def __init__(self,imgs,labels,bounds):
super(KumarPatchDataSet,self).__init__()
self.imgs = torch.tensor(imgs).float()
self.labels = torch.tensor(labels).float()
self.bounds = torch.tensor(bounds).float()
self.idx_array = list(np.arange(64).reshape(8,8)[1:7,1:7].reshape(36,).astype('int'))
# 提取patches数据出来
self.img_patches, self.img_ibs, self.img_shs, self.img_sws = extract_patches(self.imgs,256,128)
self.label_patches, self.label_ibs, self.label_shs, self.label_sws = extract_patches(self.labels,256,128)
self.bound_patches, self.bound_ibs, self.bound_shs, self.bound_sws = extract_patches(self.bounds,256,128)
def operation(chunk, kernel):
np.seterr(all='ignore')
x = util.extract_patches(chunk, kernel)
s1 = np.sum(x, axis=(-3, -2))**2
s2 = np.sum(x**2, axis=(-3, -2))
sembl = s1.sum(axis=-1) / s2.sum(axis=-1)
sembl /= kernel[0] * kernel[1]
return (sembl)
def operation(chunk, kernel):
np.seterr(all='ignore')
ki, kj, kk = kernel
patches = util.extract_patches(chunk, kernel)
out_data = []
for i in range(0, patches.shape[0]):
traces = patches[i]
traces = traces.reshape(-1, ki * kj * kk)
cov = np.apply_along_axis(cov, 1, traces, ki, kj, kk)
vals = np.linalg.eigvals(cov)
vals = np.abs(vals.max(axis=1) / vals.sum(axis=1))
out_data.append(vals)
out_data = np.asarray(out_data).reshape(patches.shape[:3])
return (out_data)
def infer(model, device, image):
aug_list = ['ori', 'rot90', 'rot180', 'rot270', 'flip_h', 'flip_w']
image_tensor = torch.FloatTensor(image)
image_tensor = image_tensor.unsqueeze(dim=0)
image_tensor = image_tensor.permute(0, 3, 1, 2)
# (0,3,1000,1000)
print('image_tensor: ', image_tensor.shape)
img_num, c, h, w = image_tensor.shape
patches_d, ibs, shs, sws = extract_patches(image_tensor, 256, 128)
sigmoid = nn.Sigmoid()
# 定义最后的结果
finalsout = np.zeros([1, h, w], dtype=np.float32)
finalcout = np.zeros([1, h, w], dtype=np.float32)
# tta操作
for type_aug in aug_list:
if type_aug == 'ori':
tta_d = patches_d.clone()
if type_aug == 'rot90':
tta_d = patches_d.rot90(1, dims=(2, 3))
if type_aug == 'rot180':
tta_d = patches_d.rot90(2, dims=(2, 3))
if type_aug == 'rot270':
tta_d = patches_d.rot90(3, dims=(2, 3))
if type_aug == 'flip_h':
tta_d = patches_d.flip(2)
if type_aug == 'flip_w':
tta_d = patches_d.flip(3)
#
spred = torch.zeros(tta_d.shape[0], 1, tta_d.shape[2], tta_d.shape[3])
cpred = torch.zeros(tta_d.shape[0], 1, tta_d.shape[2], tta_d.shape[3])
for start_batch in list(range(0, tta_d.shape[0], batch_size)):
end_batch = np.min([start_batch + batch_size, tta_d.shape[0]])
input_d = tta_d[start_batch:end_batch]
if len(input_d.shape) == 3:
input_d = input_d.unsqueeze(dim=0)
input_d = input_d.to(device)
outs = model(input_d)
sout = outs[0]
cout = outs[5]
spred[start_batch:end_batch] = sout.data.cpu()
cpred[start_batch:end_batch] = cout.data.cpu()
# Inverse TTA
if type_aug == 'rot90':
spred = spred.rot90(3,
dims=(len(spred.shape) - 2,
len(spred.shape) - 1))
cpred = cpred.rot90(3,
dims=(len(cpred.shape) - 2,
len(cpred.shape) - 1))
elif type_aug == 'rot180':
spred = spred.rot90(2,
dims=(len(spred.shape) - 2,
len(spred.shape) - 1))
cpred = cpred.rot90(2,
dims=(len(cpred.shape) - 2,
len(cpred.shape) - 1))
elif type_aug == 'rot270':
spred = spred.rot90(1,
dims=(len(spred.shape) - 2,
len(spred.shape) - 1))
cpred = cpred.rot90(1,
dims=(len(cpred.shape) - 2,
len(cpred.shape) - 1))
elif type_aug == 'flip_h':
spred = spred.flip(len(spred.shape) - 2)
cpred = cpred.flip(len(cpred.shape) - 2)
elif type_aug == 'flip_w':
spred = spred.flip(len(spred.shape) - 1)
cpred = cpred.flip(len(cpred.shape) - 1)
#
spred_map = reconstruct_from_patches_weightedall(
spred, ibs, shs, sws, 256, 128, img_num, 1, w, h,
1).squeeze().numpy()
cpred_map = reconstruct_from_patches_weightedall(
cpred, ibs, shs, sws, 256, 128, img_num, 1, w, h,
1).squeeze().numpy()
#
finalsout += spred_map
finalcout += cpred_map
finalsout /= len(aug_list)
finalcout /= len(aug_list)
return finalsout, finalcout
def operation(chunk, kernel):
x = util.extract_patches(chunk, (1, 1, kernel[-1]))
out = np.trapz(x).reshape(x.shape[:3])
return (out)
def operation(chunk, kernel):
x = util.extract_patches(chunk, kernel)
out = np.sqrt(np.mean(x**2, axis=(-3, -2, -1)))
return (out)
