[CAM]HCAM original

CAM

Author

SEOYEON CHOI

Published

September 19, 2023

https://seoyeonc.github.io/chch/cnn/feature%20extraction/big%20data%20analysis/2022/01/11/bd_9주차.html

https://seoyeonc.github.io/chch/cam/2022/01/10/bd-8주차_1.html

CNN으로 이미지 분류를 할 때 마지막 단의 출력값이 클수록 softmax를 거친 뒤 1에 가까워 진다면, 입력 이미지의 label에 해당하는 채널의 마지막 conv layer의 출력이 크게 하는 클래스에 크게 반응했다는 것이 됌..!

import

import torch 
from fastai.vision.all import *
import cv2
import numpy as np
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import cv2
import numpy as np
import matplotlib.pyplot as plt
from PIL import ImageDraw
from PIL import ImageFont
from PIL import ImageFile
from PIL import Image
ImageFile.LOAD_TRUNCATED_IMAGES = True
from torchvision.utils import save_image
import os

import rpy2
import rpy2.robjects as ro 
from rpy2.robjects.vectors import FloatVector 
from rpy2.robjects.packages import importr

def label_func(f):
    if f[0].isupper():
        return 'cat' 
    else: 
        return 'dog'

학습

path=Path('original_pet')   #랜덤박스넣은사진

files=get_image_files(path)

dls=ImageDataLoaders.from_name_func(path,files,label_func,item_tfms=Resize(512))

lrnr=cnn_learner(dls,resnet34,metrics=error_rate)
lrnr.fine_tune(1)

/home/csy/anaconda3/envs/temp_csy/lib/python3.8/site-packages/fastai/vision/learner.py:288: UserWarning: `cnn_learner` has been renamed to `vision_learner` -- please update your code
  warn("`cnn_learner` has been renamed to `vision_learner` -- please update your code")
/home/csy/anaconda3/envs/temp_csy/lib/python3.8/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead.
  warnings.warn(
/home/csy/anaconda3/envs/temp_csy/lib/python3.8/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=ResNet34_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet34_Weights.DEFAULT` to get the most up-to-date weights.
  warnings.warn(msg)

epoch	train_loss	valid_loss	error_rate	time
0	0.183123	0.020765	0.008796	40:02

0.00% [0/1 00:00<?]

epoch	train_loss	valid_loss	error_rate	time

16.30% [15/92 07:49<40:11 0.0305]

net1=lrnr.model[0]
net2=lrnr.model[1]

net2 = torch.nn.Sequential(
    torch.nn.AdaptiveAvgPool2d(output_size=1), 
    torch.nn.Flatten(),
    torch.nn.Linear(512,out_features=2,bias=False))

net=torch.nn.Sequential(net1,net2)

lrnr2=Learner(dls,net,metrics=accuracy)

lrnr2.fine_tune(5)

0.00% [0/1 00:00<?]

epoch	train_loss	valid_loss	accuracy	time

14.13% [13/92 07:12<43:45 0.4181]

interp = ClassificationInterpretation.from_learner(lrnr2)
interp.plot_confusion_matrix()

Accuracy

cat_acc_s = [] #고양이를 고양이라고 잘 맞춤
dog_acc_s = [] #강아지를 고양이라고 맞춤
cat_acc_f = [] #강아지를 강아지라고 잘 맞춤
dog_acc_f = [] #고양이를 강아지라고 맞춤

for i in range(len(path.ls())) :
    x, = first(dls.test_dl([PILImage.create(get_image_files(path)[i])]))
    camimg = torch.einsum('ij,jkl -> ikl', net2[2].weight, net1(x).squeeze())
    a,b = net(x).tolist()[0]
    catprob, dogprob = np.exp(a)/ (np.exp(a)+np.exp(b)) ,  np.exp(b)/ (np.exp(a)+np.exp(b)) 
    if catprob>dogprob: 
        if label_func(str(list(path.ls())[i]).split('/')[-1]) == 'cat' :
            cat_acc_s.append(catprob.round(5))
        else : 
            cat_acc_f.append(catprob.round(5))
    else:
        if label_func(str(list(path.ls())[i]).split('/')[-1]) == 'dog' :
            dog_acc_s.append(dogprob.round(5))
        else : 
            dog_acc_f.append(dogprob.round(5))

