nizhenliang / CGRSeg

[ECCV 2024] Context-Guided Spatial Feature Reconstruction for Efficient Semantic Segmentation
Apache License 2.0
58 stars 3 forks source link

NO change #8

Open ZhenboZhao77 opened 3 months ago

ZhenboZhao77 commented 3 months ago

Dear Author,

I am very excited to see your masterpiece and couldn't wait to add the RCM module to my model. The code is as follows: from .rcm import RCM, RCA self.rcm = RCM(dim=256) x = self.rcm(x) However, I found that the model didn't seem to change. It neither improved nor worsened. Did I use the RCM correctly, or do I need to adjust some hyperparameters?

ZhenboZhao77 commented 3 months ago

hello,can you help me?

nizhenliang commented 3 months ago

Dear Author,

I am very excited to see your masterpiece and couldn't wait to add the RCM module to my model. The code is as follows: from .rcm import RCM, RCA self.rcm = RCM(dim=256) x = self.rcm(x) However, I found that the model didn't seem to change. It neither improved nor worsened. Did I use the RCM correctly, or do I need to adjust some hyperparameters?

You can adjust the training settings, such as the learning rate and training policy. The location added to the network will also have an effect. You can try to adjust it. The number of channels can also be adjusted to fit your model.

Billy-ZTB commented 2 weeks ago

Do you see any improvement now? Where did you put this module in?

Billy-ZTB commented 2 weeks ago

Dear Author, I am very excited to see your masterpiece and couldn't wait to add the RCM module to my model. The code is as follows: from .rcm import RCM, RCA self.rcm = RCM(dim=256) x = self.rcm(x) However, I found that the model didn't seem to change. It neither improved nor worsened. Did I use the RCM correctly, or do I need to adjust some hyperparameters?

You can adjust the training settings, such as the learning rate and training policy. The location added to the network will also have an effect. You can try to adjust it. The number of channels can also be adjusted to fit your model.

Hello! Should the RCM act like the way in your article which uses multiple RCMs and multiplied in a recurrent way? I only used one RCM after the encoder to aggregate features in the last 3 stages of encoder, and my model is like the Unet, but it didn't show improvements. But it confuses me when I try to use the similar architecture in unet.

nizhenliang commented 2 weeks ago

Yes, you should use multiple RCMs. One RCM is not enough to extract multi-scale contexts well.

Billy-ZTB commented 2 weeks ago

Thanks! I will try it later.

---- Replied Message ---- | From | @.> | | Date | 10/15/2024 10:26 | | To | @.> | | Cc | Zhang @.>@.> | | Subject | Re: [nizhenliang/CGRSeg] NO change (Issue #8) |

Yes, you should use multiple RCMs. One RCM is not enough to extract multi-scale contexts well.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>