Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Random seed initialization moved to NMTF super class init #16

Merged
merged 1 commit into from
Nov 28, 2023
Merged

Random seed initialization moved to NMTF super class init #16

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 10 additions & 9 deletions adnmtf/nmtf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,7 @@ def __init__(
self.leverage = leverage
self.random_state = random_state
self.verbose = verbose
np.random.seed(random_state)

def fit_transform(
self,
Expand Down Expand Up @@ -260,9 +261,9 @@ def fit_transform(
nmf_calculate_leverage = 0
nmf_use_robust_leverage = 0

if self.random_state is not None:
random_seed = self.random_state
np.random.seed(random_seed)
# if self.random_state is not None:
# random_seed = self.random_state
# np.random.seed(random_seed)

_, mt, mw, mb, mt_pct, mw_pct, diff, add_message, err_message, cancel_pressed = r_ntf_solve(
m=m,
Expand Down Expand Up @@ -516,9 +517,9 @@ def fit_transform(
ntf_left_components = self.apply_left
ntf_right_components = self.apply_right
ntf_block_components = self.apply_block
if self.random_state is not None:
random_seed = self.random_state
np.random.seed(random_seed)
# if self.random_state is not None:
# random_seed = self.random_state
# np.random.seed(random_seed)

my_status_box = get_status_box()(verbose=log_iter)

Expand Down Expand Up @@ -600,9 +601,9 @@ def fit_transform(
nmf_calculate_leverage = 0
nmf_use_robust_leverage = 0

if self.random_state is not None:
random_seed = self.random_state
np.random.seed(random_seed)
# if self.random_state is not None:
# random_seed = self.random_state
# np.random.seed(random_seed)

if update_w:
nmf_fix_user_lhe = 0
Expand Down
68 changes: 4 additions & 64 deletions adnmtf/nmtf_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,23 +68,18 @@ def nmf_init(m, mmis, mt0, mw0, nc) -> Tuple[np.ndarray, np.ndarray]:
mw = np.copy(mw0)
if (mt.shape[0] == 0) or (mw.shape[0] == 0):
# Note that if there are missing values, SVD is performed on matrix imputed with 0's
np.random.seed(3)
if nc >= min(n, p):
# arpack does not accept to factorize at full rank -> need to duplicate in both dimensions to force it work
# noinspection PyTypeChecker
t, d, w = svds(
np.concatenate((np.concatenate((m, m), axis=1), np.concatenate((m, m), axis=1)), axis=0),
k=nc,
v0=np.random.uniform(size=2 * min(n, p)),
random_state=0
)
d /= 2
# svd causes mem allocation problem with large matrices
# t, d, w = np.linalg.svd(m)
# mt = t
# mw = w.T
else:
t, d, w = svds(m, k=nc, v0=np.random.uniform(size=min(n, p)))
# t, d, w = np.linalg.svd(m)
t, d, w = svds(m, k=nc, v0=np.random.uniform(size=min(n, p)), random_state=0)

mt = t[:n, :nc]
mw = w[:nc, :p].T
Expand Down Expand Up @@ -124,9 +119,6 @@ def nmf_init(m, mmis, mt0, mw0, nc) -> Tuple[np.ndarray, np.ndarray]:
mw = mw * ((mt.T @ m) / ((mt.T @ mt) @ mw.T + precision)).T
mt = mt * (m @ mw / (mt @ (mw.T @ mw) + precision))

# np.savetxt("C:/Users/paul_/PycharmProjects/nmtf_private/tests/data/datatest_W.csv", mt)
# np.savetxt("C:/Users/paul_/PycharmProjects/nmtf_private/tests/data/datatest_H.csv", mw)

return mt, mw


Expand All @@ -139,40 +131,14 @@ def init_ntf_type_1(m, mmis, n_blocks, nc, mt_nmf, mw_nmf, tolerance, log_iter,
else:
mt_nmf, mw_nmf = nmf_init(m=mstacked, mmis=mmis_stacked, mt0=mt_nmf, mw0=mw_nmf, nc=nc2)

# Quick NMF (canceled since multiplicative warm-up added in nmf_init)
# _, mt_nmf, mw_nmf, mb, diff, cancel_pressed = ntf_solve(
# m=mstacked,
# mmis=mmis_stacked,
# mt0=mt_nmf,
# mw0=mw_nmf,
# mb0=np.array([]),
# nc=nc2,
# tolerance=tolerance,
# log_iter=log_iter,
# status0=status0,
# max_iterations=10,
# nmf_fix_user_lhe=0,
# nmf_fix_user_rhe=0,
# nmf_fix_user_bhe=1,
# nmf_sparse_level=0,
# ntf_unimodal=0,
# ntf_smooth=0,
# ntf_left_components=0,
# ntf_right_components=0,
# ntf_block_components=0,
# n_blocks=1,
# nmf_priors=np.array([]),
# my_status_box=my_status_box,
# )

# Factorize Left vectors and distribute multiple factors if nc2 < nc
mt = np.zeros((n, nc))
mw = np.zeros((int(p / n_blocks), nc))
mb = np.zeros((n_blocks, nc))
n_fact = int(np.ceil(nc / n_blocks))
for k in range(0, nc2):
my_status_box.update_status(status="Start SVD...")
u, d, v = svds(np.reshape(mt_nmf[:, k], (int(p / n_blocks), n)).T, k=n_fact)
u, d, v = svds(np.reshape(mt_nmf[:, k], (int(p / n_blocks), n)).T, k=n_fact, random_state=0)
v = v.T
# svds returns singular vectors in reverse order
u = u[:, ::-1]
Expand Down Expand Up @@ -228,32 +194,6 @@ def init_ntf_type_2(
else:
mt_nmf, mw_nmf = nmf_init(m=m, mmis=mmis, mt0=mt_nmf, mw0=mw_nmf, nc=nc)

# Quick NMF (canceled since multiplicative warm-up added in nmf_init)
# _, mt_nmf, mw_nmf, mb, diff, cancel_pressed = ntf_solve(
# m=m,
# mmis=mmis,
# mt0=mt_nmf,
# mw0=mw_nmf,
# mb0=np.array([]),
# nc=nc,
# tolerance=tolerance,
# log_iter=log_iter,
# status0=status0,
# max_iterations=10,
# nmf_fix_user_lhe=0,
# nmf_fix_user_rhe=0,
# nmf_fix_user_bhe=1,
# nmf_sparse_level=0,
# ntf_unimodal=0,
# ntf_smooth=0,
# ntf_left_components=0,
# ntf_right_components=0,
# ntf_block_components=0,
# n_blocks=1,
# nmf_priors=np.array([]),
# my_status_box=my_status_box,
# )

# Factorize Right vectors
mt = np.zeros((n, nc))
mw = np.zeros((int(p / n_blocks), nc))
Expand All @@ -262,7 +202,7 @@ def init_ntf_type_2(
for k in range(0, nc):
my_status_box.update_status(status="Start SVD...")
# noinspection PyTypeChecker
u, d, v = svds(np.reshape(mw_nmf[:, k], (int(p / n_blocks), n_blocks)), k=1)
u, d, v = svds(np.reshape(mw_nmf[:, k], (int(p / n_blocks), n_blocks)), k=1, random_state=0)
v = v.T
u = np.abs(u)
v = np.abs(v)
Expand Down
Loading