print(len(cat_acc_s))
print(len(cat_acc_f))
print(len(dog_acc_s))
print(len(dog_acc_f))

print(sum(cat_acc_s)/len(cat_acc_s) * 100)
print(sum(cat_acc_f)/len(cat_acc_f) * 100)

print(sum(dog_acc_s)/len(dog_acc_s) * 100)
print(sum(dog_acc_f)/len(dog_acc_f) * 100)

99.35801175657008
99.40861033369225
99.6943825701619
99.6278415529907

Visualization

# # 서연 수정 code
# fig, (ax1,ax2) = plt.subplots(1,2) 
# # 
# dls_r.train.decode((x,))[0].squeeze().show(ax=ax1)
# ax1.imshow(camimg[0].to("cpu").detach(),alpha=0.7,extent=(0,511,511,0),interpolation='spline36',cmap='magma')
# #
# dls_r.train.decode((x,))[0].squeeze().show(ax=ax2)
# ax2.imshow(camimg[1].to("cpu").detach(),alpha=0.7,extent=(0,511,511,0),interpolation='spline36',cmap='magma')
# fig.set_figwidth(8)            
# fig.set_figheight(8)
# fig.tight_layout()

# fig, ax = plt.subplots(5,5) 
# k=0 
# for i in range(5):
#     for j in range(5): 
#         x, = first(dls_r.test_dl([PILImage.create(get_image_files(path_r)[k])]))
#         camimg = torch.einsum('ij,jkl -> ikl', net_2[2].weight, net_1(x).squeeze())
#         a,b = net_r(x).tolist()[0]
#         catprob, dogprob = np.exp(a)/ (np.exp(a)+np.exp(b)) ,  np.exp(b)/ (np.exp(a)+np.exp(b)) 
#         if catprob>dogprob: 
#             dls_r.train.decode((x,))[0].squeeze().show(ax=ax[i][j])
#             ax[i][j].imshow(camimg[0].to("cpu").detach(),alpha=0.7,extent=(0,512,512,0),interpolation='bilinear',cmap='bone')
#             ax[i][j].set_title("cat(%s)" % catprob.round(5))
#         else: 
#             dls_r.train.decode((x,))[0].squeeze().show(ax=ax[i][j])
#             ax[i][j].imshow(camimg[1].to("cpu").detach(),alpha=0.7,extent=(0,512,512,0),interpolation='bilinear',cmap='bone')
#             ax[i][j].set_title("dog(%s)" % dogprob.round(5))
#         k=k+1 
# fig.set_figwidth(16)            
# fig.set_figheight(16)
# fig.tight_layout()

ebayes X

fig, ax = plt.subplots(5,5) 
k=0 
for i in range(5):
    for j in range(5): 
        x, = first(dls.test_dl([PILImage.create(get_image_files(path)[k])]))
        camimg = torch.einsum('ij,jkl -> ikl', net2[2].weight, net1(x).squeeze())
        a,b = net(x).tolist()[0]
        catprob, dogprob = np.exp(a)/ (np.exp(a)+np.exp(b)) ,  np.exp(b)/ (np.exp(a)+np.exp(b))
        if catprob>dogprob: 
            test=camimg[0]-torch.min(camimg[0])
            A1=torch.exp(-0.1*test)
            X1=np.array(A1.to("cpu").detach(),dtype=np.float32)
            Y1=torch.Tensor(cv2.resize(X1,(512,512),interpolation=cv2.INTER_LINEAR))
            x1=x.squeeze().to('cpu')*Y1-torch.min(x.squeeze().to('cpu'))*Y1
            (x1*0.25).squeeze().show(ax=ax[i][j])
            ax[i][j].set_title("cat(%s)" % catprob.round(5))
        else: 
            test=camimg[1]-torch.min(camimg[1])
            A1=torch.exp(-0.1*test)
            X1=np.array(A1.to("cpu").detach(),dtype=np.float32)
            Y1=torch.Tensor(cv2.resize(X1,(512,512),interpolation=cv2.INTER_LINEAR))
            x1=x.squeeze().to('cpu')*Y1-torch.min(x.squeeze().to('cpu'))*Y1
            (x1*0.25).squeeze().show(ax=ax[i][j])
            ax[i][j].set_title("dog(%s)" % dogprob.round(5))
        k=k+1 
fig.set_figwidth(16)            
fig.set_figheight(16)
fig.tight_layout()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

