Skip to content

Latest commit

 

History

History
83 lines (68 loc) · 3.08 KB

classification_training.md

File metadata and controls

83 lines (68 loc) · 3.08 KB
Raw

Training a Classifier using ImageNet-1K

Installation of iVMCL-Release is completed successfully.

The code has been tested under Ubuntu 16.04 LTS and 18.04 LTS. It also should work in other OS for which MMCV and MMDetection support.

Dataset Preparation

ImageNet-1k

  • Download the ImageNet dataset to YOUR_IMAGENET_PATH and unzip.

    • Move validation images to labeled subfolders
    • Note: ILSVRC2015 is used in our training. For the ImageNet-1K calssification task, it is the same as ILSVRC2012.

  • Create a data subfolder under the cloned iVMCL-Release/mmdetection and a symbolic link to the ImageNet dataset

    cd iVMCL-Release/mmdetection
mkdir data
cd data
ln -s YOUR_IMAGENET_PATH ./

    E.g., the data root (relative) directory will be: data_root=data/ILSVRC2015/Data/CLS-LOC.

    The directory structure will look like

    iVMCL-Release
├── mmcv
├── mmdetection
    ├── mmdet
    ├── tools
    ├── configs
    ├── tools_ivmcl
    ├── configs_ivmcl
    ├── scripts_ivmcl
    ├── data
    │   ├── ILSVRC2015
    │   │   ├── Annotations
    │   │   ├── ImageSets
    │   │   ├── Data
    │   │       ├── CLS-LOC
    │   │           ├── train
    │   │           ├── test
    |   |           ├── val
    |   |           ├── val_orig

ImageNet-1k Reassessed

  • Paper: `Are we done with ImageNet?
  • Download the real.json file at the Reassessed ImageNet repo. 
    cd iVMCL-Release/mmdetection/data/ILSVRC2015/Data/CLS-LOC/reassessed-imagenet
wget https://raw.githubusercontent.com/google-research/reassessed-imagenet/master/real.json

Some ImageNet-X validation dataset

  • Please run the imagenet-v2.py, imagenet-sketch.py and imagenet-adv.py in the iVMCL-Release/mmdetection/tools_ivmcl folder. 
    cd iVMCL-Release/mmdetection/tools_ivmcl
python imagenet-v2.py
python imagenet-sketch.py
python imagenet-adv.py

Training a model from scratch

  • Select a configuration file at the cloned iVMCL-Release/mmdetection/configs_ivmcl, or create a new one accordingly.

    • E.g., consider aognet_12m_an_imagenet.py

  • Check data_root in a configuration file to make sure it points to the correct directory

  • Change the training hyperparameters if needed, e.g., batch_size

  • Run the script to train

    cd iVMCL-Release/mmdetection
chmod +x ./scripts_ivmcl/*.sh

    change the GPU configuration in the train_supervised_dist.sh accordingly based on your hardware environment.

    ./scripts_ivmcl/train_supervised_dist.sh configs_ivmcl/aognet_12m_an_imagenet.py