话有点多导包没问题的话可以跳过加载MNIST数据集加载经过训练的MNIST模型加载经过训练的卷积自动编码器模型可选初始化CEM解释程序以解释模型预测解释输入实例获得相关否定Pertinent

本文了如何使用AIX360的CEMBexplainer来获得对比解释的示例即对MNIST数据训练的模型所做预测的相关否定PN和相关肯定PP。

CEMBexplainer是对比解释方法的一种实现。

此案例使用经过训练的模型这些模型可从aix360/models/CEM/文件夹访问。

官方代码在https://github.com/Trusted-AI/AIX360/blob/master/examples/contrastive/CEM-MNIST.ipynb

pip

MNISTDataset经典一步一bug眼睛一睁一闭休眠升天修仙。

。

。

TensorFlow

tensorflow1.14.0看到这我真的高兴坏了之前不小心把python版本装高了没办法就是这么倒霉推倒重来官网怎么喜欢用那么老的版本为什么我的眼里常含泪水因为对知识爱得深沉。

。

。

重新创建个虚拟环境

python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6

https://blog.csdn.net/weixin_45735391/article/details/133197625

python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6python3.6

python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7

https://blog.csdn.net/weixin_45735391/article/details/133197625

python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7python3.7

pip

warnings.filterwarnings(ignore)好了导包这块终于结束了。

加载MNIST数据集

MNISTDataset()花的时间有亿点点久。

。

。

真的等不下去了。

。

。

脑子已经在修仙了。

。

。

class

os.path.join(os.path.dirname(os.path.abspath(__file__)),..,

[train-images-idx3-ubyte.gz,t10k-images-idx3-ubyte.gz,train-labels-idx1-ubyte.gz,t10k-labels-idx1-ubyte.gz]for

name

name)urllib.request.urlretrieve(http://yann.lecun.com/exdb/mnist/

name,

name)print(retrieved)train_data

60000)train_labels

train_labels[:VALIDATION_SIZE]self.train_data

:]self.train_labels

train_labels[VALIDATION_SIZE:]直接百度搜一下下载MNIST数据集找到别人分享的资源把这四个文件[train-images-idx3-ubyte.gz,t10k-images-idx3-ubyte.gz,train-labels-idx1-ubyte.gz,

t10k-labels-idx1-ubyte.gz]下载下来。

dirpath

rC:\Users\gxx\Desktop\jupter\aix360\MNIST

data

此notebook使用经过训练的MNIST模型。

此处提供了训练此模型的代码。

请注意该模型输出logits并且不使用softmax函数。

path

rC:\Users\gxx\Desktop\jupter\aix360\CEMdef

model_wt_file):#

filemodel.load_weights(model_wt_file)return

model#

load_model(os.path.join(model_path,

mnist.json),

mnist_model.summary()不出意外bug又来了。

。

。

在安装

这个notebook使用了一个经过训练的卷积自动编码器模型。

此处提供了训练此模型的代码。

load

load_model(os.path.join(model_path,

print

ae_model.summary()初始化CEM解释程序以解释模型预测

wrap

mymodel.predict_classes(np.expand_dims(input_image,

axis0)))

mymodel.predict(np.expand_dims(input_image,

axis0)))结果输出

尽管上面的图像被模型分类为数字3但是由于它与数字5具有相似性所以它也可以被分类为数字5。

我们现在使用AIX360的CEMBexplainer来计算相关的正面和负面解释这有助于我们理解为什么图像被模型分类为数字3而不是数字5。

获得相关否定Pertinent

explainer.explain_instance(np.expand_dims(input_image,

axis0),

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\aix360\algorithms\contrastive\CEM_aen.py:60:

The

instead.WARNING:tensorflow:From

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\aix360\algorithms\contrastive\CEM_aen.py:151:

The

instead.WARNING:tensorflow:From

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\aix360\algorithms\contrastive\CEM_aen.py:213:

