PathwayGrad/QualitativeAnalysis_pyfile.py at main · CAMP-eXplain-AI/PathwayGrad · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
import random
from torch.nn import functional as F
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import torch
from torchvision import datasets, models, transforms
from tqdm import tqdm
from torchray.attribution.grad_cam import grad_cam
from src import Pruner, Plot_tools
from src.attribution_methods import vanilla_saliency, integrated_gradients, grad_times_image, guided_backprop

# ImageNet as of Oct2019 can no longer be downloaded using pytorch.
# https://github.com/pytorch/vision/issues/1453
# To download ImageNet, see http://image-net.org/.

imagenet_dir = '/home/ashkan/data/ILSVRC2012/'
transform = transforms.Compose([transforms.Resize(256),
                                transforms.CenterCrop(224),
                                transforms.ToTensor(),
                                transforms.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
imagenet = datasets.ImageNet(imagenet_dir, download=False, split='val', transform=transform)
classes = imagenet.classes
mpl.rcParams['figure.dpi']= 1200


# ### Method to get attributions

def get_attribution(attribution_name, data, model, gradcam_layer, model_sparsity_threshold):
    make_single_channel = True
    class_id = model(data)
    class_id = class_id.data.max(1)[1].item()
    # print(class_id)
    model.eval()
    if attribution_name == "Gradients":
        vanilla_sal = vanilla_saliency.VanillaSaliency(model, device)
        saliency = vanilla_sal.generate_saliency(data, class_id, make_single_channel)
    elif attribution_name == "InputMCT":
        saliency = grad_times_image.generate_grad_times_image_saliency(model, data, class_id, device, make_single_channel)
    elif attribution_name == "InputIntGrad":
        integ_grad = integrated_gradients.IntegratedGradients(model, device)
        saliency = integ_grad.generate_integrated_gradients(data, class_id, 50, make_single_channel)
    elif attribution_name == "GBP":
        GB = guided_backprop.GuidedBackprop(model, device)
        saliency = GB.generate_gradients(data, class_id, make_single_channel)

    elif attribution_name == "GradCAM":
        saliency = grad_cam(model, data, class_id, saliency_layer=gradcam_layer)
        saliency = F.interpolate(saliency, 224, mode="bilinear")
        saliency = saliency.detach().cpu().numpy()

    elif attribution_name == "NeuronMCT":
        pruner = Pruner.Pruner(model, data, device)
        pruner.prune_neuron_mct(model_sparsity_threshold, debug=False)
        saliency = pruner.generate_saliency(make_single_channel=make_single_channel)
        pruner.remove_handles()

    elif attribution_name == "PrunedRandom":
        pruner = Pruner.Pruner(model, data, device)
        pruner.prune_random(model_sparsity_threshold, debug=False)
        saliency = pruner.generate_saliency(make_single_channel=make_single_channel)
        pruner.remove_handles()

    elif attribution_name == "NeuronIntGrad":
        pruner = Pruner.Pruner(model, data, device)
        pruner.prune_integrad(model_sparsity_threshold, debug=False)
        saliency = pruner.generate_saliency(make_single_channel=make_single_channel)
        pruner.remove_handles()

    elif attribution_name == "PruneGreedy":
        pruner = Pruner.Pruner(model, data, device)
        pruner.prune_greedy(1, model_sparsity_threshold, debug=False)
        saliency = pruner.generate_saliency(make_single_channel=make_single_channel)
        pruner.remove_handles()

    elif attribution_name == "CDRP":
        pruner = Pruner.Pruner(model, data, device)
        pruner.prune_dgr(model_sparsity_threshold, debug=False)
        saliency = pruner.generate_saliency(make_single_channel=make_single_channel)
        pruner.remove_handles()

    elif attribution_name == "PrunePGD":
        pruner = Pruner.Pruner(model, data, device)
        pruner.prune_neuron_mct(model_sparsity_threshold, debug=False)
        saliency = pruner.generate_saliency_pgd_l2(epsilon=500, alpha=25, num_iter=50, make_single_channel=make_single_channel, debug=False)
        pruner.remove_handles()

