bodokaiser
It's different with the standard SegNet
I think the most contribution in segnet is the idx-maxpooling. you can use the F.max_unpool2d(idx, x) to replace the nn.upsample(). The original upsample can't fit the loss of location information.
Hey, yes that makes sense! Would you mind to send a PR?
On 21. Nov 2018, at 11:07, toby [email protected] wrote:
I think the most contribution in segnet is the idx-maxpooling. you can use the F.max_unpool2d(idx, x) to replace the nn.upsample(). The original upsample can't fit the loss of location information.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.
yes, i'd like to do that. i have rewrited segnet.i will send it by my pull.
self.dec1 = features[0: 4] #[0,3]
self.dec2 = features[5: 9] #[5,8]
self.dec3 = features[10: 16] #[10,15]
self.dec4 = features[17: 23] #[17,22]
self.dec5 = features[24: -1] #[24,29]
while there's also a maxpool2d layer after each decoder block in the forward function, I wonder why the pooling operation is implemented twice before unpooling? I haven't run the code just quite confused.
@Viarow I guess this is a mistake then. Would you mind trying the correct version?
I'm actually a novice at deep learning, still trying to understand the codes. I guess it will work right only if the range of pretrained layers are set correct (as shown in the annotations above). I will tell you the results after I try the corrections.
Hi,there just one pooling at each dec. could you tell me the your found on detail.
If I understood correctly the problem can be found with the index selectors here: self.dec1 = features[0: 4] #[0,3] self.dec2 = features[5: 9] #[5,8] self.dec3 = features[10: 16] #[10,15] self.dec4 = features[17: 23] #[17,22] self.dec5 = features[24: -1] #[24,29]
So for example [0:4] will return [features[0], features[1], features[3]].
As the next layer starts with features[5] the features[4] layer is not used at all.
On 4. Feb 2019, at 10:11, toby [email protected] wrote:
Hi,there just one pooling at each dec. could you tell me the your found on detail.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bodokaiser/piwise/issues/24#issuecomment-460175873, or mute the thread https://github.com/notifications/unsubscribe-auth/ABsq8mP9fvms4uP3e0u5GZhuwdrDC8Tfks5vJ_k6gaJpZM4YsyhR.
Oh I see the point. in the FCN8 model, self.feats = nn.Sequential(*feats[0:9]) self.feat3 = nn.Sequential(*feats[10:16]) self.feat4 = nn.Sequential(*feats[17:23]) self.feat5 = nn.Sequential(*feats[24:30]) are called later in feats = self.feats(x) feat3 = self.feat3(feats) feat4 = self.feat4(feat3) feat5 = self.feat5(feat4) fconn = self.fconn(feat5) which means that some maxpooling layers are also left out in this model?