Divergence domain adaptation methods.

This example illustrates the DeepCoral method from [1] on a simple image classification task.

[1]

Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV Workshops, 2016.

# Author: Théo Gnassounou
#
# License: BSD 3-Clause
# sphinx_gallery_thumbnail_number = 4

from skorch import NeuralNetClassifier
from torch import nn

from skada.datasets import load_mnist_usps
from skada.deep import DeepCoral
from skada.deep.modules import MNISTtoUSPSNet

Load the image datasets

dataset = load_mnist_usps(n_classes=2, n_samples=0.5, return_dataset=True)
X, y, sample_domain = dataset.pack(
    as_sources=["mnist"], as_targets=["usps"], mask_target_labels=True
)
X_test, y_test, sample_domain_test = dataset.pack(
    as_sources=[], as_targets=["usps"], mask_target_labels=False
)

  0%|          | 0.00/9.91M [00:00<?, ?B/s]
100%|██████████| 9.91M/9.91M [00:00<00:00, 137MB/s]

  0%|          | 0.00/28.9k [00:00<?, ?B/s]
100%|██████████| 28.9k/28.9k [00:00<00:00, 43.3MB/s]

  0%|          | 0.00/1.65M [00:00<?, ?B/s]
100%|██████████| 1.65M/1.65M [00:00<00:00, 206MB/s]

  0%|          | 0.00/4.54k [00:00<?, ?B/s]
100%|██████████| 4.54k/4.54k [00:00<00:00, 16.4MB/s]
/home/circleci/project/skada/datasets/_mnist_usps.py:72: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.detach().clone() or sourceTensor.detach().clone().requires_grad_(True), rather than torch.tensor(sourceTensor).
  mnist_target = torch.tensor(mnist_dataset.targets)

  0%|          | 0.00/1.83M [00:00<?, ?B/s]
  2%|▏         | 32.8k/1.83M [00:00<00:23, 75.1kB/s]
  4%|▎         | 65.5k/1.83M [00:00<00:16, 105kB/s]
  7%|▋         | 131k/1.83M [00:00<00:09, 173kB/s]
 14%|█▍        | 262k/1.83M [00:01<00:05, 313kB/s]
 29%|██▊       | 524k/1.83M [00:01<00:02, 591kB/s]
 55%|█████▌    | 1.02M/1.83M [00:01<00:00, 1.10MB/s]
100%|██████████| 1.83M/1.83M [00:01<00:00, 1.16MB/s]

Train a classic model

model = NeuralNetClassifier(
    MNISTtoUSPSNet(),
    criterion=nn.CrossEntropyLoss(),
    batch_size=128,
    max_epochs=5,
    train_split=False,
    lr=1e-2,
)
model.fit(X[sample_domain > 0], y[sample_domain > 0])
model.score(X_test, y=y_test)

  epoch    train_loss      dur
-------  ------------  -------
      1        1.3530  12.9001
      2        0.1990  10.1447
      3        0.0838  8.7053
      4        0.0470  12.6023
      5        0.0345  9.0915

0.9228295819935691

Train a DeepCoral model

model = DeepCoral(
    MNISTtoUSPSNet(),
    layer_name="fc1",
    batch_size=128,
    max_epochs=5,
    train_split=False,
    reg=1,
    lr=1e-2,
)
model.fit(X, y, sample_domain=sample_domain)
model.score(X_test, y_test, sample_domain=sample_domain_test)

  epoch    train_loss      dur
-------  ------------  -------
      1        1.5227  11.5933
      2        0.3119  11.2990
      3        0.1370  13.2985
      4        0.0785  12.7983
      5        0.0707  10.6999

0.9614147909967846