This is a Tensorflow 2.0 implementation of the paper YOLACT: Real-time Instance Segmentation accepted in ICCV2019. The paper presents a fully-convolutional model for real-instance segmentation based on extending the existing architecture from object detection and its own idea of parallel prototype generation. In this Repo, I focus on reproducing the result by implementing one of the structure "ResNet50-FPN" on MS-COCO datasets. The part for training this model is ready, and the part for inference and mAP evaluation will be updated soon.
Here is the illustration of YOLACT from original paper.
COCO Dataset is used for reproducing the experiment here.
2017 Train images / 2017 Val images / 2017 Annotations
In this repo, we convert images and annotations into TFRecord through the /data/create_coco_tfrecord.py. In this script, I directly resize the image to 550 * 550 and ignore the images with only crowd annotations. Using the following command to create TFRecord.
python -m data.create_coco_tfrecord -train_image_dir 'path of train2017' -val_image_dir 'path of val2017' -train_annotations_file 'path of train annotations' -val_annotations_file 'path of val annotations' -output_dir 'path for output TFRecord'Training procedure can be conducted directly by following command:
python train.py -tfrecord_dir 'path of TFRecord files'
-weights 'path to store weights'
-train_iter 'number of iteration for training'
-batch_size 'batch_size'
-lr 'learning rate'
-momentum 'momentum for SGD'
-weight_decay 'weight_decay rate for SGD'
-print_interval 'interval for printing training result'
-save_interval 'interval for conducting validation'
-valid_iter 'number of iteration for validation'The default hyperparameters in train.py follows the original setting from the paper:
- Batch size = 8, which is recommanded by paper
- SGD optimizer with learning rate 1e-3 and divided by 10 at iterations 280K, 600K, 700K and 750K, using a momentum 0.9, a weight decay 5* 1e-4. In the original implementation of paper, a warm up learning rate 1e-4 and warm up iterations 500 are used, I put all those setting in a learning schedule object in utils/learning_rate_schedule.py.
- Random photometrics distortion, horizontal flip(mirroring) and crop are used here for data augmentation.
In Tensorflow 2.0, distibuted training with multiple GPU and TPU are straighforward to use by adding different strategy scopes, the info can be find here Distributed training with TensorFlow
- HSU, CHIH-CHAO - Professional Machine Learning Master Student at Mila