-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathvis_causal_metric_probs.py
141 lines (116 loc) · 5.75 KB
/
vis_causal_metric_probs.py
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
import os
from os.path import join, isdir, isfile
import sys
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from path_dict import PathDict
path_dict = PathDict()
proj_root = path_dict.proj_root
ds_root = path_dict.ds_root
from utils.CausalMetric import plot_causal_metric_curve, auc
from utils.ReadingDataset import get_frames, load_model_and_dataset
from process_perturb_res import vis_perturb_res, get_perturb_acc_dict
# from perturb.perturb_utils import *
from utils.ImageShow import *
import torch
from torch.utils.data import Dataset, DataLoader
import torch.nn.functional as F
from torchvision import transforms
from tqdm import tqdm
import math
import numpy as np
import pandas as pd
from skimage import transform
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--dataset", type=str, default="ucf101",
choices=["ucf101", "epic"])
parser.add_argument("--model", type=str, default="r2p1d",
choices=["v16l", "r2p1d", "r50l"])
# parser.add_argument("--vis_method", type=str,
# choices=["g", "ig", "sg", "sg2", "grad_cam", "perturb", "random", "eb", "la", "gbp"])
parser.add_argument("--mode", type=str, default="ins", choices=["ins", "del", "both"])
parser.add_argument("--order", type=str, default="most_first", choices=["most_first", "least_first"])
parser.add_argument("--new_size", type=int, default=16)
parser.add_argument('--extra_label', type=str, default="")
parser.add_argument('--save_perturbed_video', action="store_true")
parser.add_argument('--perturb_ratio', type=float, default=0.2)
args = parser.parse_args()
num_score = 128 + 1
x_coords = np.arange(num_score) / num_score
# plt.rcParams.update({'font.size': 16})
fig = plt.figure(figsize=(8,8))
plt.subplot(111)
plt.xlim(0, 1.0)
plt.ylim(0, 1.05)
plt.xlabel('Pixels Perturbed (%)', fontsize=20)
plt.ylabel('Probability (%)', fontsize=20)
frame_transform = transforms.Compose([
transforms.Resize((112,112)),
transforms.ToTensor(),
transforms.Normalize([0.43216, 0.394666, 0.37645], [0.22803, 0.22145, 0.216989]),
])
video_name = "v_Skijet_g04_c02"
vis_methods = ["grad_cam", "random"]
colors = ["orange", "blue"]
final_save_label = f"{args.dataset}_{args.model}_{args.mode}_{vis_methods[0]}_{vis_methods[1]}"
final_save_path = join(proj_root, "cm_probs_vis", final_save_label)
os.makedirs(final_save_path, exist_ok=True)
for midx, vis_method in enumerate(vis_methods):
method_save_label = f"{args.dataset}_{args.model}_{args.mode}_{vis_method}"
if args.extra_label != "":
method_save_label += f"{args.extra_label}"
if args.new_size != None:
method_save_label += f"_{args.new_size}"
probs_save_path = join(proj_root, "cm_probs", method_save_label+".csv")
probs_df = pd.read_csv(probs_save_path)
probs_dict = {}
for ridx, row in probs_df.iterrows():
row_els = list(dict(row).values())
probs_dict[row_els[0]] = row_els[1:]
video_probs = probs_dict[video_name]
plt.plot(x_coords, video_probs, color=colors[midx], linestyle='solid', linewidth=2, label=f"{vis_method}: {video_name[2:]}")
probs_array = np.array(list(probs_dict.values()))
avg_probs = probs_array.mean(axis=0)
plt.plot(x_coords, avg_probs, color=colors[midx], linestyle='dashed', linewidth=2, label=f"{vis_method}: Average")
plt.axvline(args.perturb_ratio, 0, 1, color='gray', linewidth=2)
if vis_method != 'random':
res_save_path = join(proj_root, 'exe_res', f"{args.dataset}_{args.model}_{vis_method}_full.pt")
res_dic_lst = torch.load(res_save_path)['val']
for res_dic in res_dic_lst:
if res_dic["video_name"].split("/")[-1] == video_name:
masks_tensor = torch.from_numpy(res_dic["mask"].astype('float32')).transpose(1,0) # Tx1xH'xW'
fidxs = res_dic["fidx"]
break
frames = get_frames(args.dataset, args.model, video_name, fidxs) # Tx3xHxW
frames_tensor = torch.stack([frame_transform(Image.fromarray(frame)) for frame in frames], dim=0)
nt, nch, nrow, ncol = frames_tensor.shape
else:
masks_tensor = torch.randn((nt, 1, nrow, ncol))
if masks_tensor.shape[-1] != frames_tensor.shape[-1]:
masks_tensor = F.interpolate(masks_tensor, size=(nrow, ncol), mode="bilinear")
assert nrow % args.new_size == 0
ks = nrow // args.new_size
k = torch.ones((1, 1, ks, ks)) / (ks*ks)
small_masks_tensor = F.conv2d(masks_tensor, k, stride=ks, padding=0) # Tx1x sH x sW
# T*sH*sW
sal_order = small_masks_tensor.reshape(-1).argsort(dim=-1, descending=True)
perturb_topk = int(sal_order.shape[0] * args.perturb_ratio)
pos = sal_order[:perturb_topk]
if args.mode == "del":
pmasks = torch.ones(nt, 1, args.new_size, args.new_size)
pmasks.reshape(-1)[pos] = 0
elif args.mode == "ins":
pmasks = torch.zeros(nt, 1, args.new_size, args.new_size)
pmasks.reshape(-1)[pos] = 1
masks_tensor = F.interpolate(pmasks, size=(nrow, ncol), mode='nearest') # T x 1 x H x W
pframes_tensor = frames_tensor * masks_tensor + torch.zeros_like(frames_tensor) * (1 - masks_tensor)
pframe_save_path = join(final_save_path, f"{video_name}_{vis_method}")
os.makedirs(pframe_save_path, exist_ok=True)
for it in range(nt):
pframe_tensor = pframes_tensor[it]
pframe_np = img_tensor_to_np(pframe_tensor) # 0~1
pframe_forshow = (pframe_np * 255).astype(np.uint8).transpose(1,2,0)
Image.fromarray(pframe_forshow).save(join(pframe_save_path, f'{it:02d}.pdf'))
# plt.legend(loc='upper center', bbox_to_anchor=(0.5, 1.15), ncol=2, fontsize=16)
plt.legend(loc='upper left', fontsize=18)
plt.savefig(join(final_save_path, "comparison.pdf"))