def test_DGCCA_methods_cpu(self):
latent_dims = 2
device = 'cpu'
encoder_1 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
encoder_3 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
# DTCCA
dtcca_model = DTCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2])
# hidden_layer_sizes are shown explicitly but these are also the defaults
dtcca_model = DeepWrapper(dtcca_model, device=device)
dtcca_model.fit((self.X, self.Y, self.Z))
# DGCCA
dgcca_model = DCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2, encoder_3],
objective=objectives.GCCA)
# hidden_layer_sizes are shown explicitly but these are also the defaults
dgcca_model = DeepWrapper(dgcca_model, device=device)
dgcca_model.fit((self.X, self.Y, self.Z))
# DMCCA
dmcca_model = DCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2, encoder_3],
objective=objectives.MCCA)
# hidden_layer_sizes are shown explicitly but these are also the defaults
dmcca_model = DeepWrapper(dmcca_model, device=device)
dmcca_model.fit((self.X, self.Y, self.Z))
def test_linear():
encoder_1 = architectures.LinearEncoder(latent_dims=1, feature_size=10)
encoder_2 = architectures.LinearEncoder(latent_dims=1, feature_size=12)
dcca = DCCA(latent_dims=1, encoders=[encoder_1, encoder_2])
optimizer = optim.Adam(dcca.parameters(), lr=1e-1)
dcca = CCALightning(dcca, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=50, enable_checkpointing=False)
trainer.fit(dcca, loader)
cca = CCA().fit((X, Y))
# check linear encoder with SGD matches vanilla linear CCA
assert (np.testing.assert_array_almost_equal(
cca.score((X, Y)), trainer.model.score(loader), decimal=2) is None)
def test_input_types(self):
latent_dims = 2
device = 'cpu'
encoder_1 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
# DCCA
dcca_model = DCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.CCA)
# hidden_layer_sizes are shown explicitly but these are also the defaults
dcca_model = DeepWrapper(dcca_model, device=device)
dcca_model.fit(self.train_dataset, epochs=3)
dcca_model.fit(self.train_dataset,
val_dataset=self.train_dataset,
epochs=3)
dcca_model.fit((self.X, self.Y),
val_dataset=(self.X, self.Y),
epochs=3)
def test_schedulers(self):
latent_dims = 2
device = 'cpu'
encoder_1 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims,
feature_size=10)
# DCCA
optimizers = [
optim.Adam(encoder_1.parameters(), lr=1e-4),
optim.Adam(encoder_2.parameters(), lr=1e-4)
]
schedulers = [
optim.lr_scheduler.CosineAnnealingLR(optimizers[0], 1),
optim.lr_scheduler.ReduceLROnPlateau(optimizers[1])
]
dcca_model = DCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.CCA,
optimizers=optimizers,
schedulers=schedulers)
# hidden_layer_sizes are shown explicitly but these are also the defaults
dcca_model = DeepWrapper(dcca_model, device=device)
dcca_model.fit((self.X, self.Y), epochs=20)
train_dataset = Subset(train_dataset, np.arange(n_train))
train_loader, val_loader = get_dataloaders(train_dataset,
val_dataset,
batch_size=128)
# The number of latent dimensions across models
latent_dims = 2
# number of epochs for deep models
epochs = 50
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)
# %%
# Deep CCA
dcca = DCCA(latent_dims=latent_dims, encoders=[encoder_1, encoder_2])
dcca = CCALightning(dcca)
trainer = pl.Trainer(
max_epochs=epochs,
enable_checkpointing=False,
log_every_n_steps=1,
flush_logs_every_n_steps=1,
)
trainer.fit(dcca, train_loader, val_loader)
plot_latent_label(dcca.model, train_loader)
plt.suptitle("DCCA")
plt.show()
# %%
# Deep CCA by Non-Linear Orthogonal Iterations
dcca_noi = DCCA_NOI(latent_dims=latent_dims,
from multiviewdata.torchdatasets import SplitMNISTDataset
from cca_zoo.deepmodels import DCCA, CCALightning, get_dataloaders, architectures
n_train = 500
n_val = 100
train_dataset = SplitMNISTDataset(root="",
mnist_type="MNIST",
train=True,
download=True)
val_dataset = Subset(train_dataset, np.arange(n_train, n_train + n_val))
train_dataset = Subset(train_dataset, np.arange(n_train))
train_loader, val_loader = get_dataloaders(train_dataset, val_dataset)
# The number of latent dimensions across models
latent_dims = 2
# number of epochs for deep models
epochs = 10
# TODO add in custom architecture and schedulers and stuff to show it off
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)
# Deep CCA
dcca = DCCA(latent_dims=latent_dims, encoders=[encoder_1, encoder_2])
optimizer = optim.Adam(dcca.parameters(), lr=1e-3)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, 1)
dcca = CCALightning(dcca, optimizer=optimizer, lr_scheduler=scheduler)
trainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)
trainer.fit(dcca, train_loader, val_loader)
val_dataset = Subset(train_dataset, np.arange(n_train, n_train + n_val))
train_dataset = Subset(train_dataset, np.arange(n_train))
train_loader, val_loader = get_dataloaders(train_dataset, val_dataset)
