GPUS = 4
phi = network_wrappers.FunctionFromVectorField(
networks.tallUNet(unet=networks.UNet2ChunkyMiddle, dimension=3)
)
psi = network_wrappers.FunctionFromVectorField(networks.tallUNet2(dimension=3))
pretrained_lowres_net = inverseConsistentNet.InverseConsistentNet(
network_wrappers.DoubleNet(phi, psi),
lambda x, y: torch.mean((x - y) ** 2),
100,
)
network_wrappers.assignIdentityMap(pretrained_lowres_net, input_shape)
network_wrappers.adjust_batch_size(pretrained_lowres_net, 12)
trained_weights = torch.load(
"results/dd_l400_continue_rescalegrad2/knee_aligner_resi_net1800"
)
# trained_weights = torch.load("../results/dd_knee_l400_continue_smallbatch2/knee_aligner_resi_net9300")
# rained_weights = torch.load("../results/double_deformable_knee3/knee_aligner_resi_net22200")
pretrained_lowres_net.load_state_dict(trained_weights)
hires_net = inverseConsistentNet.InverseConsistentNet(
network_wrappers.DoubleNet(
network_wrappers.DownsampleNet(pretrained_lowres_net.regis_net, dimension=3),
network_wrappers.FunctionFromVectorField(networks.tallUNet2(dimension=3)),
torch.manual_seed(1)
torch.cuda.manual_seed(1)
np.random.seed(1)
print("=" * 50)
net = inverseConsistentNet.InverseConsistentNet(
network_wrappers.DoubleNet(
network_wrappers.RandomShift(0.25),
network_wrappers.FunctionFromVectorField(
networks.tallUNet2(dimension=2)),
),
lambda x, y: torch.mean((x - y)**2),
lmbda,
)
input_shape = next(iter(d1))[0].size()
network_wrappers.assignIdentityMap(net, input_shape)
net.cuda()
optimizer = torch.optim.Adam(net.parameters(), lr=0.001)
net.train()
xs = []
for _ in range(40):
y = np.array(train.train2d(net, optimizer, d1, d2, epochs=50))
xs.append(y)
x = np.concatenate(xs)
plt.title("Loss curve for " + type(net.regis_net).__name__ + " lambda=" +
str(lmbda))
plt.plot(x[:, :3])
plt.savefig(describe.run_dir + f"loss.png")
plt.clf()
plt.title("Log # pixels with negative Jacobian per epoch")
data_size=data_size,
hollow=True,
batch_size=batch_size)
d1_t, d2_t = data.get_dataset_triangles("test",
data_size=data_size,
hollow=True,
batch_size=batch_size)
image_A, image_B = (x[0].cuda() for x in next(zip(d1, d2)))
net = inverseConsistentNet.InverseConsistentNet(
network_wrappers.FunctionFromMatrix(networks.ConvolutionalMatrixNet()),
lambda x, y: torch.mean((x - y)**2),
100,
)
network_wrappers.assignIdentityMap(net, image_A.shape)
net.cuda()
import train
optim = torch.optim.Adam(net.parameters(), lr=0.00001)
net.train().cuda()
xs = []
for _ in range(240):
y = np.array(train.train2d(net, optim, d1, d2, epochs=50))
xs.append(y)
x = np.concatenate(xs)
plt.title("Loss curve for " + type(net.regis_net).__name__)
plt.plot(x[:, :3])
plt.savefig(describe.run_dir + f"loss.png")