climber0824 / nctu_hw3

Instance segmentation using tiny dataset

NCTU_hw3

Instance segmentation using tiny dataset

Environments:
Python 3.6
PyTorch 1.5

File structure:
+- data
| - config.py: the setting of backbone of model and the path of dataset
| - dataloader.py: using to download training and testing data
+- models
| - backbone.py: the backbone(resnet101+FPN) of this model
| - box_utils.py: using to calculate IOU of bbox or mask
| - detection.py: using to calculate NMS
| - interpolate.py: class to interpolate
| - model.py: the whole predict model
| - multibox_loss.py: using to calculate the bbox loss, label loss, mask loss, and segmentation loss
| - output_utils.py: post processing for the output of model
+- utils
| - augmentation.py: function for data augmentation
| - functions.py: using to print processing bar
| - timer.py: using to print training time
train.py: using to train the model
predict.py: using to predict the mask of input images
make_submission.py: using to output .json file for submit
config.yaml: setting for training, predict, and make submission

Usage:
1.Data preparatoin:
Download the data from:https://drive.google.com/drive/folders/1fGg03EdBAxjFumGHHNhMrz2sMLLH04FK
The dataset structure is the same as below:
dataset
+- train_images
| - 2007_000033.jpg
| - 2007_000042.jpg
| - 2007_000061.jpg
| - ...
+- test_images
| - 2007_000629.jpg
| - 2007_001157.jpg
| - 2007_001239.jpg
| - ...
pascal_train.json
test.json

2.Training:
You can set up the related setting in config.yaml and run the following command to train:
$ python3 train.py --config=config.yaml

3.Prediction:
You can set up the related setting in config.yaml and run the following command to predict:
$ python3 predict.py --config=config.yaml

4.Make Submission:
You can set up the related setting in config.yaml and run the following command to make submission file:
$ python3 make_submission.py --config=config.yaml

Reference: https://github.com/dbolya/yolact