# The number of latent dimensions across models
latent_dims = 2
# number of epochs for deep models
epochs = 10
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)
# %%
# Deep MCCA
dcca = DCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.MCCA)
dcca = CCALightning(dcca)
trainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)
trainer.fit(dcca, train_loader, val_loader)
# %%
# Deep GCCA
dcca = DCCA(latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.GCCA)
dcca = CCALightning(dcca)
trainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)
trainer.fit(dcca, train_loader, val_loader)
# %%
def test_DCCA_methods():
N = len(train_dataset)
latent_dims = 2
epochs = 100
cca = CCA(latent_dims=latent_dims).fit((X, Y))
# DCCA_NOI
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=12)
dcca_noi = DCCA_NOI(latent_dims, N, encoders=[encoder_1, encoder_2], rho=0)
optimizer = optim.Adam(dcca_noi.parameters(), lr=1e-2)
dcca_noi = CCALightning(dcca_noi, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=epochs,
log_every_n_steps=10,
enable_checkpointing=False)
trainer.fit(dcca_noi, train_loader)
assert (np.testing.assert_array_less(
cca.score((X, Y)).sum(),
trainer.model.score(train_loader).sum()) is None)
# Soft Decorrelation (stochastic Decorrelation Loss)
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=12)
sdl = DCCA_SDL(latent_dims, N, encoders=[encoder_1, encoder_2], lam=1e-3)
optimizer = optim.SGD(sdl.parameters(), lr=1e-1)
sdl = CCALightning(sdl, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=epochs, log_every_n_steps=10)
trainer.fit(sdl, train_loader)
assert (np.testing.assert_array_less(
cca.score((X, Y)).sum(),
trainer.model.score(train_loader).sum()) is None)
# DCCA
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=12)
dcca = DCCA(
latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.CCA,
)
optimizer = optim.SGD(dcca.parameters(), lr=1e-1)
dcca = CCALightning(dcca, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=epochs,
log_every_n_steps=10,
enable_checkpointing=False)
trainer.fit(dcca, train_loader, val_dataloaders=val_loader)
assert (np.testing.assert_array_less(
cca.score((X, Y)).sum(),
trainer.model.score(train_loader).sum()) is None)
# DGCCA
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=12)
dgcca = DCCA(
latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.GCCA,
)
optimizer = optim.SGD(dgcca.parameters(), lr=1e-2)
dgcca = CCALightning(dgcca, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=epochs,
log_every_n_steps=10,
enable_checkpointing=False)
trainer.fit(dgcca, train_loader)
assert (np.testing.assert_array_less(
cca.score((X, Y)).sum(),
trainer.model.score(train_loader).sum()) is None)
# DMCCA
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=12)
dmcca = DCCA(
latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
objective=objectives.MCCA,
)
optimizer = optim.SGD(dmcca.parameters(), lr=1e-2)
dmcca = CCALightning(dmcca, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=epochs,
log_every_n_steps=10,
enable_checkpointing=False)
trainer.fit(dmcca, train_loader)
assert (np.testing.assert_array_less(
cca.score((X, Y)).sum(),
trainer.model.score(train_loader).sum()) is None)
# Barlow Twins
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=10)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=12)
barlowtwins = BarlowTwins(
latent_dims=latent_dims,
encoders=[encoder_1, encoder_2],
)
optimizer = optim.SGD(barlowtwins.parameters(), lr=1e-2)
barlowtwins = CCALightning(barlowtwins, optimizer=optimizer)
trainer = pl.Trainer(max_epochs=epochs,
log_every_n_steps=10,
enable_checkpointing=False)
trainer.fit(barlowtwins, train_loader)
assert (np.testing.assert_array_less(
cca.score((X, Y)).sum(),
trainer.model.score(train_loader).sum()) is None)
### Deep Learning
We also have deep CCA methods (and autoencoder variants)
- Deep CCA (DCCA)
- Deep Canonically Correlated Autoencoders (DCCAE)
We introduce a Config class from configuration.py. This contains a number of default settings for running DCCA.
"""
from cca_zoo.deepmodels import DCCA, DCCAE, DVCCA, DCCA_NOI, DeepWrapper, architectures, objectives
# %%
# DCCA
print('DCCA')
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=784)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=784)
dcca_model = DCCA(latent_dims=latent_dims, encoders=[encoder_1, encoder_2])
# hidden_layer_sizes are shown explicitly but these are also the defaults
dcca_model = DeepWrapper(dcca_model)
dcca_model.fit(train_view_1, train_view_2, epochs=epochs)
dcca_results = np.stack((dcca_model.train_correlations[0, 1],
dcca_model.predict_corr(test_view_1, test_view_2)[0,
1]))
# DGCCA
print('DGCCA')
encoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=784)
encoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=784)
dgcca_model = DCCA(latent_dims=latent_dims,