Skip to content

Latest commit

 

History

History
151 lines (108 loc) · 4.5 KB

README.md

File metadata and controls

151 lines (108 loc) · 4.5 KB
Raw

Experiment on VOC Part

Dir structure

The dir structure is

./
    requirements.txt  python environment requirements
    starting.sh       download dataset, checkpoints, and set __result__/ link.

    train/       training models
    VOC/         dataset preprocess and loading
    gradmap/     evaluating with gradient map
    activamap/   evaluating with cam/filter activation map
    mutualinfo/  evaluating with mutual information score

    checkpoints/ checkpoints in our paper
        CSG.pt   training with our CSG method
        STD.pt   training with standard method
    __data__/    -- soft link --> dataset dir
    __result__/  -- soft link --> result dir, saving experiment logs, checkpoints
        VOCPart_128x128_pretrained/
            <experiment-id>/
                <seed>/
                    checkpoints/
                    events.out.tfevents.xxx
                    log

Device

GPU Memory requirement: >=5000MB

Environment

  • Apply the python environment:

    virtualenv -p python3 venv  # we use Python 3.7.3
source venv/bin/activate venv
pip install -r requirements.txt

  • Download checkpoints, download preprocessed dataset, and set result dir:

    bash ./starting.sh

    This will

    • download the checkpoints reported in our paper to checkpoints/{STD,CSG}.pt
    • download the preprocessed dataset VOCPart used in our papaer to preprocessed_VOC_Part/ and set __data__ linked to it
    • make a dir to save trianing result and set __result__ linked to it

Datasetsss

Preprocessed dataset

The datatset we used is VOC Part dataset, which is preproceed from Pascal VOC 2010 dataset. To reproduce our paper, please download and use the preprocessed dataset with bash ./starting.sh. Then you will get the preprocessed VOC Part dataset as below:

```
./
    __data__/ --> preprocessed_VOC_Part/
    preprocessed_VOC_Part/
        processed/
            128x128/
                obj_img/
                obj_mask/
                part_mask/
            metadata/
                train.txt
                val.txt
```

Preprocessing by yoursef

If you want to preprocess Pascal VOC 2010 dataset and get VOC Part by yourself, you can run the commands below. Notice that the VOC Part dataset preprocessed by yourself might be slightly different from the aforementioned "preprocessed dataset" used in our paper, because we fix some bugs in preprocessing codes in this version.

  • Download Pascal VOC 2010 dataset dataset:

    bash VOC/download.sh

  • Then preprocess the dataset with so as to create __data__/processed/:

    python VOC/preprocess.py

  • Then you will get a dataset dir as below:

    ./
    __data__/ --> <dataset-dir>
        raw/
            VOC2010/    # traininig and validation data
            metadata/   # Pascal VOC part annotation
        processed/
            metadata/   # the division of training and validation set
            128x128/    # processed images are 128x128 pixel
                obj_img/       # cropped and resized images for each object instance
                obj_mask/      # cropped and resized segmentation ground truth for each object instance
                part_mask/     # cropped and resized segmentation ground truth for each part in an object instance

Training

To approximately reproduce the training process in our paper, please run the code below

  • Set the "training settings" in train/run.py as following, which is the default setting.

    gpu_ids = '0'
ifmask = True
train = True

    set ifmask = True and ifmask = False to train CSG/STD CNN respectively.

  • Run the following codes to train a CSG/STD CNN. It will report the classification accuracy on testing set.

    python train/run.py

Evaluation

Run the commands below to reproduce the evaluation on the mutual information score and the localization performance in our paper. This will evaluate the checkpoints we used in our paper, i.e. ./checkpoints/{STD,CSG}.pt.

  • MIS (mutual information score) between classes and channels

    python mutualinfo/mutualinfo.py

  • Localization performance of GradMap (gradient map)

    python gradmap/gradmap.py

  • Localization performance of ActivMap (activaiton map) and CAM (classification activation map)

    python activmap/activmap.py