search

cha15yq / MRC-Crowd

Implementation of "Semi-Supervised Crowd Counting with Contextual Modeling: Facilitating Holistic Understanding of Crowd Scenes"
7 stars 0 forks source link

The validation set issue #2

Open yaotao05 opened 7 months ago

yaotao05 commented 7 months ago

Hello! Thank you for finding my work meaningful for your research. When training on the ShanghaiTechA dataset, how is the validation set selected?

cha15yq commented 7 months ago

I didn't use a validation set for SHA and SHB.

yaotao05 commented 7 months ago

Is it the epoch with the best MAE or MSE during the training process?

yaotao05 commented 7 months ago

Is the best result to train in 1000 rounds? The number of rounds I have trained at shanghaitachA is 1500: MAE: 64.99; MSE: 113.07; 102.37，180.56 for 1000 rounds on QNRF;

cha15yq commented 7 months ago

I have attached the SHA training log in this repo. Please check that.

yaotao05 commented 6 months ago

Thank you for your reply. I still have some questions. I saw your log and did you use the data from the test set for verification? What data is still being used? Will using test set data for validation affect the results? Besides, the QNRF pre trained models MAE: 102.37 and MSE: 180.56 (1000 epochs), I have retrained them again now. Is there anything I can advise on?

yaotao05 commented 6 months ago

Thank you for the author's help. I have just been exposed to deep learning and learned about the application of validation sets. I am currently training on shanghaiA results and have noticed that the learning rate may have been configured incorrectly. May I ask if QNRF and shanghaiA have the same learning rate? Can you publish the QNRF log？

cha15yq commented 6 months ago

For QNRF, the learning rate is 2e-5.

On Sun, 7 Apr 2024 at 05:37, yaotao05 @.***> wrote:

Thank you for the author's help. I have just been exposed to deep learning and learned about the application of validation sets. I am currently training on shanghaiA results and have noticed that the learning rate may have been configured incorrectly. May I ask if QNRF and shanghaiA have the same learning rate? Can you publish the QNRF log？

— Reply to this email directly, view it on GitHub https://github.com/cha15yq/MRC-Crowd/issues/2#issuecomment-2041307629, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANGLTWJQGRWRFJOMQH27PIDY4DERRAVCNFSM6AAAAABFRFFXCCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDANBRGMYDONRSHE . You are receiving this because you commented.Message ID: @.***>