def train_shape_net(c_in=30, evec=False):
bs_train, bs_test = 32, 32
train_files = get_files_list('../data/eigen/eigen_train_files.txt')
test_files = get_files_list('../data/eigen/eigen_test_files.txt')
ds_train = EigenNet40Ds(train_files, evec=evec, c_in=c_in)
ds_test = EigenNet40Ds(test_files, evec=evec, c_in=c_in)
dl_train = DataLoader(ds_train, bs_train, shuffle=True)
dl_test = DataLoader(ds_test, bs_test, shuffle=True)
lr = 1e-4
l2_reg = 0
our_model = HearShapeNet(c_in=c_in)
# our_model = SpectralPointNet()
loss_fn = F.cross_entropy # This criterion combines log_softmax and nll_loss in a single function
optimizer = torch.optim.Adam(our_model.parameters(),
lr=lr,
weight_decay=l2_reg)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=20,
gamma=0.7)
trainer = NetTrainer(our_model, loss_fn, optimizer, scheduler)
expr_name = f'Eigen-t3'
if os.path.isfile(f'results/{expr_name}.pt'):
os.remove(f'results/{expr_name}.pt')
fit_res = trainer.fit(dl_train,
dl_test,
num_epochs=10000,
early_stopping=50,
checkpoints=expr_name)
return
def create_dist_eig_dataset(train, num_eigen, num_nbrs, num_points,
num_workers):
if train:
name = 'train'
else:
name = 'test'
bs = 512
dest_dir = 'dist_eig_2048'
pc_files = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/{name}_files.txt')
ds = CreateLboEig(pc_files,
num_eigen=num_eigen,
num_nbrs=num_nbrs,
num_points=num_points)
dl = DataLoader(ds, bs, shuffle=False, num_workers=num_workers)
files_list = []
dl_iter = iter(dl)
num_batches = len(dl.batch_sampler)
for batch_idx in range(num_batches):
print(f'Working on file {batch_idx}', flush=True)
eigen_vec, eigen_val = next(dl_iter)
print(f'eigen_vec.shape={eigen_vec.shape}')
print(f'eigen_val.shape={eigen_val.shape}')
file_name = f'data/{dest_dir}/spectral_data_{name}{batch_idx}.h5'
files_list.append(file_name)
with h5py.File('../' + file_name, 'w') as hf:
hf.create_dataset("eigen_vec", data=eigen_vec)
hf.create_dataset("eigen_val", data=eigen_val)
txt_name = f'../data/{dest_dir}/spectral_{name}_files.txt'
with open(txt_name, 'w') as f:
f.write('\n'.join(files_list))
return
def create_normals_dataset(train: bool, num_nbrs: int, num_points: int,
num_workers: int):
if train:
name = 'train'
else:
name = 'test'
bs = 512
dest_dir = 'normals_5nbr_2048'
pc_files = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/{name}_files.txt')
ds = CreateNormals40Ds(h5_files=pc_files,
num_nbrs=num_nbrs,
num_points=num_points)
dl = DataLoader(ds, bs, shuffle=False, num_workers=num_workers)
files_list = []
dl_iter = iter(dl)
num_batches = len(dl.batch_sampler)
print(
f'function name={inspect.stack()[0][3]}, dest_dir={dest_dir}, num files={num_batches}'
)
for batch_idx in range(num_batches):
print(f'Working on file {batch_idx}', flush=True)
normals = next(dl_iter)
print(f'normals.shape={normals.shape}')
file_name = f'data/{dest_dir}/normals_data_{name}{batch_idx}.h5'
files_list.append(file_name)
with h5py.File('../' + file_name, 'w') as hf:
hf.create_dataset("normals", data=normals)
txt_name = f'../data/{dest_dir}/normals_{name}_files.txt'
with open(txt_name, 'w') as f:
f.write('\n'.join(files_list))
return
def create_eigen_dataset(train=True, num_eigen=100, num_nbrs=5):
if train:
name = 'train'
else:
name = 'test'
bs = 2048
pc_files = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/{name}_files.txt')
ds = CreateEigenLBO(pc_files, num_eigen=num_eigen, num_nbrs=num_nbrs)
dl = DataLoader(ds, bs, shuffle=False, num_workers=20)
dl_iter = iter(dl)
num_batches = len(dl.batch_sampler)
for batch_idx in range(num_batches):