    if make_single_channel:
        saliency = torch.from_numpy(np.asarray(saliency)).view([1, 224, 224])
    else:
        saliency = torch.from_numpy(np.asarray(saliency)).view([3, 224, 224])
    saliency /= np.max(np.asarray(abs(saliency)).flatten())
    return saliency


def visualize(model, dataloader, images, classes, gradcam_layer, model_sparsity_threshold):
    global id
    model = model.to(device)
    model.eval()
    name_methods = ["Gradients", "GBP", "GradCAM", "InputMCT", "InputIntGrad", "NeuronMCT", "NeuronIntGrad"]
    num_methods = len(name_methods)
    dataiter = iter(dataloader)
    i = 0
    fig = plt.figure(figsize=(7, num_samples)) #figsize=(7, 10*2))
    acts = []
    for chosen in tqdm(range(num_samples)):
        data, _ = dataiter.next()
        data = data.to(device)
        output = model(data.clone())
        output = torch.nn.functional.softmax(output.detach(), dim=1)
        predicted_logit = output.data.max(1)[1].item()
        predicted_prob = output.data.max(1)[0].item()
        image = Plot_tools.reverse_preprocess_imagenet_image(data.clone())
        ax = fig.add_subplot(num_samples, num_methods+1, chosen*(num_methods+1) + 1)
        if chosen == 0:
            ax.set_title("Original Image", fontsize=7)
        ax.text(-0.13, 0.5, classes[predicted_logit][0]+"\n"+str("%.2f" % round(predicted_prob*100, 2)+"%"), fontsize=7, rotation=90, horizontalalignment='center', verticalalignment='center', transform=ax.transAxes)
        plt.axis('off')
        plt.imshow(image)
        for j in range(num_methods):
            ax = fig.add_subplot(num_samples, num_methods+1, chosen*(num_methods+1) + 1 + j + 1)
            attribution = get_attribution(name_methods[j], data.clone(), model, gradcam_layer,  model_sparsity_threshold)
            attribution = np.asarray(attribution.squeeze(0))
            if chosen == 0:
                if name_methods[j] == "NeuronMCT" or name_methods[j] == "NeuronIntGrad":
                    ax.set_title(name_methods[j], fontsize=7, fontweight='bold')
                else:
                    ax.set_title(name_methods[j], fontsize=7)
            plt.imshow(abs(attribution), cmap='jet', vmin=0, vmax=1)
            plt.axis('off')
        i += 1
    plt.tight_layout()
    plt.subplots_adjust(wspace=0.05, hspace=0.01)
    plt.savefig('appendix_attribution_'+str(indices)+'.png', dpi=200) #'./fig'+str(id)+'.png', dpi=200)
    id += 1


# # Evaluate on Resnet50

num_samples = 9
id = 0
indices = [5054, 36203, 43518, 20813, 7054, 15574, 15598, 41442, 9678] #RESNET APPENDIX
dataset = torch.utils.data.Subset(imagenet, indices)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=1)

model = models.resnet50(pretrained=True)
device = 'cuda:0'
resnet_gradcam_layer = 'layer4'
model_sparsity_threshold = 75
visualize(model, dataloader, indices, classes, resnet_gradcam_layer, model_sparsity_threshold)


# # Evaluate on VGG16

num_samples = 9 #1
# indices = [40639] # airplane
indices = [39834, 40639, 3158, 8105, 2094, 23684, 12659, 23147, 4581]  #VGG16 APPENDIX
dataset = torch.utils.data.Subset(imagenet, indices)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=1)

model = models.vgg16(pretrained=True)
vgg_gradcam_layer = 'features'
model_sparsity_threshold = 90
visualize(model, dataloader, indices, classes, vgg_gradcam_layer, model_sparsity_threshold)