Domain Adaptation

Dependencies

There are 2 requirements.txt, one for Ubuntu and another for Windows.

Use as follows: pip intall -r requirements.txt

Pretrained weights

In the save/ directory, I have pretrained weights (se_weights_epoch_0.pth) that achieve :

• Source accuracy: 98 %
• Target accuracy: 97 %
• F1 target: 1.000
• Precision target: 1.000
• Recall target: 1.000

Python script

Usage

usage: domain_adaptation.py [-h] [--resume RESUME] [--data DATA] [--save SAVE]
                            [--eval EVAL] [--batch_s BATCH_S]
                            [--batch_t BATCH_T] [--disp]

SVHN to MNIST domain adaptation

optional arguments:
  -h, --help         show this help message and exit
  --resume RESUME    Weights path to resume from
  --data DATA        Datasets path
  --save SAVE        Result path
  --eval EVAL        Evaluate model
  --batch_s BATCH_S  Source domain batch size
  --batch_t BATCH_T  Target domain batch size
  --disp             Display predictions during eval

Evaluate

python domain_adaptation.py --resume=./save/se_weights_epoch_0.pth --eval 100 --batch_s 10 --batch_t 10

Evaluate and display samples

python domain_adaptation.py --resume=./save/se_weights_epoch_0.pth --eval 1 --disp --batch_s 10 --batch_t 10

Close the matplotlib windows (alt-f4 or close button) to get new batch prediction. Ctrl-C to stop.

Train

python domain_adaptation.py --resume=./save/se_weights_epoch_0.pth --save=./weights/ --batch_s 100  --batch_t 1000

Batch sizes have to be big according to the paper, to keep a good distribution of labels/samples (e.g. 100 per class)

Train from scratch

python domain_adaptation.py --save=./weights/ --batch_s 100  --batch_t 1000