print(f'Working on file {batch_idx}', flush=True)
eigval, eigvec, label = next(dl_iter)
print(f'eigval.shape={eigval.shape}')
print(f'eigvec.shape={eigvec.shape}')
print(f'label.shape={label.shape}')
with h5py.File(f'../data/eigen/eigen_data_{name}{batch_idx}.h5',
'w') as hf:
hf.create_dataset("eigval", data=eigval)
hf.create_dataset("eigvec", data=eigvec)
hf.create_dataset("label", data=label)
return
def time_check():
lp = LineProfiler()
lp.add_function(create_distance_mat) # add additional function to profile
lp_wrapper = lp(get_item)
num_points = 1024
tot_examples = 0
examples = []
train = 'train'
h5_files = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/{train}_files.txt')
for h5_file in h5_files:
current_data, current_label = load_h5(h5_file)
current_data = current_data[:, 0:num_points, :]
for i in range(current_data.shape[0]):
examples.append((current_data[i, :, :], current_label[i, :]))
tot_examples += current_data.shape[0]
sd, label = lp_wrapper(examples=examples, index=83)
lp.print_stats()
return
def train_net(expr_name: str, num_eigen: int, num_nbrs: int):
num_points = 1024
bs_train, bs_test = 32, 32
train_files_pc = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/train_files.txt')
test_files_pc = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/test_files.txt')
# train_files_pc = get_files_list(f'../data/normals_1024/normals_train_files.txt')
# test_files_pc = get_files_list(f'../data/normals_1024/normals_test_files.txt')
train_files_lbo = get_files_list(
f'../data/lbo_eig_{num_points}/spectral_train_files.txt')
test_files_lbo = get_files_list(
f'../data/lbo_eig_{num_points}/spectral_test_files.txt')
train_files_gdd = get_files_list(
f'../data/gdd_{num_points}/gdd_train_files.txt')
test_files_gdd = get_files_list(
f'../data/gdd_{num_points}/gdd_test_files.txt')
ds_train = SfmDs(lbo_files=train_files_lbo,
pc_files=train_files_pc,
num_points=num_points,
num_eigen=num_eigen,
num_nbrs=num_nbrs,
gdd_files=train_files_gdd)
ds_test = SfmDs(lbo_files=test_files_lbo,
pc_files=test_files_pc,
num_points=num_points,
num_eigen=num_eigen,
num_nbrs=num_nbrs,
gdd_files=test_files_gdd)
dl_train = DataLoader(ds_train, bs_train, shuffle=True, num_workers=0)
dl_test = DataLoader(ds_test, bs_test, shuffle=True, num_workers=0)
lr = 5e-4
min_lr = 1e-5
l2_reg = 0
model = SfmModel(num_eigen=num_eigen, num_nbs=num_nbrs)
loss_fn = F.cross_entropy
optimizer = torch.optim.Adam(model.parameters(),
lr=lr,
weight_decay=l2_reg)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=20,
gamma=0.7)
trainer = SfmNetTrainer(model,
loss_fn,
optimizer,
scheduler,
min_lr=min_lr,
exp_name=expr_name)
if os.path.isfile(f'results/{expr_name}.pt'):
os.remove(f'results/{expr_name}.pt')
_ = trainer.fit(dl_train,
dl_test,
num_epochs=400,
early_stopping=50,
checkpoints=expr_name)
return
print(f'eigen_val.shape={eigen_val.shape}')
file_name = f'data/{dest_dir}/spectral_data_{name}{batch_idx}.h5'
files_list.append(file_name)
with h5py.File('../' + file_name, 'w') as hf:
hf.create_dataset("eigen_vec", data=eigen_vec)
hf.create_dataset("eigen_val", data=eigen_val)
txt_name = f'../data/{dest_dir}/spectral_{name}_files.txt'
with open(txt_name, 'w') as f:
f.write('\n'.join(files_list))
return
if __name__ == '__main__':
name = 'train'
pc_files = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/{name}_files.txt')
num_nbrs = 5
num_eigen = 32
num_points = 2048
ds = CreateSpectralDist(h5_files=pc_files,
num_nbrs=num_nbrs,
num_eigen=num_eigen,
num_points=num_points)
for index in range(5):
pc, label = ds.get_pc(index)
