da_setting adda seems no effect

[xiaoyuan1996]

python train_DA.py -d digits --source svhn --target mnist --da_setting adda

Training target model... Train Epoch: 22, Batch:937, S: [c: 0.000, d: 1.354], Time: 12.886 Testing target model on [source]... Test, loss: 4.267, prec: 6.4%, Time: 2.465 Testing target model on [target]... Test, loss: 4.293, prec: 9.8%, Time: 1.192 73%|████████████████████████████████████████████████████████████████████████████████████▎ | 22/30 [06:11<02:15, 16.93s/it]Training target model... Train Epoch: 23, Batch:937, S: [c: 0.000, d: 1.429], Time: 12.536 Testing target model on [source]... Test, loss: 6.942, prec: 7.8%, Time: 2.410 Testing target model on [target]... Test, loss: 7.577, prec: 10.3%, Time: 1.235 77%|████████████████████████████████████████████████████████████████████████████████████████▏ | 23/30 [06:27<01:57, 16.75s/it]Training target model... Train Epoch: 24, Batch:937, S: [c: 0.000, d: 1.375], Time: 13.038 Testing target model on [source]... Test, loss: 14.831, prec: 9.7%, Time: 2.508 Testing target model on [target]... Test, loss: 15.324, prec: 9.8%, Time: 1.543 80%|████████████████████████████████████████████████████████████████████████████████████████████ | 24/30 [06:44<01:41, 16.91s/it]Training target model... Train Epoch: 25, Batch:937, S: [c: 0.000, d: 1.367], Time: 12.633 Testing target model on [source]... Test, loss: 6.635, prec: 6.1%, Time: 2.474 Testing target model on [target]... Test, loss: 7.577, prec: 10.1%, Time: 1.280 83%|███████████████████████████████████████████████████████████████████████████████████████████████▊ | 25/30 [07:01<01:23, 16.80s/it]Training target model... Train Epoch: 26, Batch:937, S: [c: 0.000, d: 1.390], Time: 12.824 Testing target model on [source]... Test, loss: 9.628, prec: 15.9%, Time: 2.390 Testing target model on [target]... Test, loss: 10.641, prec: 10.3%, Time: 1.202 87%|███████████████████████████████████████████████████████████████████████████████████████████████████▋ | 26/30 [07:17<01:06, 16.73s/it]Training target model... Train Epoch: 27, Batch:937, S: [c: 0.000, d: 1.447], Time: 12.648 Testing target model on [source]... Test, loss: 8.766, prec: 11.1%, Time: 3.055 Testing target model on [target]... Test, loss: 8.690, prec: 10.1%, Time: 1.376

.. bad effect, any problems?

[hou-yz]

Hi,

Thank you for your interest. However, we have no official support for digits datasets using ADDA. The current discriminator architecture in ADDA is designed for stronger classifiers like ResNet50 in VisDA datasets. Directly applying such a strong discriminator might introduce difficulties to the 3-layerl-convolution-classifier in digits datasets in terms of adversarial training.

Feel free to adjust the discriminator architecture a little bit (probably remove a few layers or see other ADDA implementations) to facilitate the digits datasets.

Yunzhong