Introduction

Codes for RSIPAC track#2. Leave an issue if you have any problems.

How to run

Data perparing

Data directory should be like:

.
└── data
    ├── test
    │   ├── A
    │   ├── B
    │   └── instances_test.json
    └─── train
        ├── A
        ├── B
        └── label

Config

./configs/base.yaml is a baseline config, you can refer this for further development.

The first several lines define some global items and reusable items. Like name and version, these two is used for model log saving. Specifically, model checkpoints and logs will be saved to ./logs/{name}/{version}/*.{ckpt|yaml|csv}.

The following sections define each component in training:

1. model

I have implemented some common segmentation models by default, like segmentation_models_pytorch (smp) and mmsegmentation (mmseg). The superior model type is defined by model.type item in YAML.

1.1 SMPModel

For smp, I always use its Unet, so by default it will give you a Unet model. You can change this and other optional arguments in ./src/models/smp_models.py.

A reference config is like:

model:
  type: SMPModel
  model_type: Unet
  model_name: timm-efficientnet-b0
  pretrained_weight: noisy-student
  num_classes: 1

1.2 MMSegModel

For mmseg, the config defination is almost the same, but I added and deleted some components for my personal perferrence.

I used this instead of original mmseg is because it is time-consuming to install mmseg on Kaggle platform, so I channged every module that is imported from mmcv into pytorch native version. You can easily convert an mmseg model file into this repository by changing the import path.

The converted mmseg is at ./src/models/mmseg. There are already some files, including most frequently used segformer's MiT, segformer_head and uper_head, but not fully covered the whole mmseg.

For example:

model:
  type: MMSegModel
  backbone:
    type: timm
    model_name: resnet50
    pretrained: True
    in_chans: 6
  decode_head:
    type: UPerHead
    pool_scales: [1, 2, 3, 6]
    channels: 512
    dropout_ratio: 0.1
    num_classes: 1
    norm_cfg: {type: BN, requires_grad: True}
    align_corners: False

Here, I added timm backbone for every model that supports feature extraction. Arguments excepts type: timm will be passed to timm.create_model().

2. others

Code is simple, and you can refer the base.yaml and ./src/* for your custum changing.

Train

GPUS=0 # or 0,1 for multi-gpu training
python Solver --config /path/to/config --gpus $GPUS

Submission

First, define names in Submission.py, which is a list of your model names, that is, the name item in model config YAML file.

python Submission.py

The results will be saved at ./results/test.segm.json and ./results.zip.