Closed pingjun18-li closed 3 years ago
Thanks for your attention. Please followed the settings of our paper before getting normal results. The value of COCmap is limited to between -25 and 25 during training (see the paper for more details).
I think the results are far from the expectations.
@pingjun18-li I got the same result like Figure 1. Have you addressed the problem?
I think you should attach the training logs to better understand your questions (e.g. TensorBoard log and training log).
train_COCMAP_PARAM_TRAIN_210706-115411.log Here is the training log. And I just follow the original setting.
The training log is quite different form mine at the begin. Normally, the initial loss is approximately 10.
models/kernel_de_bparam_net.py
Does you uncomment it?
Yes, but the original setting is from -25 to 25. And I will retrain the model soon.
The setting (-25, 25)
may cause training to crash to some extent. Changing it to (-24, 24)
is a simple solution. Of course, you can modify the code of the loss function to avoid boundary exception.
Changing it to (-24, 24) doesn't work.
If you still suffer from a terrible initial loss value at the beginning, I suggest reimplement the repo in other devices (e.g. NVIDIA 3090 GPU).
Do you mean the loss is correlated with the type of device?
What I want to say is that the initial parameters of your network are poor, which leads to the initial solution of the network is too poor and can not be optimized at the beginning. In my experiment settings, the GPU we chose was NVIDIA 3090.
Would you mind uploading the pretrained model for COC map estimation?
Leave your email, I will provide you with an initial model with proper parameters.
Many thanks and my mail address is yarqian24@gmail.com.
Thanks for projects again! I used the config you given to train 20 epochs Cocmap on dataset of Canon by the preprocessing code 'image_to_patch_filter.py' , and the inference results are shown in the figure below Figure 1, all pixels is black with the value >1000.There are several questions: 1.The pixel value of the image is uint16, and the effect of the sub-area described in the paper cannot be seen. I would like to know whether my training is wrong or yours is the same. 2.Why three channel and not a single channel, and what's the meaning of the pixel value on each result graph? In addition, I modified the config "niter: 500000 epoch: 300 #20" , trained it at about 45 epoch(70 000 iter), and the phenomenon of unsupervised loss=0 would appear. Then, the test results were shown in Figure 2, is that normal?