cedricgsh / isdnet Goto Github PK

Hi,
First of all, thank you for your nice work. However, I have a small question about Deepglobe dataset.

As far as I know, the original validation and test sets do not provide labels. Are all the val and test sets in the current data split from the original training set?
According to your and GLNet's split file (.txt), the total number of train samples is 454 rather than 455. Why do you and your previous papers on related work explain that 455？

Thank you for your attention!

你好，请问ISDNET的模型还没上传吗？

代码网络结构未改过
显示不匹配

作者您好，
非常感谢您的工作，并将代码进行了开源。
我近期希望基于Deepglobe预训练的模型展开一些研究工作，在使用您的代码时遇到了一些问题。

1. 在使用提供的isdnet_r18-d8_1224x1224_80k_DeepGlobe.py和deepglobe.pth时，通过init_segmentor会出现警告The model and loaded state dict do not match exactly。
   既有
   unexpected key in source state_dict:
   decode_head.image_pool.1.bn.weight, decode_head.image_pool.1.bn.bias, decode_head.image_pool.1.bn.running_mean, decode_head.image_pool.1.bn.running_var, decode_head.image_pool.1.bn.num_batches_tracked, decode_head.aspp_modules.0.bn.weight, decode_head.aspp_modules.0.bn.bias, decode_head.aspp_modules.0.bn.running_mean, decode_head.aspp_modules.0.bn.running_var,……………………
   又有
   missing keys in source state_dict:
   decode_head.image_pool.1.conv.bias, decode_head.aspp_modules.0.conv.bias, decode_head.aspp_modules.1.conv.bias, decode_head.aspp_modules.2.conv.bias, decode_head.aspp_modules.3.conv.bias, decode_head.bottleneck.conv.bias,…………
   我不知道这是否是由于模型训练、推理阶段的模型结构差异导致的？

2. 在尝试使用demo中的方法，利用deepglobe预训练的ISDNet进行图像分割时，提示EncoderDecoderRefine类缺少out_channel属性。如果直接调用model.encode_decode函数，分类预测结果则非常不理想，不知道是否是问题1导致的。

目前始终无法debug完成正确推理，希望你能解答我的困惑。如果可以的话，能否提供一个利用deepglobe预训练的ISDNet的demo？

During training of ISDNet on DeepGlobe with the following script:

./tools/dist_train.sh configs/isdnet/isdnet_r18-d8_1224x1224_80k_DeepGlobe.py 1

where /tools/dist_train.sh has been just slightly modified to our setup:

#!/usr/bin/env bash

CONFIG=$1
GPUS=$2

PYTHONPATH="$(dirname $0)/..":$PYTHONPATH \
python3 -m torch.distributed.launch --nproc_per_node=$GPUS \
    $(dirname "$0")/train.py $CONFIG --launcher pytorch ${@:3}

We get the following error:

[...]
  File "/path/ISDNet/mmseg/models/decode_heads/refine_decode_head.py", line 217, in losses
    seg_logit = resize(
  File "/path/ISDNet/mmseg/ops/wrappers.py", line 26, in resize
    return F.interpolate(input, size, scale_factor, mode, align_corners)
  File "/usr/local/lib/python3.8/site-packages/torch/nn/functional.py", line 3163, in interpolate
    return torch._C._nn.upsample_bilinear2d(input, output_size, align_corners, sfl[0], sfl[1])
RuntimeError: It is expected output_size equals to 2, but got size 3

We have neither modified train.py nor the ISDnet model itself (taken as is from this git directory).
The DeepGlobe dataset has been split like requested, so we don't think this might be the issue. Or is there some preprocessing needed that we missed?
Do you have any idea why this error occurred, and where it might have come from?
Thanks again in advance for your help!

请问总loss是把这么多loss相加吗？

Thanks for the source code! I tried to reproduce results in the paper, but I only achieved 72.3% on DeepGlobe dataset. Could you please provide your training log?

您好，最近在研究您的工作，训练时遇到一些错误难以解决，错误如下图：

请问loss权重调参有什么策略吗，是调到一个量级吗？

The predicted result of this network output is RGB or grayscale image，there was a problem when I tested with the code you provided

Hi,

Thank you for your interesting work. Will the code be released any time soon?

你好， 请问有这个数据集的下载链接吗，官网上没有。
谢谢

您好，最近在研究您的工作，发现在复现过程中存在以下几点问题：

1. 论文中提到初始学习率为0.001，但是代码中确实却是保持mmsegmentation的默认配置，即0.01，请问具体训练的时候，学习率是怎样设置的呢？
2. 关于inria aerial数据集，给出的训练配置中，先resize到了5000*0.8，再进行的训练，但是deepglobe数据集确实直接使用原图大小2448进行的训练，这一点论文中并没给出，希望可以得到您的解答？
3. 关于对比实验部分，如deeplabv3+在test集上的预测结果，请问deepglobe的global训练和local训练，图像的数据集具体是多大呢，以及论文中给出的关于inria aerial数据集的对比实验，请问是local训练的结果还是global训练的结果呢，inrai aerial数据集的大小在其他对比模型中的大小又是多少呢？
4. 使用您给出的deepglobe数据集的配置文件进行训练，unknown那一类的iou结果始终为0，请问这个是什么原因呢？因为使用您提供的预训练权重在test集上进行测试，unknown这一类的iou 是有数值的，但是根据您提供的代码进行训练，unknown这一类的iou就没有值。
   您的工作非常出色，希望您在百忙之中能回答我的这些疑惑，这将对我的研究工作具有很大的帮助

你好，我使用自己的数据集训练，结果要比切成patch后训练差不少，是不是因为不切片数据量太少的原因，训练集大概401幅1000 x 1000像元。

你好，作者，请问RAF这个模块定义在哪个py文件下呢？

Hi, I'm trying your code. I'm new to MMCV library. I'm having a bit of difficulty with the package versions. Can you share the package versions that you used in this paper?

I tried with the latest versions and (MMSeg 0.16 and MMCV 1.3.7) but I still get an error every time. And also, I'm working on Windows, did you run this project on which system?

Thank you for your time.

你好！我想用ISDNet在新的数据集上进行训练，该数据有4通道，24个标签类别。我尝试在你的代码上进行修改后训练，但是无法达到基线方法的精度。请问应该如何修改代码？能否麻烦提供一下在新数据集上训练的教程？