Step by step

CAT

x, = first(dls.test_dl([PILImage.create(get_image_files(path)[2])]))

camimg = torch.einsum('ij,jkl -> ikl', net2[2].weight, net1(x).squeeze())

fig, (ax1,ax2,ax3) = plt.subplots(1,3) 
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.set_title("Input image")
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax2)
ax2.imshow((camimg[0]).to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("CAT PART")
#
dls.train.decode((x,))[0].squeeze().show(ax=ax3)
ax3.imshow((camimg[1]).to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax3.set_title("DOG PART")
#
fig.set_figwidth(12)            
fig.set_figheight(12)
fig.tight_layout()

판단 근거가 강할 수록 파란색 -> 보라색

a,b = net(x).tolist()[0]

np.exp(a)/ (np.exp(a)+np.exp(b)) ,  np.exp(b)/ (np.exp(a)+np.exp(b))

(0.9999560817987819, 4.3918201218046276e-05)

mode 1

test=camimg[0]-torch.min(camimg[0])
A1=torch.exp(-0.01*(test))
A2 = 1 - A1

fig, (ax1,ax2) = plt.subplots(1,2) 
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.imshow(A2.data.to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("MODE1 WEIGHTT")
#
dls.train.decode((x,))[0].squeeze().show(ax=ax2)
ax2.imshow(A1.data.to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("MODE1 RES WEIGHT")
#
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

# mode 1 res
X1=np.array(A1.to("cpu").detach(),dtype=np.float32)
Y1=torch.Tensor(cv2.resize(X1,(512,512),interpolation=cv2.INTER_LINEAR))
x1=x.squeeze().to('cpu')*Y1-torch.min(x.squeeze().to('cpu'))*Y1

# mode 1
X12=np.array(A2.to("cpu").detach(),dtype=np.float32)
Y12=torch.Tensor(cv2.resize(X12,(512,512),interpolation=cv2.INTER_LINEAR))
x12=x.squeeze().to('cpu')*Y12-torch.min(x.squeeze().to('cpu'))*Y12

- 1st CAM 분리

fig, (ax1) = plt.subplots(1,1) 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.set_title("ORIGINAL")
fig.set_figwidth(4)            
fig.set_figheight(4)
fig.tight_layout()
#
fig, (ax1, ax2) = plt.subplots(1,2) 
(x12*0.3).squeeze().show(ax=ax1)  #MODE1
(x1*0.3).squeeze().show(ax=ax2)  #MODE1_res
ax1.set_title("MODE1")
ax2.set_title("MODE1 RES")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

x1 = x1.reshape(1,3,512,512)

net1.to('cpu')
net2.to('cpu')

Sequential(
  (0): AdaptiveAvgPool2d(output_size=1)
  (1): Flatten(start_dim=1, end_dim=-1)
  (2): Linear(in_features=512, out_features=2, bias=False)
)

camimg1 = torch.einsum('ij,jkl -> ikl', net2[2].weight, net1(x1).squeeze())

a1,b1 = net(x1).tolist()[0]

np.exp(a1)/ (np.exp(a1)+np.exp(b1)) ,  np.exp(b1)/ (np.exp(a1)+np.exp(b1))

(0.9986349047084874, 0.0013650952915126677)

