基于SHAP框架的MobileNetv2模型事后可解释性分析
发布时间
阅读量：
阅读量
基于SHAP框架的MobileNetv2模型事后可解释性分析
-
0. MobileNetv2模型
-
-
1.1 预测单张数据
-
- 1.1.1 加载包
- 1.1.2 加载torchvision的预训练模型
-
1.2 批量预测数据
-
- 1.2.1 加载包
- 1.2.2 加载模型并训练
-
-
2. 训练与可解释性分析
-
- 2.1 单张预测结果
- 2.2 批量预测结果
0. MobileNetv2模型
MobileNetV2是对MobileNetV1的一种优化改进，并被设计为一种轻量级神经网络模型。该模型继承了V1版本的深度可分离卷积结构，并在此基础上引入了线性瓶颈（Linear Bottleneck）和倒残差（Inverted Residual）两种关键组件以提升性能。在整体网络架构中采用了固定扩展率设计，在实验研究中发现，在5到10之间的扩展率设置能够带来几乎一致的理想性能曲线。A类网络在使用较低扩展率时表现出更好的性能表现，在整体网络架构中采用了固定扩展率设计，在实验研究中发现，在5到10之间的扩展率设置能够带来几乎一致的理想性能曲线。A类网络在使用较低扩展率时表现出更好的性能表现
MobileNetV2主要采用一个扩张因子为6的设计策略来处理输入张量的尺寸变化。具体而言，在经过该网络架构时，在经过瓶颈层（Bottleneck layer）之前的特征图通常会先通过一个中间扩张层（Intermediate expansion layer）进行放大处理。例如，在一个使用了64个输入通道并最终生成128个输出通道的瓶颈层中，中间扩张层将具有64 \times 6 = 384个输出通道的数量。
1.1 预测单张数据
1.1.1 加载包
import numpy as np
import torchvision
import torch
import torch.nn as nn
import shap
from PIL import Image
import json
AI助手1.1.2 加载torchvision的预训练模型
###
'''
预测单张数据
'''
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = torchvision.models.mobilenet_v2(pretrained=True, progress=False)
model.to(device)
model.eval()
X, y = shap.datasets.imagenet50()
# Prepare data transformation pipeline
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
topk = 4
batch_size = 50
n_evals = 10000
############################################################################################################
url = "https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json"
with open(shap.datasets.cache(url)) as file:
class_names = [v[1] for v in json.load(file).values()]
##########################################################################################################
def nhwc_to_nchw(x: torch.Tensor) -> torch.Tensor:
if x.dim() == 4:
x = x if x.shape[1] == 3 else x.permute(0, 3, 1, 2)
elif x.dim() == 3:
x = x if x.shape[0] == 3 else x.permute(2, 0, 1)
return x
def nchw_to_nhwc(x: torch.Tensor) -> torch.Tensor:
if x.dim() == 4:
x = x if x.shape[3] == 3 else x.permute(0, 2, 3, 1)
elif x.dim() == 3:
x = x if x.shape[2] == 3 else x.permute(1, 2, 0)
return x
transform= [
torchvision.transforms.Lambda(nhwc_to_nchw),
torchvision.transforms.Lambda(lambda x: x*(1/255)),
torchvision.transforms.Normalize(mean=mean, std=std),
torchvision.transforms.Lambda(nchw_to_nhwc),
]
inv_transform= [
torchvision.transforms.Lambda(nhwc_to_nchw),
torchvision.transforms.Normalize(
mean = (-1 * np.array(mean) / np.array(std)).tolist(),
std = (1 / np.array(std)).tolist()
),
torchvision.transforms.Lambda(nchw_to_nhwc),
]
transform = torchvision.transforms.Compose(transform)
inv_transform = torchvision.transforms.Compose(inv_transform)
def predict(img: np.ndarray) -> torch.Tensor:
img = nhwc_to_nchw(torch.Tensor(img))
img = img.to(device)
output = model(img)
return output
###########################################################################################################
# Check that transformations work correctly
Xtr = transform(torch.Tensor(X))
out = predict(Xtr[1:3])
classes = torch.argmax(out, axis=1).cpu().numpy()
print(f'Classes: {classes}: {np.array(class_names)[classes]}')
########################################################################################################
