# cifar10_test_featuremaps

**Repository Path**: jackeywang7/cifar10_test_featuremaps

## Basic Information

- **Project Name**: cifar10_test_featuremaps
- **Description**: cifar10 testdata featuremaps
- **Primary Language**: Python
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-04-26
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Introduction about featuremaps extracted from cifar10

>we extract the featuremaps from a trained CifarNet in cifar-10 dataset.
>
>The CifarNet architecutre is refer from 
>
>[1]. Ma, B. Li, Y. Wang, S. M. Erfani, S. Wijewickrema, M. E. Houle, G. Schoenebeck, D. Song, and J. Bailey. Characterizing adversarial subspaces using local intrinsic dimensionality. arXiv preprint arXiv:1801.02613, 2018. 2, 5, 6, 7, 11, 12
>
>[2] Krishna Kanth Nakka,Mathieu Salzmann Interpretable BoW Network from Adversarial Example Detection. arXiv:1901.02229v1

You can download the featuremaps and model from :https://gitee.com/jackeywang7/cifar10_test_featuremaps

**Enviroment:** pytorch1.0+,HDF5,Numpy

## 1.0 Data

**We train a CifarNet model in cifar-10 and the test accuracy can archieve 87.29%(the best result in that network architecture)**

Cifar-10 dataset:http://www.cs.toronto.edu/~kriz/cifar.html?usg=alkjrhjqbhw2llxlo8emqns-tbk0at96jq

The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images.

![](https://jackeyimage.oss-cn-beijing.aliyuncs.com/img20200426180644.png)



## 2.0 CNN network architecture

```
CifarNet(
  (features): Sequential(
    (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))  #[3,32,32]
    (1): ReLU(inplace=True)
    (2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): ReLU(inplace=True)
    (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    #feature map[32,16,16]
    (5): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))#[64,16,16]
    (6): ReLU(inplace=True)
    (7): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (8): ReLU(inplace=True)
    (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)#[64,8,8]
    (10): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))  
    #feature map size[128,8,8]
    (13): ReLU(inplace=True) 
    #extract featuremap over here [128,8,8]
    (14): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    #feture map size[128,4,4]
  )
  (classifier): Sequential(
    (0): Linear(in_features=2048, out_features=1024, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=1024, out_features=512, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=512, out_features=10, bias=True)
  )
)
```

## 3.0 Feature maps

**we extract the feature maps from the target layer "13"(see above architecture)**

> from train_data:
>
> Class:airplane featuremap.size:(1000, 128, 8, 8) label:0 objNum:1000
> Class:automobile featuremap.size:(1000, 128, 8, 8) label:1 objNum:1000
> Class:bird featuremap.size:(1000, 128, 8, 8) label:2 objNum:1000
> Class:cat featuremap.size:(1000, 128, 8, 8) label:3 objNum:1000
> Class:deer featuremap.size:(1000, 128, 8, 8) label:4 objNum:1000
> Class:dog featuremap.size:(1000, 128, 8, 8) label:5 objNum:1000
> Class:frog featuremap.size:(1000, 128, 8, 8) label:6 objNum:1000
> Class:horse featuremap.size:(1000, 128, 8, 8) label:7 objNum:1000
> Class:ship featuremap.size:(1000, 128, 8, 8) label:8 objNum:1000
> Class:truck featuremap.size:(1000, 128, 8, 8) label:9 objNum:1000
>
> from test_data:
>
> Class:airplane featuremap.size:(1000, 128, 8, 8) label:0 objNum:1000
> Class:automobile featuremap.size:(1000, 128, 8, 8) label:1 objNum:1000
> Class:bird featuremap.size:(1000, 128, 8, 8) label:2 objNum:1000
> Class:cat featuremap.size:(1000, 128, 8, 8) label:3 objNum:1000
> Class:deer featuremap.size:(1000, 128, 8, 8) label:4 objNum:1000
> Class:dog featuremap.size:(1000, 128, 8, 8) label:5 objNum:1000
> Class:frog featuremap.size:(1000, 128, 8, 8) label:6 objNum:1000
> Class:horse featuremap.size:(1000, 128, 8, 8) label:7 objNum:1000
> Class:ship featuremap.size:(1000, 128, 8, 8) label:8 objNum:1000
> Class:truck featuremap.size:(1000, 128, 8, 8) label:9 objNum:1000

You can read the hdf5 from following code:

```python
import h5py
import numpy

classes = ('airplane', 'automobile', 'bird', 'cat',

           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

#You should modify the hdf5 save_path

save_path = "9.truck_testset_data.h5"

def read_rougCNN_hdf5(save_path, target_layer=["13"]):

    #only extact the target layer "13" from the pretrain_model

    f = h5py.File(save_path, 'r')

    layerNum = len(target_layer)

    label = f['label'][:]

    for idx, layer in enumerate(target_layer):

        curgroup = f[layer]

        curlayer_featuremap = curgroup['layerfeaturemaps'][:]

    f.close()

    return curlayer_featuremap,  label

if __name__=="__main__":
    #test
    featuremap,label = read_rougCNN_hdf5(save_path)
    print(featuremap.shape,classes[label[0]])
```