Skip to content

This repo covers the implementation of the following ICSE 2020 paper, "DISSECTOR: Input Validation for Deep Learning Applications by Crossing-layer Dissection " (DISSECTOR).

License

Notifications You must be signed in to change notification settings

ParagonLight/dissector

Repository files navigation

Dissector

Understanding behaviors of convolutional neural networks on image classification.

This repo covers the implementation of the following ICSE 2020 paper:

"DISSECTOR: Input Validation for Deep Learning Applications by Crossing-layer Dissection ".

Supported Models and Datasets

ImageNet:

  • ResNet-101
  • ResNet-50
  • VGG-16

CIFAR-100:

  • ResNeXt-29, 8x64
  • VGG-16
  • DenseNet-BC (L=100, k=12)

CIFAR-10:

  • WRN-28-10 (drop 0.3)
  • VGG-16
  • DenseNet-BC (L=100, k=12)

MNIST:

  • LeNet4
  • LeNet5
  • DNN2

**ToDo: support for submodel training.

Installation

  • Install PyTorch and TorchVision (pytorch.org)

    pip install torch torchvision
    
  • Install requirements

    pip install lmdb msgpack progress pillow sklearn
    
  • Prepare pretrained model and dataset

    ImageNet validation dataset, all supported models and corresponding submodels could be found through this link: https://1drv.ms/u/s!Anr26WqGCJOLsSICmSnSpZgvJM0K

Dissector Example for ResNet101 on ImageNet

Fetch the example data folder for ImageNet dataset

https://1drv.ms/u/s!Anr26WqGCJOLsSICmSnSpZgvJM0K

ILSVRC-val.lmdb is ImageNet validation set. You should change the dataset path in utils.py.

imagenet_val_path = YOURPATH

imagenet_pub is the root folder of the target. Pretrained submodels and layer info are all in imagenet_pub/models/resnet101.

tensor_pub is the root folder for outputs of dissector.

How to use

Suppose we have ImageNet dataset and pretrained ResNet101 model and corresponding pretrained 6 submodels.

  1. Create the root folder, such as YOURPROOT.

  2. Create several folders in YOURROOT folder.

  • YOURROOT/imagenet: root folder of imagenet dataset.

  • YOURROOT/imagenet/data: root folder of imagenet dataset files.

  • YOURROOT/imagenet/models/resnet101: root folder of ResNet101 sub models for imagenet dataset.

  • YOURROOT/imagenet/tensor_pub: root folder for anatomy outputs.

    • YOURROOT/imagenet/tensor_pub/res_layer1: folder for output of submodel res_layer1.
    • YOURROOT/imagenet/tensor_pub/res_layer2: folder for output of submodel res_layer2.
    • YOURROOT/imagenet/tensor_pub/res_block8: folder for output of submodel res_block8.
    • YOURROOT/imagenet/tensor_pub/res_block16: folder for output of submodel res_block16.
    • YOURROOT/imagenet/tensor_pub/res_layer3: folder for output of submodel res_layer3.
    • YOURROOT/imagenet/tensor_pub/res_layer4: folder for output of submodel res_layer4.
    • YOURROOT/imagenet/tensor_pub/out: folder for output of ResNet101.
  1. Put pretrained submodels model in folder data/imagenet/models/resnet101.

  2. Create file layer_info to write layers' info, which the layers are used for anatomy.

    For each row, write layer_name,layer's output size

  3. Run anatomy to produce results from each submodel for all instances.

    sh imagenet.sh
    
  4. Run merge_raw_layer_outputs.py to merge results from all layers.

    sh profile.sh
    

    this is for running imagenet using Dissector-linear as an example. Use --help to see arguments.

What to expect

For our example Imagenet+resnet101, AUC results are as follows:

Env Dissector-linear Dissector-log Dissector-exp
[Python 2.7.15, Pytorch 0.4.1] (our ICSE'20 paper setting) 0.8250 0.8223 0.8562
[Python 3.6.9, Pytorch 1.4.0] 0.8212 0.8237 0.8547

Note: AUC results may vary due to different versions of Pytorch and Python. According to our test on two different servers, the impact to effectiveness of Dissector is limited.

Citation

If you find this repo useful for your research, please consider citing the paper.

@inproceedings{Wang2019Dissector,
  title={Dissector: Input Validation for Deep Learning Applications by Crossing-layer Dissection},
  author={Huiyan Wang and Jingwei Xu and Chang Xu and Xiaoxing Ma and Jian Lu},
  booktitle={The 42th International Conference on Software Engineering},
  year={2020}
}

For any questions, please contact Huiyan Wang (cocowhy1013@gmail.com) and Jingwei Xu (jingweix@nju.edu.cn).

About

This repo covers the implementation of the following ICSE 2020 paper, "DISSECTOR: Input Validation for Deep Learning Applications by Crossing-layer Dissection " (DISSECTOR).

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •