Skip to content

Instantiate tensors on correct devices #467

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

Open
jakobnissen wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Instantiate tensors on correct devices #467

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
30 changes: 12 additions & 18 deletions vamb/aamb_encode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,9 +122,6 @@

std = torch.exp(logvar / 2)
sampled_z = Variable(Tensor(self.rng.normal(0, 1, (mu.size(0), self.ld))))

if self.usecuda:
sampled_z = sampled_z.cuda()
z = sampled_z * std + mu

return z
Expand Down Expand Up @@ -248,7 +245,7 @@
# Define adversarial loss for the discriminators
adversarial_loss = torch.nn.BCELoss()
if self.usecuda:
adversarial_loss.cuda()
adversarial_loss = adversarial_loss.cuda()

#### Optimizers
optimizer_E = torch.optim.Adam(enc_params, lr=lr)
Expand Down Expand Up @@ -290,16 +287,15 @@
)

# Sample noise as discriminator Z,Y ground truth
device = "cuda" if self.usecuda else "cpu"

if self.usecuda:
z_prior = torch.cuda.FloatTensor(nrows, self.ld).normal_()
z_prior.cuda()
ohc = RelaxedOneHotCategorical(
torch.tensor([T], device="cuda"),
torch.ones([nrows, self.y_len], device="cuda"),
torch.tensor([T], device=device),
torch.ones([nrows, self.y_len], device=device),
)
y_prior = ohc.sample()
y_prior = y_prior.cuda()

else:
z_prior = Variable(Tensor(self.rng.normal(0, 1, (nrows, self.ld))))
Expand All @@ -308,9 +304,8 @@

del ohc

if self.usecuda:
depths_in = depths_in.cuda()
tnfs_in = tnfs_in.cuda()
depths_in = depths_in.to(device)
tnfs_in = tnfs_in.to(device)

# -----------------
# Train Generator
Expand Down Expand Up @@ -461,26 +456,25 @@
with torch.no_grad():
for depths_in, tnfs_in, _, _ in new_data_loader:
nrows, _ = depths_in.shape
device = "cuda" if self.usecuda else "cpu"

if self.usecuda:
z_prior = torch.cuda.FloatTensor(nrows, self.ld).normal_()
z_prior.cuda()
ohc = RelaxedOneHotCategorical(
torch.tensor([0.15], device="cuda"),
torch.ones([nrows, self.y_len], device="cuda"),
torch.tensor([0.15], device=device),
torch.ones([nrows, self.y_len], device=device),
)
y_prior = ohc.sample()
y_prior = y_prior.cuda()

else:
z_prior = Variable(Tensor(self.rng.normal(0, 1, (nrows, self.ld))))

Check failure on line 470 in vamb/aamb_encode.py

View workflow job for this annotation

GitHub Actions / lint

Ruff (F841) 

vamb/aamb_encode.py:470:21: F841 Local variable `z_prior` is assigned to but never used
ohc = RelaxedOneHotCategorical(
0.15, torch.ones([nrows, self.y_len])
)
y_prior = ohc.sample()

Check failure on line 474 in vamb/aamb_encode.py

View workflow job for this annotation

GitHub Actions / lint

Ruff (F841) 

vamb/aamb_encode.py:474:21: F841 Local variable `y_prior` is assigned to but never used

if self.usecuda:
depths_in = depths_in.cuda()
tnfs_in = tnfs_in.cuda()
depths_in = depths_in.to(device)
tnfs_in = tnfs_in.to(device)

mu, _, _, _, y_sample = self(depths_in, tnfs_in)[0:5]

Expand Down
19 changes: 8 additions & 11 deletions vamb/cluster.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,9 +224,9 @@ def _check_params(
def _init_histogram_kept_mask(self, N: int) -> tuple[_Tensor, _Tensor]:
"N is number of contigs"

kept_mask = _torch.ones(N, dtype=_torch.bool)
if self.cuda:
kept_mask = kept_mask.cuda()
kept_mask = _torch.ones(
N, dtype=_torch.bool, device="cuda" if self.cuda else "cpu"
)

histogram = _torch.empty(_ceil(_XMAX / _DELTA_X))
return histogram, kept_mask
Expand Down Expand Up @@ -258,10 +258,9 @@ def __init__(
_normalize(torch_matrix, inplace=True)

# Move to GPU
torch_lengths = _torch.Tensor(lengths)
if cuda:
torch_matrix = torch_matrix.cuda()
torch_lengths = torch_lengths.cuda()
device = "cuda" if cuda else "cpu"
torch_matrix = torch_matrix.to(device)
torch_lengths = _torch.tensor(lengths, dtype=_torch.float32, device=device)

self.maxsteps: int = maxsteps
self.minsuccesses: int = minsuccesses
Expand All @@ -274,9 +273,7 @@ def __init__(
self.indices = _torch.arange(len(matrix))
self.order = _np.argsort(lengths)[::-1]
self.order_index = 0
self.lengths = _torch.Tensor(lengths)
if cuda:
self.lengths = self.lengths.cuda()
self.lengths = torch_lengths
self.n_emitted_clusters = 0
self.n_remaining_points = len(torch_matrix)
self.peak_valley_ratio = 0.1
Expand Down Expand Up @@ -321,7 +318,7 @@ def pack(self):
cpu_kept_mask = self.kept_mask.cpu()
self.matrix = _vambtools.torch_inplace_maskarray(
self.matrix.cpu(), cpu_kept_mask
).cuda()
).to("cuda")
self.indices = self.indices[cpu_kept_mask]

else:
Expand Down
25 changes: 12 additions & 13 deletions vamb/encode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -274,10 +274,9 @@ def _encode(self, tensor: Tensor) -> Tensor:

# sample with gaussian noise
def reparameterize(self, mu: Tensor) -> Tensor:
epsilon = _torch.randn(mu.size(0), mu.size(1))

if self.usecuda:
epsilon = epsilon.cuda()
epsilon = _torch.randn(
mu.size(0), mu.size(1), device="cuda" if self.usecuda else "cpu"
)

epsilon.requires_grad = True

Expand Down Expand Up @@ -392,11 +391,11 @@ def trainepoch(
tnf_in.requires_grad = True
abundance_in.requires_grad = True

if self.usecuda:
depths_in = depths_in.cuda()
tnf_in = tnf_in.cuda()
abundance_in = abundance_in.cuda()
weights = weights.cuda()
device = "cuda" if self.usecuda else "cpu"
depths_in = depths_in.to(device)
tnf_in = tnf_in.to(device)
abundance_in = abundance_in.to(device)
weights = weights.to(device)

optimizer.zero_grad()

Expand Down Expand Up @@ -465,10 +464,10 @@ def encode(self, data_loader) -> _np.ndarray:
with _torch.no_grad():
for depths, tnf, ab, _ in new_data_loader:
# Move input to GPU if requested
if self.usecuda:
depths = depths.cuda()
tnf = tnf.cuda()
ab = ab.cuda()
device = "cuda" if self.usecuda else "cpu"
depths = depths.to(device)
tnf = tnf.to(device)
ab = ab.to(device)

# Evaluate
_, _, _, mu = self(depths, tnf, ab)
Expand Down
70 changes: 32 additions & 38 deletions vamb/semisupervised_encode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -238,9 +238,7 @@

def forward(self, labels):
mu = self._encode(labels)
logsigma = _torch.zeros(mu.size())
if self.usecuda:
logsigma = logsigma.cuda()
logsigma = _torch.zeros(mu.size(), device="cuda" if self.usecuda else "cpu")
latent = self.reparameterize(mu)
labels_out = self._decode(latent)
return labels_out, mu, logsigma
Expand Down Expand Up @@ -281,8 +279,7 @@
labels_in = labels_in[0]
labels_in.requires_grad = True

if self.usecuda:
labels_in = labels_in.cuda()
labels_in = labels_in.to("cuda" if self.usecuda else "cpu")

optimizer.zero_grad()

Expand Down Expand Up @@ -343,8 +340,7 @@
for labels in new_data_loader:
labels = labels[0]
# Move input to GPU if requested
if self.usecuda:
labels = labels.cuda()
labels = labels.to("cuda" if self.usecuda else "cpu")

# Evaluate
out_labels, mu, logsigma = self(labels)
Expand Down Expand Up @@ -504,9 +500,7 @@
def forward(self, depths, tnf, abundance, labels):
tensor = _torch.cat((depths, tnf, abundance, labels), 1)
mu = self._encode(tensor)
logsigma = _torch.zeros(mu.size())
if self.usecuda:
logsigma = logsigma.cuda()
logsigma = _torch.zeros(mu.size(), device="cuda" if self.usecuda else "cpu")
latent = self.reparameterize(mu)
depths_out, tnf_out, abundance_out, labels_out = self._decode(latent)

Expand Down Expand Up @@ -596,12 +590,12 @@
abundance_in.requires_grad = True
labels_in.requires_grad = True

if self.usecuda:
depths_in = depths_in.cuda()
tnf_in = tnf_in.cuda()
abundance_in = abundance_in.cuda()
weights = weights.cuda()
labels_in = labels_in.cuda()
device = "cuda" if self.usecuda else "cpu"
depths_in = depths_in.to(device)
tnf_in = tnf_in.to(device)
abundance_in = abundance_in.to(device)
weights = weights.to(device)
labels_in = labels_in.to(device)

optimizer.zero_grad()

Expand Down Expand Up @@ -677,11 +671,11 @@
with _torch.no_grad():
for depths, tnf, ab, weights, labels in new_data_loader:
# Move input to GPU if requested
if self.usecuda:
depths = depths.cuda()
tnf = tnf.cuda()
ab = ab.cuda()
labels = labels.cuda()
device = "cuda" if self.usecuda else "cpu"
depths = depths.to(device)
tnf = tnf.to(device)
ab = ab.to(device)
labels = labels.to(device)