pc = pc[0:num_points, :]
src, target, wt = ds.get_knn(pc)
src, target, wt = ds.connect_graph(pc, src, target, wt)
L, D = ds.create_laplacian(src, target, wt)
dist_mat = distance_mat.get_distance_m(num_points, src, target, wt)
eigen_val_lbo, eigen_vec_lbo = eigh(a=L, b=D, eigvals=(1, num_eigen))
sign1 = np.random.choice([-1, 1], size=self.size)
sign2 = np.random.choice([-1, 1], size=self.size)
sign1 = np.reshape(sign1, (self.size, 1))
sign2 = np.reshape(sign2, (self.size, 1))
mat = np.multiply(sign1.transpose(), mat)
mat = np.multiply(mat, sign2)
return mat
def add_noise(self, mat):
noise = np.random.normal(1, 0.0001, (self.size, self.size))
return mat * noise
matrix_size = 32
bs_train, bs_test = 32, 32
train_files_pc = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/train_files.txt')
test_files_pc = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/test_files.txt')
# train_files_lbo = get_files_list('../data/lbo_eig_2048/spectral_train_files.txt')
# test_files_lbo = get_files_list('../data/lbo_eig_2048/spectral_test_files.txt')
#
# train_files_dist = get_files_list('../data/dist_eig_2048/spectral_train_files.txt')
# test_files_dist = get_files_list('../data/dist_eig_2048/spectral_test_files.txt')
# ds_train = Spectral40Ds(train_files, size=matrix_size)
# ds_test = Spectral40Ds(test_files, size=matrix_size)
# ds_train = HybMat40Ds(lbo_files=train_files_lbo, dist_files=train_files_dist, pc_files=train_files_pc, size=matrix_size)
ds_train = training.PicNet40Ds(train_files_pc, num_slices=1)
def train_spectral_net(matrix_size):
bs_train, bs_test = 32, 32
train_files_pc = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/train_files.txt')
test_files_pc = get_files_list(
f'../data/modelnet40_ply_hdf5_2048/test_files.txt')
train_files_lbo = get_files_list(
'../data/lbo_eig_2048/spectral_train_files.txt')
test_files_lbo = get_files_list(
'../data/lbo_eig_2048/spectral_test_files.txt')
train_files_dist = get_files_list(
'../data/dist_eig_2048/spectral_train_files.txt')
test_files_dist = get_files_list(
'../data/dist_eig_2048/spectral_test_files.txt')
# ds_train = Spectral40Ds(train_files, size=matrix_size)
# ds_test = Spectral40Ds(test_files, size=matrix_size)
ds_train = HybMat40Ds(lbo_files=train_files_lbo,
dist_files=train_files_dist,
pc_files=train_files_pc,
size=matrix_size)
ds_test = HybMat40Ds(lbo_files=test_files_lbo,
dist_files=test_files_dist,
pc_files=test_files_pc,
size=matrix_size)
# ds_train = SpectralWithEig40Ds(train_files, size=matrix_size)
# ds_test = SpectralWithEig40Ds(test_files, size=matrix_size)
dl_train = DataLoader(ds_train, bs_train, shuffle=True, num_workers=4)
dl_test = DataLoader(ds_test, bs_test, shuffle=True, num_workers=4)
lr = 1e-4
min_lr = 5e-6
l2_reg = 0
our_model = ResNet(BasicBlock, [2, 2, 2, 2])
# our_model = SpectralSimpleNet(mat_size=matrix_size)
# our_model = SpectralSimpleNetBn(mat_size=matrix_size)
# our_model = SpectralFcNet(c_in=matrix_size)
# our_model = SpectralConv1dNet(c_in=matrix_size)
loss_fn = F.cross_entropy
optimizer = torch.optim.Adam(our_model.parameters(),
lr=lr,
weight_decay=l2_reg)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=100,
gamma=0.9)
trainer = NetTrainer(our_model,
loss_fn,
optimizer,
scheduler,
min_lr=min_lr)
# trainer = EigTrainer(our_model, loss_fn, optimizer, scheduler, min_lr=min_lr)
expr_name = f'Spectral-resnet-brs-lr{lr}-noise'
if os.path.isfile(f'results/{expr_name}.pt'):
os.remove(f'results/{expr_name}.pt')
_ = trainer.fit(dl_train,
dl_test,
num_epochs=10000,
early_stopping=50,
checkpoints=expr_name)
return