Hi @GeoZcx, I find a phenomenon that when I initialization of network parameters, the network can't detect any change in Pytorch Version. But when I don't initialize of network parameters the network works well. The follow is the code of initialization

`
def _initialize_weights(self):

    for m in self.modules():
        if isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d) or isinstance(m, nn.Linear):
            nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            if m.bias is not None:
                m.bias.data.zero_()
        elif isinstance(m, nn.BatchNorm2d):
            m.weight.data.fill_(1)
            m.bias.data.zero_()

`

In the pytorch version code:

1. 'Sigmoid' is used when you build DSIFN model, but it's used again in your loss function.

2. Then, in loss.py, why use 'Sigmoid' function for 'target' ?

Hi @GeoZcx First of all thanks for your great repo.

Unfortunately, Beidu drive isn't available in my country. Could you please provide a Google drive or even Onedrive version of it?

Sorry, my English is not very good. About your codes in pytorch version, I am comfused about a problem need your help. Why is the convolution output of the last layer of the network single channel? In remote sensing change detection. Generally, there are two types of change and non change. If output only single channel, how do we get the results of change?

During training, VGG backbone is frozen with pretrained weights by self.features = nn.ModuleList(features).eval() so that only the weight parameters of DDN will be updated.

Environment:

Ubuntu 16.04
Python==3.6.8
PyTorch==1.4
a RTX2080Ti GPU

Hyperparameters setting:

batch_size = 8
initial learning rate=0.0001
adam optimizer

Would you mind releasing the complete training code? Thanks.

I want to do some research on change detection in big map. But there are few public change detection dataset of big map. I see the raw big maps are shown in readme. Does it mean that the raw image can be obtained by stitching, or the raw image will be made public in the future?

Hello, I am very interested in your paper, but I have some problems in understanding your paper and code. Your paper said that the change detection task is divided into two stages. First, the bitemporal image is sent to the pre-trained DFEN (two VGG networks) to generate deep features, and then the extracted deep features are input into DDN to distinguish the changed areas. What do you mean by saying that these two parts should be trained independently? Does it mean to use the feature extraction method in transfer learning to extract features and then directly put them into DDN for training? You did not use the fine-tuning method, right? Hope to hear from you , thank you very much!

The 512x512 resultion masks in the train and val set that you released are all black. Did you train your model by upsampling the masks that resultion is 256x256, or some other methods ?

When defining the model，generally this is the form："class Unet(nn.Module): def init(self, input_nbr, label_nbr):". Why is there no choice of input and output channels in your Pytorch codes. In your codes, “class DSIFN(nn.Module): def init(self, model_A, model_B):”, model_A and model_B represent what? vgg16_base? ? Looking forward to your answer, it will be very helpful to me. Thank you!

Hello, what is the role of model_A and model_B or where is its definition, and what should be input? What did they do to "t1_input" and "t2_input" ? Thanks.

Hi, I am confused about your training process. Based on my understanding，you load 'imagenet' pretrained weights on Vgg16, then you freeze Vgg16 parameters and only train DDN part. Is my understanding right？According to the paper, I noticed that the deep feature extraction network (DFEN) and the difference discrimination network (DDN) in the proposed method are independently trained during the training process, but if you just load 'imagenet' pretrained weights on Vgg16, you don't train Vgg16. In addition to this, if you just use pretrained weights, I question the effect of pretrained weights.

Hello, I recently came across your paper. I am unable to find the training code for your model. Can you please share your training code for keras? Can you provide it here or in the google drive link for the dataset? Thank you!