feevos / ceecnet Goto Github PK

hi feevos, the result in your paper has IOU metric in table. Is the IOU metric for change only ? Or the IOU euqal mIOU for 2 classes

Hi feevos, should dists multiply 100 to calculate the loss? in chopchop2rec.py, the dists multiplied 100.

hi feevos, the segmentation head in this paper, the distance is between feature1 and feature2 ? and it likes to Euclidean distance ？ and the Boundary is how to produce? I haven't used mxnet , so It's difficult to read your code for me. But want to learn your ideas in segmentation head. Can you explain your ideas briefly ? thanks

I use these environment:

1. docker pull bitnami/mxnet:1.8.0-debian-10-r188
2. docker pull bitnami/mxnet:1.6.0
3. pip install mxnet
4. pip install mxnet-cu100
   ...

to run these code:

import sys
sys.path.append('/mount/home/bluemoon/nas_194/repos-read/')
from mxnet import nd 
from ceecnet.models.changedetection.mantis.mantis_dn import *

# D6nf32 example 
depth=6
norm_type='GroupNorm'
norm_groups=4
ftdepth=5
NClasses=2
nfilters_init=32
psp_depth=4
nheads_start=4

net = mantis_dn_cmtsk(nfilters_init=nfilters_init, NClasses=NClasses,depth=depth, ftdepth=ftdepth, model='CEECNetV1',psp_depth=psp_depth,norm_type=norm_type,norm_groups=norm_groups,nheads_start=nheads_start)
net.initialize()

BatchSize = 4
img_size=256
NChannels = 3

input_img_1 = nd.random.uniform(shape=[BatchSize, NChannels, img_size, img_size])
input_img_2 = nd.random.uniform(shape=[BatchSize, NChannels, img_size, img_size])

outs = net(input_img_1, input_img_2)

, but all these environment will report error at this step “outs = net(input_img_1, input_img_2)”， the report error info is "Illegal instruction (core dumped)"

Hello, boss, can you write a little more detailed readme file? Thank you.

Can I get a trained model for this work? I need to perform some inferences to understand if I can use this for my own work.

I have been training CeecnetV1 on LEVIRCD Dataset, even after 120 epochs I am getting an average f1-score between 0.35-0.45 for segments while average loss is 0.22. Around how many epochs should I see f1 score going up?

The layer in nn/layers/attention.py does not have multihead attention as described in the paper.

Haha, I have seen the review report of your paper on Remote Sensing, and I think reviewer 3 is crazy and irrational.

Fortunately, your work has been published. Congratulations!

hi, boss, i trained on levircd datasets, when i just overfit one image, all losses could decline quickly and got a pretty results. but when i training on all after 60 epoches, the boundary loss and distance loss do not decline, and the segmentation loss is fluctuation,
the segmentation prediction tends to be all zeros. i have 8 gpus, 2 batch sizes per gpu, i changed lr from 1e-3 to 1e-7,but all got bad results. is there any suggestions to avoid the segmentation prediction tends to be all zeros?