-dir ./data/VOCdevkit --data-year 2007 --num-epochs 100

• how to train SSD512 using PASCAL VOC2012 for 120 epochs on GPU 1 with batch size 8 and save weights to ./checkpoints_512:

python train.py --data-dir ./data/VOCdevkit --data-year 2012 --arch ssd512 --num-epochs 120 --batch-size 8 --checkpoint_dir ./checkpoints_512 --gpu-id 1

Testing

Arguments for the testing script:

>> python test.py --help
usage: test.py [-h] [--data-dir DATA_DIR] [--data-year DATA_YEAR]
               [--arch ARCH] [--num-examples NUM_EXAMPLES]
               [--pretrained-type PRETRAINED_TYPE]
               [--checkpoint-dir CHECKPOINT_DIR]
               [--checkpoint-path CHECKPOINT_PATH] [--gpu-id GPU_ID]

Arguments explanation:

• --data-dir dataset directory (must specify to VOCdevkit folder)

• --data-year the year of the dataset (2007 or 2012)

• --arch SSD network architecture (ssd300 or ssd512)

• --num-examples number of examples to test (-1: test all)

• --checkpoint-dir checkpoint directory

• --checkpoint-path path to a specific checkpoint

• --pretrained-type pretrained weight type (latest: automatically look for newest checkpoint in checkpoint_dir, specified: use the checkpoint specified in checkpoint_path)

• --gpu-id GPU ID

• how to test the first training pattern above using the latest checkpoint:

python test.py --data-dir ./data/VOCdevkit --data-year 2007 --checkpoint_dir ./checkpoints

• how to test the second training pattern above using the 100th epoch's checkpoint, using only 40 examples:

python test.py --data-dir ./data/VOCdevkit --data-year 2012 --arch ssd512 --checkpoint_path ./checkpoints_512/ssd_epoch_100.h5 --num-examples 40

Reference

• Single Shot Multibox Detector paper: paper
• Caffe original implementation: code
• Pytorch implementation: [code] (https://github.com/ChunML/ssd-pytorch)