def slice(self):
coords, colors, pcd = load_file("1.ply")
voxel_size = 0.02
colors = torch.from_numpy(colors).float()
bcoords = batched_coordinates([coords / voxel_size],
dtype=torch.float32)
tfield = TensorField(colors, bcoords)
network = nn.Sequential(
MinkowskiLinear(3, 16),
MinkowskiBatchNorm(16),
MinkowskiReLU(),
MinkowskiLinear(16, 32),
MinkowskiBatchNorm(32),
MinkowskiReLU(),
MinkowskiToSparseTensor(),
MinkowskiConvolution(32, 64, kernel_size=3, stride=2, dimension=3),
MinkowskiConvolutionTranspose(64,
32,
kernel_size=3,
stride=2,
dimension=3),
)
otensor = network(tfield)
ofield = otensor.slice(tfield)
self.assertEqual(len(tfield), len(ofield))
self.assertEqual(ofield.F.size(1), otensor.F.size(1))
ofield = otensor.cat_slice(tfield)
self.assertEqual(len(tfield), len(ofield))
self.assertEqual(ofield.F.size(1),
(otensor.F.size(1) + tfield.F.size(1)))
def setUp(self):
coords, colors, pcd = load_file("1.ply")
voxel_size = 0.02
colors = torch.from_numpy(colors).float()
bcoords = batched_coordinates([coords / voxel_size],
dtype=torch.float32)
self.tensor_field = TensorField(coordinates=bcoords, features=colors)
def test_forward(self):
coords, colors, pcd = load_file("1.ply")
device = "cuda"
X = []
Y = []
W = []
for IC in [3, 8, 16, 24, 32, 48, 64, 96, 128]:
for OC in [3, 8, 16, 24, 32, 48, 64, 96, 128, 192, 256]:
for batch_size in [1, 5, 10, 15, 20]:
for voxel_size in [0.2, 0.1, 0.075, 0.05, 0.025]:
min_times = []
for mode in [
_C.ConvolutionMode.DIRECT_GEMM,
_C.ConvolutionMode.COPY_GEMM,
]:
min_time = 100000
dcoords = torch.from_numpy(
np.floor(coords / voxel_size)
).int()
bcoords = batched_coordinates(
[dcoords for i in range(batch_size)]
)
in_feats = torch.rand(len(bcoords), IC).to(0)
sinput = SparseTensor(
in_feats, coordinates=bcoords, device=device
)
conv = MinkowskiConvolution(
in_channels=IC,
out_channels=OC,
kernel_size=3,
stride=2,
convolution_mode=mode,
dimension=3,
).to(device)
soutput = conv(sinput)
loss = soutput.F.sum()
for i in range(10):
stime = time.time()
loss.backward()
min_time = min(time.time() - stime, min_time)
min_times.append(min_time)
X.append(
[
IC,
OC,
len(sinput),
len(soutput),
]
)
Y.append(np.argmin(min_times))
W.append(np.abs(min_times[0] - min_times[1]))
print(X[-1], Y[-1], W[-1])
import pickle as pkl
with open("forward-speed.pkl", "wb") as f:
pkl.dump([X, Y, W], f)
def test_sum(self):
coords, colors, pcd = load_file("1.ply")
device = "cuda"
D = 3
batch_size = 16
voxel_size = 0.02
channels = [3, 64, 128]
dcoords = torch.from_numpy(np.floor(coords / voxel_size)).int()
bcoords = batched_coordinates([dcoords for i in range(batch_size)])
in_feats = torch.rand(len(bcoords), 3).to(0)
layer = MinkowskiStackSum(
ME.MinkowskiConvolution(
channels[0],
channels[1],
kernel_size=3,
stride=1,
dimension=3,
),
nn.Sequential(
ME.MinkowskiConvolution(
channels[0],
channels[1],
kernel_size=3,
stride=2,
dimension=3,
),
ME.MinkowskiStackSum(
nn.Identity(),
nn.Sequential(
ME.MinkowskiConvolution(
channels[1],
channels[2],
kernel_size=3,
stride=2,
dimension=3,
),
ME.MinkowskiConvolutionTranspose(
channels[2],
channels[1],
kernel_size=3,
stride=1,
dimension=3,
),
ME.MinkowskiPoolingTranspose(
kernel_size=2, stride=2, dimension=D
),
),
),
ME.MinkowskiPoolingTranspose(kernel_size=2, stride=2, dimension=D),
),
).cuda()
for i in range(1000):
torch.cuda.empty_cache()
sinput = ME.SparseTensor(in_feats, coordinates=bcoords, device=device)
layer(sinput)
def test_pcd(self):
coords, colors, pcd = load_file("1.ply")
voxel_size = 0.02
colors = torch.from_numpy(colors)
bcoords = batched_coordinates([coords / voxel_size])
tfield = TensorField(colors, bcoords)
self.assertTrue(len(tfield) == len(colors))
stensor = tfield.sparse()
print(stensor)
def test_conv(self):
IC, OC = 3, 16
coords, colors, pcd = load_file("1.ply")
kernel_size = [3, 3, 3]
kernel_stride = [2, 2, 2]
kernel_dilation = [1, 1, 1]
# size, in, out
kernel = torch.rand(np.prod(kernel_size), IC, OC).to(0)
kernel_generator = KernelGenerator(
kernel_size=kernel_size,
stride=kernel_stride,
dilation=kernel_dilation,
