import torch.nn as nn
import torch
class Simple(nn.Module):
def __init__(self):
super().__init__()
self.feature = nn.Sequential(
nn.Conv2d(3, 16, 3, 1, 1, bias=True),
nn.MaxPool2d(2),
nn.Conv2d(16, 32, 3, 1, 1, bias=True),
nn.MaxPool2d(2),
nn.Conv2d(32, 64, 3, 1, 1, bias=True),
nn.MaxPool2d(2),
nn.Conv2d(64, 128, 3, 1, 1, bias=True),
nn.MaxPool2d(2),
nn.Conv2d(128, 256, 3, 1, 1, bias=True),
nn.MaxPool2d(2),
nn.Conv2d(256, 512, 3, 1, 1, bias=True)
)
self.fc = nn.Sequential(
nn.Linear(512*16, 1000),
nn.ReLU(),
nn.Linear(1000, 100),
nn.ReLU(),
nn.Linear(100, 6)
)
def forward(self, x):
f = self.feature(x)
f = f.view(-1, 512*16)
return self.fc(f)
net=Simple()
paras=torch.load('', map_location='cuda')
net.load_state_dict(paras)
net.cuda().eval()
raw = cv2.imread('')
im = cv2.resize(raw, (150, 150))
im = im / 255.0
im = np.transpose(im, (2,0,1))
im = im.reshape(1,3,150,150)
im = torch.from_numpy(im).float().cuda()
for ch in net.feature.children():
im = ch(im)
b, c, h, w = im.shape
for ci in range(c):
fm = im[0][ci]
ma = torch.max(fm)
mi = torch.min(fm)
fm = 255 * (fm - mi) / (ma - mi)
fm = fm.cpu().detach().numpy().astype('uint8')
fm = fm.reshape(h, w, 1)
cv2.imwrite('log/feature/{}-{}.jpg'.format(h, ci), fm)
本文内容由网友自发贡献,版权归原作者所有,本站不承担相应法律责任。如您发现有涉嫌抄袭侵权的内容,请联系:hwhale#tublm.com(使用前将#替换为@)