- mode1 res

fig, (ax1,ax2) = plt.subplots(1,2) 
# 
(x1*0.3).squeeze().show(ax=ax1)
ax1.imshow(camimg1[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("CAT PART")
#
(x1*0.3).squeeze().show(ax=ax2)
ax2.imshow(camimg1[1].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("DOG PART")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

- 첫번째 CAM 결과와 비교

fig, (ax1,ax2) = plt.subplots(1,2) 
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.imshow(camimg[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("1ST CAM")
#
dls.train.decode((x,))[0].squeeze().show(ax=ax2)
ax2.imshow(camimg1[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("2ND CAM")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

- 2nd CAM 분리

test1=camimg1[0]-torch.min(camimg1[0])
A3 = torch.exp(-0.015*(test1))
A4 = 1 - A3

fig, (ax1,ax2) = plt.subplots(1,2) 
# 
x1.squeeze().show(ax=ax2)
dls.train.decode((x1,))[0].squeeze().show(ax=ax1)
ax1.imshow(A3.data.to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("MODE2 RES WEIGHT")
#
x1.squeeze().show(ax=ax2)
dls.train.decode((x1,))[0].squeeze().show(ax=ax2)
ax2.imshow(A4.data.to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("MODE2 WEIGHT")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

X2=np.array(A3.to("cpu").detach(),dtype=np.float32)
Y2=torch.Tensor(cv2.resize(X2,(512,512),interpolation=cv2.INTER_LINEAR))
x2=(x1)*Y2-torch.min((x1)*Y2)

X22=np.array(A4.to("cpu").detach(),dtype=np.float32)
Y22=torch.Tensor(cv2.resize(X22,(512,512),interpolation=cv2.INTER_LINEAR))
x22=(x1)*Y22-torch.min((x1)*Y22)

fig, (ax1) = plt.subplots(1,1) 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.set_title("ORIGINAL")
fig.set_figwidth(4)            
fig.set_figheight(4)
fig.tight_layout()
#
fig, (ax1, ax2) = plt.subplots(1,2) 
(x12*0.3).squeeze().show(ax=ax1)  #MODE1
(x1*0.2).squeeze().show(ax=ax2)  #MODE1_res
ax1.set_title("MODE1")
ax2.set_title("MODE1 RES")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()
#
fig, (ax1, ax2) = plt.subplots(1,2) 
(x22*0.5).squeeze().show(ax=ax1)  #MODE2
(x2*0.3).squeeze().show(ax=ax2)  #MODE2_res
ax1.set_title("MODE2")
ax2.set_title("MODE2 RES")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

x2 = x2.reshape(1,3,512,512)

net1.to('cpu')
net2.to('cpu')

Sequential(
  (0): AdaptiveAvgPool2d(output_size=1)
  (1): Flatten(start_dim=1, end_dim=-1)
  (2): Linear(in_features=512, out_features=2, bias=False)
)

camimg2 = torch.einsum('ij,jkl -> ikl', net2[2].weight, net1(x2).squeeze())

a2,b2 = net(x2).tolist()[0]
np.exp(a2)/(np.exp(a2)+np.exp(b2)), np.exp(b2)/(np.exp(a2)+np.exp(b2))

(0.19813533943993217, 0.8018646605600679)

- mode2 res 에 CAM 결과 올리기

fig, (ax1, ax2) = plt.subplots(1,2) 
#
x2.squeeze().show(ax=ax1)
ax1.imshow(camimg2[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("CAT PART")
#
x2.squeeze().show(ax=ax2)
ax2.imshow(camimg2[1].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("DOG PART")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

fig, (ax1,ax2,ax3) = plt.subplots(1,3) 
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.imshow(camimg[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("1ST CAM")
#
dls.train.decode((x,))[0].squeeze().show(ax=ax2)
ax2.imshow(camimg1[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("2ND CAM")
#
dls.train.decode((x,))[0].squeeze().show(ax=ax3)
ax3.imshow(camimg2[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax3.set_title("3RD CAM")
fig.set_figwidth(12)            
fig.set_figheight(12)
fig.tight_layout()

mode 3 만들기

test2=camimg2[0]-torch.min(camimg2[0])