expand_coordinates=False,
dimension=3,
)
for batch_size in [1, 5, 10, 20, 40]:
for voxel_size in [0.05, 0.035, 0.02]:
min_time = 100000
dcoords = torch.from_numpy(np.floor(coords / voxel_size)).int()
bcoords = batched_coordinates(
[dcoords for i in range(batch_size)])
for i in range(10):
manager = _C.CoordinateMapManagerGPU_c10()
# batch insert
in_key, (unique_map, inverse_map) = manager.insert_and_map(
bcoords.to(0), [1, 1, 1], "")
in_feats = torch.rand(manager.size(in_key), IC).to(0)
out_key = _C.CoordinateMapKey(4)
stime = time.time()
out_features = _C.ConvolutionForwardGPU(
in_feats,
kernel,
kernel_generator.kernel_size,
kernel_generator.kernel_stride,
kernel_generator.kernel_dilation,
kernel_generator.region_type,
kernel_generator.region_offsets,
kernel_generator.expand_coordinates,
in_key,
out_key,
manager,
)
min_time = min(time.time() - stime, min_time)
print(
f"{batch_size}\t{manager.size(in_key)}\t{manager.size(out_key)}\t{min_time}"
)
def stride_slice(self):
coords, colors, pcd = load_file("1.ply")
voxel_size = 0.02
colors = torch.from_numpy(colors).float()
bcoords = batched_coordinates([coords / voxel_size],
dtype=torch.float32)
tfield = TensorField(colors, bcoords)
network = nn.Sequential(
MinkowskiToSparseTensor(),
MinkowskiConvolution(3, 8, kernel_size=3, stride=4, dimension=3),
MinkowskiReLU(),
MinkowskiConvolution(8, 16, kernel_size=3, stride=4, dimension=3),
)
otensor = network(tfield)
ofield = otensor.slice(tfield)
def test_decomposition(self):
coords, colors, pcd = load_file("1.ply")
colors = torch.from_numpy(colors)
for batch_size in [1, 5, 10, 20, 40]:
for voxel_size in [0.02]:
dcoords = torch.from_numpy(np.floor(coords / voxel_size)).int()
bcoords = batched_coordinates(
[dcoords for i in range(batch_size)])
feats = torch.cat([colors for b in range(batch_size)], 0)
sinput = SparseTensor(feats, bcoords)
(
decomposed_coords,
decomposed_feats,
) = sinput.decomposed_coordinates_and_features
print([len(c) for c in decomposed_coords])
print([len(f) for f in decomposed_feats])
self.assertEqual(len(decomposed_coords), batch_size)
self.assertEqual(len(decomposed_feats), batch_size)
def test_network_device(self):
coords, colors, pcd = load_file("1.ply")
voxel_size = 0.02
colors = torch.from_numpy(colors)
bcoords = batched_coordinates([coords / voxel_size])
tfield = TensorField(colors, bcoords, device=0).float()
network = nn.Sequential(
MinkowskiLinear(3, 16),
MinkowskiBatchNorm(16),
MinkowskiReLU(),
MinkowskiLinear(16, 32),
MinkowskiBatchNorm(32),
MinkowskiReLU(),
MinkowskiToSparseTensor(),
MinkowskiConvolution(32, 64, kernel_size=3, stride=2, dimension=3),
).to(0)
print(network(tfield))
def field_to_sparse(self):
coords, colors, pcd = load_file("1.ply")
voxel_size = 0.02
colors = torch.from_numpy(colors).float()
bcoords = batched_coordinates([coords / voxel_size],
dtype=torch.float32)
tfield = TensorField(colors, bcoords)
network = nn.Sequential(
MinkowskiToSparseTensor(),
MinkowskiConvolution(3, 8, kernel_size=3, stride=4, dimension=3),
MinkowskiReLU(),
MinkowskiConvolution(8, 16, kernel_size=3, stride=4, dimension=3),
)
otensor = network(tfield)
field_to_sparse = tfield.sparse(
coordinate_map_key=otensor.coordinate_map_key)
self.assertTrue(len(field_to_sparse.F) == len(otensor))
def test_decomposition_gpu(self):
print(f"{self.__class__.__name__}: test_decomposition_gpu")
if not torch.cuda.is_available():
return
coords, colors, pcd = load_file("1.ply")
colors = torch.from_numpy(colors)
for batch_size in [5, 10, 20, 40]:
for voxel_size in [0.02]:
dcoords = torch.from_numpy(np.floor(coords / voxel_size)).int()
bcoords = batched_coordinates([dcoords for i in range(batch_size)])
feats = torch.cat([colors for b in range(batch_size)], 0)
sinput = SparseTensor(feats.to(0), bcoords.to(0))
(
decomposed_coords,
decomposed_feats,
) = sinput.decomposed_coordinates_and_features
print([len(c) for c in decomposed_coords])
print([len(f) for f in decomposed_feats])
self.assertEqual(len(decomposed_coords), batch_size)
self.assertEqual(len(decomposed_feats), batch_size)