How could I modify the network so it can take images of greater size as input?
Opened this issue · 2 comments
JavierUrenaPhDProjects commented
Hello, so I'm doing a comparative study with other similar networks and I want all of them to take as input 384x384 images. I tried to register a new model, similar to CCTrans_large, which allows input of 256x256, but modifying the hyperparameters so it could fit 384x384 images. The problem is I dont know what to tweak in order to accomplish it. What should I put attention to, or change, in order to configure the model correctly for images of 384x384? (or any desired size?)
wfs123456 commented
Two solution:
(1) The input size must be divided by 256，so u can set 256, 512, 768
(2) If the input size must be 384, update
wss=[8，8，8，8] to wsd=[8, 8, 8, 6] in this function (def alt_gvt_large).
…---Original---
From: ***@***.***>
Date: Fri, Jun 2, 2023 20:10 PM
To: ***@***.***>;
Cc: ***@***.***>;
Subject: [wfs123456/CCTrans] How could I modify the network so it can takeimages of greater size as input? (Issue #31)
Hello, so I'm doing a comparative study with other similar networks and I want all of them to take as input 384x384 images. I tried to register a new model, similar to CCTrans_large, which allows input of 256x256, but modifying the hyperparameters so it could fit 384x384 images. The problem is I dont know what to tweak in order to accomplish it. What should I put attention to, or change, in order to configure the model correctly for images of 384x384? (or any desired size?)
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
JavierUrenaPhDProjects commented
Ok, it seems that it works! thanks! can I ask exactly what was changed here so the model could work correctly?