# Evaluate
_, _, _, _, mu, _ = self(depths, tnf, ab, labels)
Expand Down Expand Up @@ -882,17 +876,17 @@
abundance_in_unsup.requires_grad = True
labels_in_unsup.requires_grad = True

if self.VAEVamb.usecuda:
depths_in_sup = depths_in_sup.cuda()
tnf_in_sup = tnf_in_sup.cuda()
abundance_in_sup = abundance_in_sup.cuda()
weights_in_sup = weights_in_sup.cuda()
labels_in_sup = labels_in_sup.cuda()
depths_in_unsup = depths_in_unsup.cuda()
tnf_in_unsup = tnf_in_unsup.cuda()
abundance_in_unsup = abundance_in_unsup.cuda()
weights_in_unsup = weights_in_unsup.cuda()
labels_in_unsup = labels_in_unsup.cuda()
device = "cuda" if self.VAEVamb.usecuda else "cpu"
depths_in_sup = depths_in_sup.to(device)
tnf_in_sup = tnf_in_sup.to(device)
abundance_in_sup = abundance_in_sup.to(device)
weights_in_sup = weights_in_sup.to(device)
labels_in_sup = labels_in_sup.to(device)
depths_in_unsup = depths_in_unsup.to(device)
tnf_in_unsup = tnf_in_unsup.to(device)
abundance_in_unsup = abundance_in_unsup.to(device)
weights_in_unsup = weights_in_unsup.to(device)
labels_in_unsup = labels_in_unsup.to(device)

optimizer.zero_grad()

Expand All @@ -913,15 +907,15 @@
abundance_out_unsup,
mu_vamb_unsup,
) = self.VAEVamb(depths_in_unsup, tnf_in_unsup, abundance_in_unsup)
logsigma_vamb_unsup = _torch.zeros(mu_vamb_unsup.size())
if self.usecuda:
logsigma_vamb_unsup = logsigma_vamb_unsup.cuda()
logsigma_vamb_unsup = _torch.zeros(

Check failure on line 910 in vamb/semisupervised_encode.py

View workflow job for this annotation

GitHub Actions / lint

Ruff (F841) 

vamb/semisupervised_encode.py:910:13: F841 Local variable `logsigma_vamb_unsup` is assigned to but never used
mu_vamb_unsup.size(), device="cuda" if self.VAEVamb.usecuda else "cpu"
)
_, _, _, mu_vamb_sup_s = self.VAEVamb(
depths_in_sup, tnf_in_sup, abundance_in_sup
)
logsigma_vamb_sup_s = _torch.zeros(mu_vamb_sup_s.size())
if self.usecuda:
logsigma_vamb_sup_s = logsigma_vamb_sup_s.cuda()
logsigma_vamb_sup_s = _torch.zeros(
mu_vamb_sup_s.size(), device="cuda" if self.VAEVamb.usecuda else "cpu"
)
labels_out_unsup, mu_labels_unsup, logsigma_labels_unsup = self.VAELabels(
labels_in_unsup
)
Expand Down
34 changes: 17 additions & 17 deletions vamb/taxvamb_encode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -900,11 +900,11 @@ def predict(self, data_loader) -> Iterable[tuple[_np.ndarray, _np.ndarray]]:
with _torch.no_grad():
for depths, tnf, abundances, weights in new_data_loader:
# Move input to GPU if requested
if self.usecuda:
depths = depths.cuda()
tnf = tnf.cuda()
abundances = abundances.cuda()
weights = weights.cuda()
device = "cuda" if self.usecuda else "cpu"
depths = depths.to(device)
tnf = tnf.to(device)
abundances = abundances.to(device)
weights = weights.to(device)

# Evaluate
labels = self(depths, tnf, abundances, weights)
Expand All @@ -931,12 +931,12 @@ def predict_with_ground_truth(

with _torch.no_grad():
for depths_in, tnf_in, abundances_in, weights, labels_in in new_data_loader:
if self.usecuda:
depths_in = depths_in.cuda()
tnf_in = tnf_in.cuda()
abundances_in = abundances_in.cuda()
weights = weights.cuda()
labels_in = labels_in.cuda()
device = "cuda" if self.usecuda else "cpu"
depths_in = depths_in.to(device)
tnf_in = tnf_in.to(device)
abundances_in = abundances_in.to(device)
weights = weights.to(device)
labels_in = labels_in.to(device)

labels_out = self(depths_in, tnf_in, abundances_in, weights)
loss, correct_labels = self.calc_loss(labels_in, labels_out)
Expand Down Expand Up @@ -1004,12 +1004,12 @@ def trainepoch(
abundances_in.requires_grad = True
labels_in.requires_grad = True

if self.usecuda:
depths_in = depths_in.cuda()
tnf_in = tnf_in.cuda()
abundances_in = abundances_in.cuda()
weights = weights.cuda()
labels_in = labels_in.cuda()
device = "cuda" if self.usecuda else "cpu"
depths_in = depths_in.to(device)
tnf_in = tnf_in.to(device)
abundances_in = abundances_in.to(device)
weights = weights.to(device)
labels_in = labels_in.to(device)

optimizer.zero_grad()

Expand Down
Loading