Skip to content

Latest commit

 

History

History
229 lines (158 loc) · 6.18 KB

GETTING_STARTED.md

File metadata and controls

229 lines (158 loc) · 6.18 KB
Raw

Getting Started

This page provides basic tutorials about the usage of MMFashion.

Inference with pretrained models

We provide testing scripts to evaluate a whole dataset (Category and Attribute Prediction Benchmark, In-Shop Clothes Retrieval Benchmark, Fashion Landmark Detection Benchmark etc.), and also some high-level apis for easier integration to other projects.

Test an image

You can use the following commands to test an image.

python demo/test_*.py --input ${INPUT_IMAGE_FILE}

Examples:

Assume that you have already downloaded the checkpoints to checkpoints/.

  1. Test an attribute predictor(coarse prediction).

    # Prepare `Anno/list_attr_cloth.txt` which is specified in `configs/attribute_predict/global_predictor_vgg_attr.py`
python demo/test_attr_predictor.py \
    --input demo/imgs/attr_pred_demo1.jpg

    Test a category and attribute predictor(more accurate prediction).

     # Prepare `Anno/list_attr_cloth.txt` which is specified in `configs/category_attribute_predict/global_predictor_vgg_attr.py`
 python demo/test_cate_attr_predictor.py \
     --input demo/imgs/attr_pred_demo1.jpg

  2. Test an in-shop / Consumer-to_shop clothes retriever.

    # Prepare the gallery data which is specified in `configs/retriever_in_shop/global_retriever_vgg_loss_id.py`
python demo/test_retriever.py \
    --input demo/imgs/retrieve_demo1.jpg

  3. Test a landmark detector.

    python demo/test_landmark_detector.py \
    --input demo/imgs/04_1_front.jpg

  4. Test a fashion-compatibility predictor.

    python demo/test_fashion_recommender.py \
    --input_dir demo/imgs/fashion_compatibility/set2

Test a dataset

You can use the following commands to test a dataset.

python tools/test_*.py --config ${CONFIG_FILE} --checkpoint ${CHECKPOINT_FILE}

Examples:

Assume that you have already downloaded the checkpoints to checkpoints/ and prepared the dataset in data/.

  1. Test an attribute predictor.

    python tools/test_predictor.py \
    --config configs/attribute_predict/roi_predictor_vgg_attr.py \
    --checkpoint checkpoint/Predict/vgg/roi/latest.pth

    Test a category and attribute predictor.

    python test tools/test_cate_attr_predictor.py \
     --config configs/category_attribute_predict/roi_predictor_vgg.py \
     --checkpoint checkpoint/CateAttrPredict/vgg/roi/latest.pth

  2. Test an in-shop / Consumer-to_shop clothes retriever.

    python tools/test_retriever.py \
    --config configs/retriever_in_shop/roi_retriever_vgg.py \
    --checkpoint checkpoint/Retrieve_in_shop/vgg/latest.pth
    python tools/test_retriever.py \
    --config configs/retriever_consumer_to_shop/roi_retriever_vgg.py \
    --checkpoint checkpoint/Retrieve_consumer_to_shop/vgg/latest.pth

  3. Test a landmark detector.

    python tools/test_landmark_detector.py \
    --config configs/landmark_detect/landmark_detect_vgg.py
    --checkpoint checkpoint/LandmarkDetect/vgg/latest.pth

  4. Test a fashion-compatibility predictor.

    python tools/test_fashion_recommender.py \
    --config configs/fashion_recommendation/type_aware_recommendation_polyvore_disjoint.py
    --checkpoint checkpoint/FashionRecommend/TypeAware/latest.pth

  5. Test a virtual try-on module.

    Step 1, use the geometric matching module(GMM) to generate warp-cloth and warp-mask,

     python tools/test_virtual_tryon.py \
     --config configs/virtual_tryon/cp_vton.py \
     --stage GMM

    Step 2, use the tryon module(TOM) to generate the results. The default result directory is data/VTON/result, you can modify the path in config file(line 103).

     python tools/test_virtual_tryon.py \
     --config configs/virtual_tryon/cp_vton.py \
     --stage TOM

Train a model

You can use the following commands to train a model.

python tools/train_*.py --config ${CONFIG_FILE}

Examples:

  1. Train an attribute predictor.

    python tools/train_predictor.py \
    --config configs/attribute_predict/roi_predictor_vgg_attr.py

  2. Train an in-shop clothes / Consumer-to-shop retriever.

    python tools/train_retriever.py \
    --config configs/retriever_in_shop/roi_retriever_vgg.py
    python tools/train_retriever.py \
    --config configs/retriever_consumer_to_shop/roi_retriever_vgg.py

  3. Train a landmark detector.

    python tools/train_landmark_detector.py \
    --config configs/landmark_detect/landmark_detect_vgg.py

  4. Train a fashion-compatibility predictor.

    python tools/train_fashion_recommender.py \
    --config configs/fashion_recommendation/type_aware_recommendation_polyvore_disjoint.py

  5. Train a fashion detector.

    python mmdetection/tools/train.py \
    configs/fashion_parsing_segmentation/mask_rcnn_r50_fpn_1x.py

  6. Train a virtual try-on module.

    Step 1, train a geometric matching module(GMM)

     python tools/train_virtual_tryon.py \
     --config configs/virtual_tryon/cp_vton.py \
     --stage GMM

    After training GMM, you need to generate the training images which are required by Step 2.

    Note that you need to modify the config file configs/virtual_tryon/cp_vton.py as follows,

    Line 87 should be datamode='train',

    Line 89 should be data_list='test_pairs.txt',

    Line 93 should be save_dir=os.path.join(data_root, 'vton_resize', 'train')

    Then,

     python tools/test_virtual_tryon.py \
     --config configs/virtual_tryon/cp_vton.py \
     --stage GMM

    Step 2, train a tryon module(TOM),

     python tools/train_virtual_tryon.py \
     --config configs/virtual_tryon/cp_vton.py \
     --stage TOM

Use custom datasets

The simplest way is to prepare your dataset to existing dataset formats (AttrDataset, InShopDataset, ConsumerToShopDataset or LandmarkDetectDataset).

Please refer to DATA_PREPARATION.md for the dataset specifics.