The performance difference when target domain is Clipart on office home dataset.

[mzmzdcr]

Hi！

Recently, when I was running your method, I encountered a problem. When I used baseline, the performance of all tasks whose target domain is Clipart decreased greatly. Can you help me solve this problem? It's more important to me.

[viyjy]

Hi, thanks for you interest in our work.
Do you mean that you can get reasonable performance on other tasks? Clipart is the only exception?

[mzmzdcr]

Right, I encountered this strange phenomenon. Can you give me some advice？

[viyjy]

You only changed the batch size? Can you exactly follow the hyper-parameters as I used?

[mzmzdcr]

@viyjy. I reduced the bs to 32 and adjusted the input resolution to 224. Other hyperparameters are completely set by default

[viyjy]

why do you make these changes? Have you tried the command line in script?

[mzmzdcr]

I need to control the consumption of GPU memory below 24g, so I use this setting.

[viyjy]

can you also decrease the lr by half and see what happens?

[mzmzdcr]

Let me test whether it is feasible to reduce the learning rate

[mzmzdcr]

I'm sorry to reply to you so slowly because recently there was a problem with our GPU server device. First, whether the learning rate is increased or decreased, the results obtained on the task whose target domain is clipart are reduced.

Could you spare time to test locally and share the training log with me? The specific settings are as follows:
--img_size 224 --train_batch_size 32 --beta 0.1 --gamma 0. --use_im --theta 0.1