A5 = torch.exp(-0.05*(test2))

A6 = 1 - A5

fig, (ax1, ax2) = plt.subplots(1,2) 
#
x2.squeeze().show(ax=ax1)
ax1.imshow(camimg2[0].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax1.set_title("CAT PART")
#
x2.squeeze().show(ax=ax2)
ax2.imshow(camimg2[1].to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("DOG PART")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

#mode 3 res
X3=np.array(A5.to("cpu").detach(),dtype=np.float32)
Y3=torch.Tensor(cv2.resize(X3,(512,512),interpolation=cv2.INTER_LINEAR))
x3=x2*Y3-torch.min(x2*Y3)
# mode 3
X32=np.array(A6.to("cpu").detach(),dtype=np.float32)
Y32=torch.Tensor(cv2.resize(X32,(512,512),interpolation=cv2.INTER_LINEAR))
x32=x2*Y32-torch.min(x2*Y32)

fig, (ax1) = plt.subplots(1,1) 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.set_title("ORIGINAL")
fig.set_figwidth(4)            
fig.set_figheight(4)
fig.tight_layout()
#
fig, (ax1, ax2) = plt.subplots(1,2) 
(x12*0.3).squeeze().show(ax=ax1)  #MODE1
(x1*0.2).squeeze().show(ax=ax2)  #MODE1_res
ax1.set_title("MODE1")
ax2.set_title("MODE1 RES")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()
#
fig, (ax1, ax2) = plt.subplots(1,2) 
(x22*0.5).squeeze().show(ax=ax1)  #MODE2
(x2*0.3).squeeze().show(ax=ax2)  #MODE2_res
ax1.set_title("MODE2")
ax2.set_title("MODE2 RES")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()
#
fig, (ax1, ax2) = plt.subplots(1,2) 
(x32*0.3).squeeze().show(ax=ax1)  #MODE3
(x3*0.2).squeeze().show(ax=ax2)  #MODE3_res
ax1.set_title("MODE3")
ax2.set_title("MODE3 RES")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

fig, (ax1) = plt.subplots(1,1) 
(x12*0.3).squeeze().show(ax=ax1)  #MODE1
ax1.set_title("MODE1")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

fig, (ax1) = plt.subplots(1,1) 
(x12*0.3 + x22*0.3).squeeze().show(ax=ax1)  #MODE1+MODE2
ax1.set_title("MODE1+MODE2")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

fig, (ax1) = plt.subplots(1,1) 
(x12*0.3 + x22*0.3 + x32*0.2).squeeze().show(ax=ax1)  #MODE1+MODE2+MODE3
ax1.set_title("MODE3+MODE2+MODE3")
fig.set_figwidth(8)            
fig.set_figheight(8)
fig.tight_layout()

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

DOG

x, = first(dls.test_dl([PILImage.create(get_image_files(path)[12])]))

camimg = torch.einsum('ij,jkl -> ikl', net2[2].weight, net1(x).squeeze())

fig, (ax1,ax2,ax3) = plt.subplots(1,3) 
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax1)
ax1.set_title("Input image")
# 
dls.train.decode((x,))[0].squeeze().show(ax=ax2)
ax2.imshow((camimg[0]).to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax2.set_title("CAT PART")
#
dls.train.decode((x,))[0].squeeze().show(ax=ax3)
ax3.imshow((camimg[1]).to("cpu").detach(),alpha=0.5,extent=(0,511,511,0),interpolation='bilinear',cmap='cool')
ax3.set_title("DOG PART")
#
fig.set_figwidth(12)            
fig.set_figheight(12)
fig.tight_layout()

판단 근거가 강할 수록 파란색 -> 보라색

a,b = net(x).tolist()[0]

np.exp(a)/ (np.exp(a)+np.exp(b)) ,  np.exp(b)/ (np.exp(a)+np.exp(b))

(1.7996574339365895e-06, 0.9999982003425661)