Deep-Unsupervised-Domain-Adaptation

Pytorch implementation of four neural network based domain adaptation techniques: DeepCORAL, DDC, CDAN and CDAN+E. Evaluated on benchmark dataset Office31.

Paper: Evaluation of Deep Neural Network Domain Adaptation Techniques for Image Recognition

Abstract

It has been well proved that deep networks are efficient at extracting features from a given (source) labeled dataset. However, it is not always the case that they can generalize well to other (target) datasets which very often have a different underlying distribution. In this report, we evaluate four different domain adaptation techniques for image classification tasks: Deep CORAL, Deep Domain Confusion (DDC), Conditional Adversarial Domain Adaptation (CDAN) and CDAN with Entropy Conditioning (CDAN+E). The selected domain adaptation techniques are unsupervised techniques where the target dataset will not carry any labels during training phase. The experiments are conducted on the office-31 dataset.

Results

Accuracy performance on the Office31 dataset for the source and domain data distributions (with and without transfer losses).

Deep CORAL	DDC

CDAN	CDAN+E

Target accuracies for all six domain shifts in Office31 dataset (amazon, webcam and dslr)

Method	A → W	A → D	W → A	W → D	D → A	D → W
No Adaptaion	43.1 ± 2.5	49.2 ± 3.7	35.6 ± 0.6	94.2 ± 3.1	35.4 ± 0.7	90.9 ± 2.4
DeepCORAL	49.5 ± 2.7	40.0 ± 3.3	38.3 ± 0.4	74.4 ± 4.3	38.5 ± 1.5	89.1 ± 4.4
DDC	41.7 ± 9.1	---	---	---	---	---
CDAN	44.9 ± 3.3	49.5 ± 4.6	34.8 ± 2.4	93.3 ± 3.4	32.9 ± 3.4	88.3 ± 3.8
CDAN+E	48.7 ± 7.5	53.7 ± 4.7	35.3 ± 2.7	93.6 ± 3.4	33.9 ± 2.2	87.7 ± 4.0

Training and inference

To train the model in your computer you must download the Office31 dataset and put it in your data folder.

Execute training of a method by going to its folder (e.g. DeepCORAL):

cd DeepCORAL/
python main.py --epochs 100 --batch_size_source 128 --batch_size_target 128 --name_source amazon --name_target webcam

Loss and accuracy plots

Once the model is trained, you can generate plots like the ones shown above by running:

cd DeepCORAL/
python plot_loss_acc.py --source amazon --target webcam --no_epochs 10

The following is a list of the arguments the usuer can provide:

--epochs number of training epochs
--batch_size_source batch size of source data
--batch_size_target batch size of target data
--name_source name of source dataset
--name_target name of source dataset
--num_classes no. classes in dataset
--load_model flag to load pretrained model (AlexNet by default)
--adapt_domain bool argument to train with or without specific transfer loss

Requirements

tqdm
PyTorch
matplotlib
numpy
pickle
scikit-image
torchvision

References

DeepCORAL paper
DDC paper
CDAN paper

hyacinthschatten / deep-unsupervised-domain-adaptation Goto Github PK