The

tf.compat.v1.train.polynomial_decay

instead.WARNING:tensorflow:From

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\tensorflow\python\keras\optimizer_v2\learning_rate_schedule.py:409:

div

tensorflow.python.ops.math_ops)

deprecated

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\aix360\algorithms\contrastive\CEM_aen.py:216:

The

tf.train.GradientDescentOptimizer

deprecated.

tf.compat.v1.train.GradientDescentOptimizer

instead.WARNING:tensorflow:From

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\tensorflow\python\ops\math_grad.py:1250:

add_dispatch_support.locals.wrapper

(from

tensorflow.python.ops.array_ops)

deprecated

C:\Users\gxx\anaconda3\envs\tf-py37\lib\site-packages\aix360\algorithms\contrastive\CEM_aen.py:230:

The

tf.compat.v1.variables_initializer

instead.iter:0

max_nontarget_lab_score:14.4428iter:500

const:[10.]

max_nontarget_lab_score:14.4428iter:0

const:[100.]

max_nontarget_lab_score:29.0375iter:500

const:[100.]

max_nontarget_lab_score:28.5018iter:0

const:[55.]

max_nontarget_lab_score:27.1471iter:500

const:[55.]

max_nontarget_lab_score:19.4013iter:0

const:[32.5]

max_nontarget_lab_score:17.5513iter:500

const:[32.5]

max_nontarget_lab_score:14.4428iter:0

const:[21.25]

max_nontarget_lab_score:14.5784iter:500

const:[21.25]

max_nontarget_lab_score:14.4428iter:0

const:[26.875]

max_nontarget_lab_score:15.6829iter:500

const:[26.875]

max_nontarget_lab_score:14.4428iter:0

const:[24.0625]

max_nontarget_lab_score:14.8760iter:500

const:[24.0625]

max_nontarget_lab_score:14.4428iter:0

const:[25.46875]

max_nontarget_lab_score:15.3106iter:500

const:[25.46875]

max_nontarget_lab_score:14.4428iter:0

const:[24.765625]

max_nontarget_lab_score:15.1403iter:500

const:[24.765625]

max_nontarget_lab_score:14.4428print(info_pn)结果输出

[INFO]kappa:0.9,

explainer.explain_instance(np.expand_dims(input_image,

axis0),

max_nontarget_lab_score:1.0441iter:500

const:[10.]

max_nontarget_lab_score:1.0441iter:0

const:[100.]

max_nontarget_lab_score:1.1579iter:500

const:[100.]

max_nontarget_lab_score:5.3354iter:0

const:[55.]

max_nontarget_lab_score:1.0441iter:500

const:[55.]

max_nontarget_lab_score:1.0441iter:0

const:[77.5]

max_nontarget_lab_score:1.0441iter:500

const:[77.5]

max_nontarget_lab_score:1.0441iter:0

const:[88.75]

max_nontarget_lab_score:1.1082iter:500

const:[88.75]

max_nontarget_lab_score:8.5652iter:0

const:[83.125]

max_nontarget_lab_score:1.0918iter:500

const:[83.125]

max_nontarget_lab_score:5.3355iter:0

const:[80.3125]

max_nontarget_lab_score:1.0846iter:500

const:[80.3125]

max_nontarget_lab_score:2.9340iter:0

const:[78.90625]

max_nontarget_lab_score:1.0807iter:500

const:[78.90625]

max_nontarget_lab_score:4.7113iter:0

const:[78.203125]

max_nontarget_lab_score:1.0788iter:500

const:[78.203125]

max_nontarget_lab_score:5.0136print(info_pp)结果输出

[INFO]kappa:0.9,

str(mymodel.predict_classes(np.expand_dims(input_image,

axis0))[0])

str(mymodel.predict_classes(adv_pn)[0])

PP)

str(mymodel.predict_classes(delta_pp)[0])

cmapgray)

PP突出显示图像中存在的最小像素集以便将其分类为数字3。

注意原始图像和PP都被分类器分类为数字3。

PN在顶部突出显示一条小水平线该水平线的存在会将原始图像的分类改变为数字5因此应该不存在以便分类保持为数字3。