def fft_ar(R):
# imgs = []
# for i in range(R.shape[0]):
print("R.shape", R.shape)
n_levels = 8
R_thr = -10
h = R.shape[1]
w = R.shape[2]
print("R[0]")
# loga(R[0])
mask_R0 = R[0] >= R_thr
mask_R1 = R[1] >= R_thr
mask_R2 = R[2] >= R_thr
mask_R = mask_R0 * mask_R1 * mask_R2
mask_R = mask_R[0,0].astype(np.float32)
res = fft_decompose(R, ar_order=2, n_cascade_levels=n_levels, R_thr=R_thr)
out_imgs = []
print(">>> run fft...")
# for each level
for i in range(0, n_levels):
img0 = res[0]["cascade_levels"][i]
img1 = res[1]["cascade_levels"][i]
img2 = res[2]["cascade_levels"][i]
# for each image
img0 = np.reshape(img0, (1, 1, h, w))
img1 = np.reshape(img1, (1, 1, h, w))
img2 = np.reshape(img2, (1, 1, h, w))
img0 = torch.from_numpy(img0)
img1 = torch.from_numpy(img1)
img2 = torch.from_numpy(img2)
if use_cuda:
img0 = img0.cuda()
img1 = img1.cuda()
img2 = img2.cuda()
# load model
corr_module = Corr(allones=True)
if use_cuda:
corr_module = corr_module.cuda()
print("img0", img0.shape)
print("img1", img1.shape)
print("img2", img2.shape)
img3 = corr_module(img0, img1, img2, mask_R)
#plt.imshow(img3[0,0])
#plt.show()
print(img3.shape)
out_imgs.append(img3[0,0])
out = fft_recompose(out_imgs)
print("out.shape", out.shape)
img1 = R[1]
img2 = R[2]
res = fft_decompose(R, ar_order=2, n_cascade_levels=8, R_thr=-10)
print(">>> ori img0")
#loga(img0)
print(res[0].keys())
for i in range(8):
img = res[0]["cascade_levels"][i]
#plt.imshow(img)
#plt.show()
R = res[0]["cascade_levels"]
print("res[0].keys()", res[0].keys())
out_img = fft_recompose(R)
print(">>> out_img")
# loga(out_img)
print("img0", img0.shape)
h = img0.shape[0]
w = img0.shape[1]
img0 = np.reshape(img0, (1, 1, h, w))
img1 = np.reshape(img1, (1, 1, h, w))
img2 = np.reshape(img2, (1, 1, h, w))
img0 = torch.from_numpy(img0)
img1 = torch.from_numpy(img1)
img2 = torch.from_numpy(img2)
if use_cuda:
def fft_ar(R):
# load model
# corr_module = Corr(image_level=True)
corr_module = Corr(window_size=3, sigma=1)
print("R.shape", R.shape)
n_levels = 8
R_thr = -10
h = R.shape[1]
w = R.shape[2]
print("R[0]")
# loga(R[0])
mask_R0 = R[0] >= R_thr
mask_R1 = R[1] >= R_thr
mask_R2 = R[2] >= R_thr
mask_R = mask_R0 * mask_R1 * mask_R2
print(">>> mask_R.shape", mask_R.shape)
# mask_R = mask_R[0,0].astype(np.float32)
# attention here
mask_R = mask_R.astype(np.int)
mask_R = torch.from_numpy(mask_R)
print(">>> mask_R type", type(mask_R))
print(mask_R)
res = fft_decompose(R, ar_order=2, n_cascade_levels=n_levels, R_thr=R_thr)
print("R.shape", R.shape)
print("res.shape", res[0]["cascade_levels"].shape)
# origin_img = fft_recompose(res[0]["cascade_levels"])
# plt.imshow(origin_img)
# plt.show()
# raise ValueError()
out_imgs = []
print(">>> run fft...")
# for each level
for i in range(0, n_levels):
img0 = res[0]["cascade_levels"][i]
img1 = res[1]["cascade_levels"][i]
img2 = res[2]["cascade_levels"][i]
# for each image
img0 = np.reshape(img0, (1, 1, h, w))
img1 = np.reshape(img1, (1, 1, h, w))
img2 = np.reshape(img2, (1, 1, h, w))
img0 = torch.from_numpy(img0)
img1 = torch.from_numpy(img1)
img2 = torch.from_numpy(img2)
if use_cuda:
img0 = img0.cuda()
img1 = img1.cuda()
img2 = img2.cuda()
if use_cuda:
corr_module = corr_module.cuda()
print(">>> diff")
print(torch.sum(torch.abs(img0 - img1)))
# print(">>> maskR")
# loga(mask_R)
# raise ValueError()
# print(">>> loga img0, img1, img2")
# loga(img0)
# loga(img1)
# loga(img2)
# print("pre mask_R")
# loga(mask_R)
img3 = corr_module(img0, img1, img2, mask_R)
# print("after mask_R")
# loga(mask_R)
print("img3", img3.shape)
print(img3.shape)
out_imgs.append(img3[0, 0].data.numpy())
# loga(img3) #shape:(1, 1, 320, 355) type:(float32 of torch.Tensor) max: -5.1484, min: -15.0, mean: -14.286
# raise ValueError()
for i in range(8):
print("current i = ", i)
print(out_imgs[i][56, 192])
# loga(out_imgs[i])
out = fft_recompose(out_imgs)
print("out,", out[56, 192])
print("R[0]")
# loga(R[0])
vis_radar(R[0], "R0.png")
print("R[1]")
# loga(R[1])
vis_radar(R[1], "R1.png")
print("R[2]")
# loga(R[2])
vis_radar(R[2], "R2.png")
vis_radar(out, "out.png")
print("out.shape", out.shape)
print("mask_R.shape", mask_R.shape)
# loga(mask_R)
# raise ValueError()
print("type out", type(out))
mask_R = mask_R.data.numpy()
print("type mask_R", type(mask_R))
assert type(out) == type(mask_R)
out[mask_R <= 0.5] = -15
# loga(out)
vis_radar(out)
plt.imshow(out)
plt.show()
print("out.shape", out.shape)