Open In Colab

Let’s install the txv package.#

[1]:

# !pip install txv

Import necessary libraries.#

[2]:

import sys
sys.path.append('..')
from txv.exp import *
from txv.utils import *
from txv.vit import *
import matplotlib.pyplot as plt

Model Initialization#

We will use vit-base-patch16-224 model in this tutorial. You can try this with other models in Available Models section as well. Note that this package works with models in Available Models section only. We initialize 2 models, one with lrp=True and other with lrp=False. LRP based explanation methods like LRP and Beyond Attention require relevance propogation which is enabled through lrp=True.#

[3]:

device = torch.device('cuda:0')
model1 = vit_base_patch16_224(lrp=True).to(device)
model2 = vit_base_patch16_224().to(device)

[4]:

# Read the image
image = read_image('../images/zebra-elephant.png').to(device)

[5]:

# Generate the output
output = model1(image)

[6]:

# Print the top classes
print_top_classes(output)

Top 10 classes:
        340 : zebra                                             value = 12.749   prob = 94.2%
        386 : African elephant, Loxodonta africana              value = 9.485    prob = 3.6%
        101 : tusker                                            value = 7.923    prob = 0.8%
        385 : Indian elephant, Elephas maximus                  value = 7.684    prob = 0.6%
        352 : impala, Aepyceros melampus                        value = 5.946    prob = 0.1%
        351 : hartebeest                                        value = 5.460    prob = 0.1%
        343 : warthog                                           value = 5.416    prob = 0.1%
        9 : ostrich, Struthio camelus                           value = 5.292    prob = 0.1%
        353 : gazelle                                           value = 5.248    prob = 0.1%
        293 : cheetah, chetah, Acinonyx jubatus                 value = 4.881    prob = 0.0%

Note that we use model1 for LRP which is initialized with lrp=True which is not used in further explanation methods(except Beyond Attention).#

LRP#

[7]:

lrp = LRP(model1)
mask_lrp = lrp.explain(image)
inp,mask_lrp = postprocess(image, mask_lrp)
cam = show_exp_on_image(inp[0], mask_lrp[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_13_0.png

Integrated Gradients#

[8]:

ig = IntegratedGradients(model2)
mask_ig = ig.explain(image)
inp,mask_ig = postprocess(image, mask_ig)
cam = show_exp_on_image(inp[0], mask_ig[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_15_0.png

Raw Attention#

[9]:

ra = RawAttention(model2)
mask_ra = ra.explain(image,7)

# Take mean over all the heads and then visualize CLS token
mask_ra = mask_ra.mean(dim=1)[:,0,1:]

#Uncomment the below line for distilled models as they have DIST token also.
#You can try visualizing the DIST token as well.
# mask_ra = mask_ra.mean(dim=1)[:,0,2:]

inp,mask_ra = postprocess(image, mask_ra)
cam = show_exp_on_image(inp[0], mask_ra[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_17_0.png

Attention Rollout#

[10]:

ar = AttentionRollout(model2)
mask_ar = ar.explain(image,layer=1)
inp,mask_ar = postprocess(image, mask_ar)
cam = show_exp_on_image(inp[0], mask_ar[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_19_0.png

GradSAM#

[11]:

gradsam = GradSAM(model2)
mask_gradsam = gradsam.explain(image)
inp,mask_gradsam = postprocess(image, mask_gradsam)
cam = show_exp_on_image(inp[0], mask_gradsam[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_21_0.png

Beyond Attention#

[12]:

ba = BeyondAttention(model1)
mask_ba = ba.explain(image)
inp,mask_ba = postprocess(image,mask_ba)
cam = show_exp_on_image(inp[0],mask_ba[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_23_0.png

Generic Attention#

[13]:

ga = GenericAttention(model2)
mask_ga = ga.explain(image)
inp,mask_ga = postprocess(image,mask_ga)
cam = show_exp_on_image(inp[0],mask_ga[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_25_0.png

Transition Attention Map(TAM)#

[14]:

#Note that TAM works with ViT-based models only.
tam = TAM(model2)
mask_tam = tam.explain(image)
inp,mask_tam = postprocess(image, mask_tam)
cam = show_exp_on_image(inp[0], mask_tam[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_27_0.png

Beyond Intuition - token-wise#

[15]:

bi = BeyondIntuition(model2)
mask_bi = bi.explain(input=image,method='token')
inp,mask_bi = postprocess(image,mask_bi)
cam = show_exp_on_image(inp[0],mask_bi[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_29_0.png

Beyond Intuition - head-wise#

[16]:

bi = BeyondIntuition(model2)
mask_bi = bi.explain(image,method='head')
inp,mask_bi = postprocess(image,mask_bi)
cam = show_exp_on_image(inp[0],mask_bi[0])
plt.imshow(cam)
plt.axis('off')
plt.show()
../_images/tutorials_explanation_methods_visualization_31_0.png