-
Notifications
You must be signed in to change notification settings - Fork 13
/
vis_swin_l.py
253 lines (217 loc) · 8.92 KB
/
vis_swin_l.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
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
import torch
import warnings
torch.autograd.set_detect_anomaly(True)
warnings.simplefilter("ignore")
import torchvision
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import cv2
import os
import argparse
import timm
import matplotlib
import matplotlib.pyplot as plt
import seaborn as sns
from utils.config_utils import load_yaml
from vis_utils import ImgLoader, get_cdict
global module_id_mapper
global features
global grads
def forward_hook(module: nn.Module, inp_hs, out_hs):
global features, module_id_mapper
layer_id = len(features) + 1
module_id_mapper[module] = layer_id
features[layer_id] = {}
features[layer_id]["in"] = inp_hs
features[layer_id]["out"] = out_hs
# print('forward_hook, layer_id:{}, hs_size:{}'.format(layer_id, out_hs.size()))
def backward_hook(module: nn.Module, inp_grad, out_grad):
global grads, module_id_mapper
layer_id = module_id_mapper[module]
grads[layer_id] = {}
grads[layer_id]["in"] = inp_grad
grads[layer_id]["out"] = out_grad
# print('backward_hook, layer_id:{}, hs_size:{}'.format(layer_id, out_grad[0].size()))
def build_model(pretrainewd_path: str,
img_size: int,
fpn_size: int,
num_classes: int,
num_selects: dict,
use_fpn: bool = True,
use_selection: bool = True,
use_combiner: bool = True,
comb_proj_size: int = None):
from models.pim_module import PluginMoodel
model = \
PluginMoodel(img_size = img_size,
use_fpn = use_fpn,
fpn_size = fpn_size,
proj_type = "Linear",
upsample_type = "Conv",
use_selection = use_selection,
num_classes = num_classes,
num_selects = num_selects,
use_combiner = use_combiner,
comb_proj_size = comb_proj_size)
if pretrainewd_path != "":
ckpt = torch.load(pretrainewd_path)
model.load_state_dict(ckpt['model'])
model.eval()
### hook original layer1~4
model.backbone.layers[0].register_forward_hook(forward_hook)
model.backbone.layers[0].register_full_backward_hook(backward_hook)
model.backbone.layers[1].register_forward_hook(forward_hook)
model.backbone.layers[1].register_full_backward_hook(backward_hook)
model.backbone.layers[2].register_forward_hook(forward_hook)
model.backbone.layers[2].register_full_backward_hook(backward_hook)
model.backbone.layers[3].register_forward_hook(forward_hook)
model.backbone.layers[3].register_full_backward_hook(backward_hook)
### hook original FPN layer1~4
model.fpn_down.Proj_layer1.register_forward_hook(forward_hook)
model.fpn_down.Proj_layer1.register_full_backward_hook(backward_hook)
model.fpn_down.Proj_layer2.register_forward_hook(forward_hook)
model.fpn_down.Proj_layer2.register_full_backward_hook(backward_hook)
model.fpn_down.Proj_layer3.register_forward_hook(forward_hook)
model.fpn_down.Proj_layer3.register_full_backward_hook(backward_hook)
model.fpn_down.Proj_layer4.register_forward_hook(forward_hook)
model.fpn_down.Proj_layer4.register_full_backward_hook(backward_hook)
### hook original FPN_UP layer1~4
model.fpn_up.Proj_layer1.register_forward_hook(forward_hook)
model.fpn_up.Proj_layer1.register_full_backward_hook(backward_hook)
model.fpn_up.Proj_layer2.register_forward_hook(forward_hook)
model.fpn_up.Proj_layer2.register_full_backward_hook(backward_hook)
model.fpn_up.Proj_layer3.register_forward_hook(forward_hook)
model.fpn_up.Proj_layer3.register_full_backward_hook(backward_hook)
model.fpn_up.Proj_layer4.register_forward_hook(forward_hook)
model.fpn_up.Proj_layer4.register_full_backward_hook(backward_hook)
return model
def cal_backward(args, out, sum_type: str = "softmax"):
assert sum_type in ["none", "softmax"]
target_layer_names = ['layer1', 'layer2', 'layer3', 'layer4',
'FPN1_layer1', 'FPN1_layer2', 'FPN1_layer3', 'FPN1_layer4', 'comb_outs']
sum_out = None
for name in target_layer_names:
if name != "comb_outs":
tmp_out = out[name].mean(1)
else:
tmp_out = out[name]
