MoreauGrad/data_util.py at main · buyeah1109/MoreauGrad · 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
import torch
import numpy as np
import torchvision
import torchvision.transforms as transforms
import glob
import os
import json
from PIL import Image
import matplotlib.pyplot as plt

torch.backends.cudnn.deterministic=True
torch.backends.cudnn.benchmark=False

def agg_default(x):
	if x.ndim == 4:
		return np.abs(x).sum((0, 1))

	elif x.ndim == 3:
		return np.abs(x).sum(0)

epsilon = 1e-10

def clip(x, top_clip=True):
	if x.ndim == 3:
		batch_size, height, width = x.shape
		x = x.reshape(batch_size, -1)
		if top_clip:
			vmax = np.percentile(x, 99, axis=1, keepdims=True)
		else:
			vmax = np.max(x, axis=1, keepdims=True)
		vmin = np.min(x, axis=1, keepdims=True)
		vdiff = vmax - vmin
		for i, v in enumerate(vdiff):
			v = max(0, np.abs(v))
			if np.abs(v) < epsilon:
				x[i] = np.zeros_like(x[i])
			else:
				x[i] = np.clip((x[i] - vmin[i]) / v, 0, 1)
		x = x.reshape(batch_size, height, width)
	elif x.ndim == 2:
		height, width = x.shape
		x = x.ravel()
		x = np.nan_to_num(x)
		vmax = np.percentile(x, 99) if top_clip else np.max(x)
		vmin = np.min(x)
		vdiff = max(0, np.abs(vmax - vmin))
		if np.abs(vdiff) < epsilon:
			x = np.zeros_like(x)
		else:
			x = np.clip((x - vmin) / (vmax - vmin), 0, 1)
		x = x.reshape(height, width)
	return x

def agg_clip(x, top_clip=True):
	return clip(agg_default(x), top_clip=top_clip)

def plot_saliency(mp, file):
	result1 = np.uint8(255*agg_clip(mp.cpu().numpy()))
	plt.figure()
	plt.axis('off')
	plt.imshow(result1, cmap='gray')
	plt.savefig(file, bbox_inches='tight', pad_inches=0)
	plt.close()

def pil_loader(path):
	with open(path, 'rb') as f:
		with Image.open(f) as img:
			return img.convert('RGB')

def setup_imagenet(batch_size=16, example_ids=None,
				   n_batches=-1, n_examples=-1,
				   shuffle=True, dump_name=None):
	model = torchvision.models.resnet50(pretrained=True)
	model.eval()
	model.cuda()
	print('model loaded')

	home_dir = './ILSVRC_val/'
	image_path = './ILSVRC_val/**/*.JPEG'
	image_files = list(glob.iglob(image_path, recursive=True))
#    print(len(image_files))
	image_files = sorted(image_files, key=lambda x: os.path.basename(x))
	real_ids = [os.path.basename(x) for x in image_files]

	if example_ids is not None:
		examples = {r: (r, m)
					for r, m in zip(real_ids, image_files)}
		examples = [examples[x] for x in example_ids]
	else:
		examples = list(zip(real_ids, image_files))

	if shuffle:
		np.random.seed(0)
		np.random.shuffle(examples)

	if n_examples > 0:
		examples = examples[:n_examples]
	elif n_batches > 0:
		examples = examples[:batch_size * n_batches]
	else:
		print('using all images')

	selected_files = sorted([x[0] for x in examples])
	if dump_name is not None:
		with open(dump_name, 'w') as f:
			f.write(json.dumps(selected_files))
#    print('\n'.join(selected_files))

	def batch_loader(batch):
		batch = list(map(list, zip(*batch)))
		ids, xs = batch
		return (ids, [pil_loader(x) for x in xs])

	batch_indices = list(range(0, len(examples), batch_size))
	batches = [examples[i: i + batch_size] for i in batch_indices]
	batches = map(batch_loader, batches)
#    print('image loaded', len(batch_indices))

	return model, batches