【Experiment】Single-Model VS Multi-Model

[bobo0810]

Data source

Private data based on 4 scenarios，a total of about 62,000 data.

Conclusion

(1) Under the same parameters, the multi-task model will decrease by ~2%. (2) Note, however, that this is not fair. Due to the mutual influence of each subtask, the multi-task model should be trained more times in theory

[piercus]

Hi @bobo0810 seems very interesting result.

Can you help me understand more precisely what is task1, task2, task3 and task4, which config file correspond to each task ? Maybe it would be helpful to give them more expressive names ?!

Is it still true if instead of 62000 data you limit yourself to less data (for example 600 data, or 6000 data) ?

[bobo0810]

Can you help me understand more precisely what is task1, task2, task3 and task4, which config file correspond to each task ? Maybe it would be helpful to give them more expressive names ?!

@piercus Thank you for noticing my work. I'm very sorry, because it involves the internal data of the company, so there is less information to post.

	num_class	data size
Task 1	8	44283
Task 2	17	10708
Task 3	2	5700
Task 4	2	12065

• backbone=SwinTransformerV2
• epoch=50
• batch=8*128

config.zip

Is it still true if instead of 62000 data you limit yourself to less data (for example 600 data, or 6000 data) ?

untested