Note: please install nightly build gluoncv for less bugs
pip install gluoncv --pre --upgrade
iMaterialist (Fashion) 2019 at FGVC6
- Convert imaterialist to COCO style
python tools/convert_imaterialist_to_coco.py --data_root datasets/imaterialist
- files structure
datasets
└─imaterialist
├── annotations
│ ├── resize_rle_instances_val.json
│ ├── rle_instances_train.json
| └── rle_instances_val.json
├── test
├── train
├── val
├── label_descriptions.json
├── sample_submission.csv
└── train.csv
Note: all using COCO pretrained models, only the categories less than 13 (main apparels) for segmentation.
python train_mask_rcnn.py --save-prefix train_logs/xxx/ --val-interval 1
python train_segmentation.py --model deeplabv3 --dataset imaterialist \
--model-zoo deeplab_resnet101_coco --aux \
--checkname res101 --epochs 30 --lr 0.001 --ngpus 1 \
--workers 2 --batch-size 2 --test-batch-size 2
python train_segmentation.py --model deeplabv3plus --dataset imaterialist \
--model-zoo deeplab_plus_xception_coco --aux \
--base-size 576 --crop-size 512 \
--checkname xception --epochs 30 --lr 0.001 --ngpus 1 \
--workers 1 --batch-size 1 --test-batch-size 1