for i in range(dataset.data.NB):
x, y = dataset.data.getImageAndLabel(i, torchformat=True)
x, y = x.cuda(), y.cuda()
h, w = y.shape[0], y.shape[1]
globalresize = torch.nn.AdaptiveAvgPool2d((h, w))
power2resize = torch.nn.AdaptiveAvgPool2d(
((h // 64) * 64, (w // 64) * 64))
x = power2resize(x)
z = largeforward(net, x.unsqueeze(0))
z = globalresize(z)
z = (z[0, 1, :, :] > z[0, 0, :, :]).float()
for size in ["0", "1", "2"]:
cm[size] += noisyairs.confusion(y, z, size=int(size))
cm["bordonly"] = cm["0"] - cm["2"]
if False:
nextI = len(os.listdir("build"))
debug = noisyairs.torchTOpil(globalresize(x))
debug = PIL.Image.fromarray(numpy.uint8(debug))
debug.save("build/" + str(nextI) + "_x.png")
debug = y.cpu().numpy() * 255
debug = PIL.Image.fromarray(numpy.uint8(debug))
debug.save("build/" + str(nextI) + "_y.png")
debug = z.cpu().numpy() * 255
debug = PIL.Image.fromarray(numpy.uint8(debug))
debug.save("build/" + str(nextI) + "_z.png")
for size in ["0", "1", "2", "bordonly"]:
zz, _ = sobel(torch.unsqueeze(z[:, 1, :, :] - z[:, 0, :, :], dim=1))
gradientdiff = torch.sum(zz * yy, dim=1)
gradientdiff = torch.mean((1 - gradientdiff) * border)
gradientdiff = gradientdiff * size / nbborder
segloss = CE + dice
regloss = gradientdiff
loss = segloss + regloss
with torch.no_grad():
printloss[0] += segloss.clone().detach()
printloss[1] += regloss.clone().detach()
z = (z[:, 1, :, :] > z[:, 0, :, :]).clone().detach().float()
for j in range(batchsize):
stats += noisyairs.confusion(y[j], z[j], size=1)
if i < 10:
print(i, "/", nbbatchs, printloss)
if i < 1000 and i % 100 == 99:
print(i, "/", nbbatchs, printloss / 100)
printloss = torch.zeros(2).cuda()
if i >= 1000 and i % 300 == 299:
print(i, "/", nbbatchs, printloss / 300)
printloss = torch.zeros(2).cuda()
if i % 1000 == 999:
torch.save(net, "build/model.pth")
perf = noisyairs.perf(stats)
print(i, "perf", perf)
stats = torch.zeros((2, 2)).cuda()
with torch.no_grad():
for i in range(dataset.data.NB):
x, y = dataset.data.getImageAndLabel(i, torchformat=True)
x, y = x.cuda(), y.cuda()
h, w = y.shape[0], y.shape[1]
globalresize = torch.nn.AdaptiveAvgPool2d((h, w))
power2resize = torch.nn.AdaptiveAvgPool2d(
((h // 64) * 64, (w // 64) * 64))
x = power2resize(x)
z = largeforward(net, x.unsqueeze(0))
z = globalresize(z)
z = (z[0, 1, :, :] > z[0, 0, :, :]).float()
cm += noisyairs.confusion(y, z, size=size)
if False:
nextI = len(os.listdir("build"))
debug = noisyairs.torchTOpil(globalresize(x))
debug = PIL.Image.fromarray(numpy.uint8(debug))
debug.save("build/" + str(nextI) + "_x.png")
debug = y.float()
debug = debug * 2 * (1 - noisyairs.isborder(y, size=size))
debug = debug + noisyairs.isborder(y, size=size)
debug *= 127
debug = debug.cpu().numpy()
debug = PIL.Image.fromarray(numpy.uint8(debug))
debug.save("build/" + str(nextI) + "_y.png")
debug = z.cpu().numpy() * 255
debug = PIL.Image.fromarray(numpy.uint8(debug))