# define a masker that is used to mask out partitions of the input image.
masker_blur = shap.maskers.Image("blur(128,128)", Xtr[0].shape)
# create an explainer with model and image masker
explainer = shap.Explainer(predict, masker_blur, output_names=class_names)
# feed only one image
# here we explain two images using 100 evaluations of the underlying model to estimate the SHAP values
shap_values = explainer(Xtr[1:4], max_evals=n_evals, batch_size=batch_size,
outputs=shap.Explanation.argsort.flip[:topk])
shap_values.data = inv_transform(shap_values.data).cpu().numpy()[0]
shap_values.values = [val for val in np.moveaxis(shap_values.values[0],-1, 0)]
shap.image_plot(shap_values=shap_values.values,
pixel_values=shap_values.data,
labels=shap_values.output_names,
true_labels=[class_names[132]])
AI助手1.2 批量预测数据
1.2.1 加载包
import numpy as np
import torchvision
import torch
import torch.nn as nn
import shap
from PIL import Image
import json
AI助手1.2.2 加载模型并训练
'''
批量化预测Imagenet数据集
'''
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = torchvision.models.mobilenet_v2(pretrained=True, progress=False)
model.to(device)
model.eval()
X, y = shap.datasets.imagenet50()
# Prepare data transformation pipeline
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
topk = 4
batch_size = 50
n_evals = 10000
############################################################################################################
url = "https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json"
with open(shap.datasets.cache(url)) as file:
class_names = [v[1] for v in json.load(file).values()]
##########################################################################################################
def nhwc_to_nchw(x: torch.Tensor) -> torch.Tensor:
if x.dim() == 4:
x = x if x.shape[1] == 3 else x.permute(0, 3, 1, 2)
elif x.dim() == 3:
x = x if x.shape[0] == 3 else x.permute(2, 0, 1)
return x
def nchw_to_nhwc(x: torch.Tensor) -> torch.Tensor:
if x.dim() == 4:
x = x if x.shape[3] == 3 else x.permute(0, 2, 3, 1)
elif x.dim() == 3:
x = x if x.shape[2] == 3 else x.permute(1, 2, 0)
return x
transform= [
torchvision.transforms.Lambda(nhwc_to_nchw),
torchvision.transforms.Lambda(lambda x: x*(1/255)),
torchvision.transforms.Normalize(mean=mean, std=std),
torchvision.transforms.Lambda(nchw_to_nhwc),
]
inv_transform= [
torchvision.transforms.Lambda(nhwc_to_nchw),
torchvision.transforms.Normalize(
mean = (-1 * np.array(mean) / np.array(std)).tolist(),
std = (1 / np.array(std)).tolist()
),
torchvision.transforms.Lambda(nchw_to_nhwc),
]
transform = torchvision.transforms.Compose(transform)
inv_transform = torchvision.transforms.Compose(inv_transform)
def predict(img: np.ndarray) -> torch.Tensor:
img = nhwc_to_nchw(torch.Tensor(img))
img = img.to(device)
output = model(img)
return output
###########################################################################################################
# Check that transformations work correctly
Xtr = transform(torch.Tensor(X))
out = predict(Xtr[1:5])
classes = torch.argmax(out, axis=1).cpu().numpy()
print(f'Classes: {classes}: {np.array(class_names)[classes]}')
########################################################################################################
# define a masker that is used to mask out partitions of the input image.
masker_blur = shap.maskers.Image("blur(128,128)", Xtr[0].shape)
# create an explainer with model and image masker
explainer = shap.Explainer(predict, masker_blur, output_names=class_names)
# feed only one image
# here we explain two images using 100 evaluations of the underlying model to estimate the SHAP values
shap_values = explainer(Xtr[1:5], max_evals=n_evals, batch_size=batch_size,
outputs=shap.Explanation.argsort.flip[:topk])
shap_values.data = inv_transform(shap_values.data).cpu().numpy()
shap_values.values = [val for val in np.moveaxis(shap_values.values,-1, 0)]
shap.image_plot(shap_values=shap_values.values,
pixel_values=shap_values.data,
labels=shap_values.output_names)
AI助手2. 训练与可解释性分析
2.1 单张预测结果
The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead.
Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=MobileNet_V2_Weights.IMAGENET1K_V1`. You can also use `weights=MobileNet_V2_Weights.DEFAULT` to get the most up-to-date weights.
Classes: [132 814]: ['American_egret' 'speedboat']
Partition explainer: 33%|███▎ | 1/3 [00:00<?, ?it/s]
0%| | 0/9998 [00:00<?, ?it/s]
20%|██ | 2042/9998 [00:00<00:00, 15163.07it/s]
36%|███▌ | 3592/9998 [00:04<00:09, 693.75it/s]
...
99%|█████████▉| 9942/9998 [00:24<00:00, 295.75it/s]
100%|█████████▉| 9992/9998 [00:25<00:00, 295.50it/s]
10042it [00:25, 295.84it/s]
Partition explainer: 4it [01:32, 30.73s/it]
Process finished with exit code 0
AI助手
2.2 批量预测结果
The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead.
Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=MobileNet_V2_Weights.IMAGENET1K_V1`. You can also use `weights=MobileNet_V2_Weights.DEFAULT` to get the most up-to-date weights.
Classes: [132 814 746 42]: ['American_egret' 'speedboat' 'puck' 'agama']
Partition explainer: 25%|██▌ | 1/4 [00:00<?, ?it/s]
0%| | 0/9998 [00:00<?, ?it/s]
19%|█▉ | 1942/9998 [00:00<00:00, 13616.69it/s]
33%|███▎ | 3342/9998 [00:03<00:09, 708.17it/s]
39%|███▉ | 3942/9998 [00:05<00:10, 576.55it/s]
43%|████▎ | 4292/9998 [00:06<00:10, 521.64it/s]
45%|████▌ | 4542/9998 [00:07<00:11, 487.35it/s]
47%|████▋ | 4742/9998 [00:07<00:11, 460.08it/s]
49%|████▉ | 4892/9998 [00:08<00:11, 440.01it/s]
...
98%|█████████▊| 9842/9998 [00:24<00:00, 293.41it/s]
99%|█████████▉| 9892/9998 [00:25<00:00, 292.31it/s]
99%|█████████▉| 9942/9998 [00:25<00:00, 292.30it/s]
100%|█████████▉| 9992/9998 [00:25<00:00, 292.03it/s]
10042it [00:25, 291.35it/s]
Partition explainer: 5it [02:04, 31.11s/it]
Process finished with exit code 0
AI助手
全部评论 (0)
还没有任何评论哟~