if sum_type == "softmax":
tmp_out = torch.softmax(tmp_out, dim=-1)
if sum_out is None:
sum_out = tmp_out
else:
sum_out = sum_out + tmp_out # note that use '+=' would cause inplace error
with torch.no_grad():
if args.use_label:
print("use label as target class")
pred_score = torch.softmax(sum_out, dim=-1)[0][args.label]
backward_cls = args.label
else:
pred_score, pred_cls = torch.max(torch.softmax(sum_out, dim=-1), dim=-1)
pred_score = pred_score[0]
pred_cls = pred_cls[0]
backward_cls = pred_cls
print(sum_out.size())
print("pred: {}, gt: {}, score:{}".format(backward_cls, args.label, pred_score))
sum_out[0, backward_cls].backward()
@torch.no_grad()
def get_grad_cam_weights(grads):
weights = {}
for grad_name in grads:
_grad = grads[grad_name]['out'][0][0]
L, C = _grad.size()
H = W = int(L ** 0.5)
_grad = _grad.view(H, W, C).permute(2, 0, 1)
C, H, W = _grad.size()
weights[grad_name] = _grad.mean(1).mean(1)
print(weights[grad_name].max())
return weights
@torch.no_grad()
def plot_grad_cam(features, weights):
act_maps = {}
for name in features:
hs = features[name]['out'][0]
L, C = hs.size()
H = W = int(L ** 0.5)
hs = hs.view(H, W, C).permute(2, 0, 1)
C, H, W = hs.size()
w = weights[name]
w = w.view(-1, 1, 1).repeat(1, H, W)
weighted_hs = F.relu(w * hs)
a_map = weighted_hs
a_map = a_map.sum(0)
# a_map /= abs(a_map).max()
act_maps[name] = a_map
return act_maps
if __name__ == "__main__":
global module_id_mapper, features, grads
module_id_mapper, features, grads = {}, {}, {}
"""
Please add
pretrained_path to yaml file.
"""
# ===== 0. get setting =====
parser = argparse.ArgumentParser("Visualize SwinT Large")
parser.add_argument("-pr", "--pretrained_root", type=str,
help="contain {pretrained_root}/best.pt, {pretrained_root}/config.yaml")
parser.add_argument("-img", "--image", type=str)
parser.add_argument("-sn", "--save_name", type=str)
parser.add_argument("-lb", "--label", type=int)
parser.add_argument("-usl", "--use_label", default=False, type=bool)
parser.add_argument("-sum_t", "--sum_features_type", default="softmax", type=str)
args = parser.parse_args()
load_yaml(args, args.pretrained_root + "/config.yaml")
# ===== 1. build model =====
model = build_model(pretrainewd_path = args.pretrained_root + "/best.pt",
img_size = args.data_size,
fpn_size = args.fpn_size,
num_classes = args.num_classes,
num_selects = args.num_selects)
# ===== 2. load image =====
img_loader = ImgLoader(img_size = args.data_size)
img, ori_img = img_loader.load(args.image)
# ===== 3. forward and backward =====
img = img.unsqueeze(0) # add batch size dimension
out = model(img)
cal_backward(args, out, sum_type="softmax")
# ===== 4. check result =====
grad_weights = get_grad_cam_weights(grads)
act_maps = plot_grad_cam(features, grad_weights)
# ===== 5. show =====
# cv2.imwrite("./vis_imgs/{}_ori.png".format(args.save_name), ori_img)
sum_act = None
resize = torchvision.transforms.Resize((args.data_size, args.data_size))
for name in act_maps:
layer_name = "layer: {}".format(name)
_act = act_maps[name]
_act /= _act.max()
r_act = resize(_act.unsqueeze(0))
act_m = _act.numpy() * 255
act_m = act_m.astype(np.uint8)
act_m = cv2.resize(act_m, (args.data_size, args.data_size))
# cv2.namedWindow(layer_name, 0)
# cv2.imshow(layer_name, act_m)
if sum_act is None:
sum_act = r_act
else:
sum_act *= r_act
sum_act /= sum_act.max()
sum_act = torchvision.transforms.functional.adjust_gamma(sum_act, 1.0)
sum_act = sum_act.numpy()[0]
# sum_act *= 255
# sum_act = sum_act.astype(np.uint8)
plt.cla()
cdict = get_cdict()
cmap = matplotlib.colors.LinearSegmentedColormap("jet_revice", cdict)
plt.imshow(ori_img[:, :, ::-1] / 255)
plt.imshow(sum_act, alpha=0.5, cmap=cmap) # , alpha=0.5, cmap='jet'
plt.axis('off')
plt.savefig("./{}.jpg".format(args.save_name),
bbox_inches='tight', pad_inches=0.0, transparent=True)
plt.show()
# cv2.namedWindow("ori", 0)
# cv2.imshow("ori", ori_img)
# cv2.namedWindow("heat", 0)
# cv2.imshow("heat", sum_act)
# cv2.waitKey(0)