iMaterialist-2020

This is a segmentation framework modified from MaskRCNN.

requirements

• Python (>= 3.6)
• PyTorch pip3 install -r requirements.txt

Demo

we provide a clothing parsing system using our trained model as segmentation model.

Train

This code support two types of model: without attribute and with attribute

./train.sh (lr schedule) (lr) # train without attr
#./train_attr.sh (attribute-wise weight) (attribute-loss weight) (lr) # train with attr

Val

predict on validation data (output mIOU and mF1)

./val.sh (experiment_name) (checkpoint_idx)

Test

test without attribute

# ./test.sh experiment_name mask_thresh checkpoint_idx
./test.sh torch_MaskRCNN_40e_lr0.1 0.5 21

test with attribute

# ./test_attr.sh experiment_name mask_thresh checkpoint_idx attr_score_thresh
./test_attr.sh torch_MaskRCNN_20e_lr0.01_attr_1000weight_3aweight 0.5 0 0.7

code log

train.py # train with train/val.csv, not support for online validation (validate while training)。[something wrong in code, seems to be the cuda device incorrespondence] train_attr.py # train with attribute data，using binary cross entropy with specified pos weight and loss weight.