the output is strange after training on my own dataset

[milk-bottle-liyu]

Hi,friend, it's me again. <_>
I use my own data to train my network. Because I only recorded the steer angle, so I modified it to output only one float, the steer angle. However something strange happened. The mean loss in training stage stays high. Then I check the output on the val dataset and find the values are all positive. Though my dataset is not as big as yours, it still contains thousands samples from 4 scenes. Some of my outputs are in the picture. I tried to normalized the data, but it did not work. What should I do to solve the problem?

[wangjksjtu]

Hi, it is pretty strange. There must be some bugs while transplanting to your dataset. I could not ensure the problems with the information provided. Have you modified the loss by the way or adjust the hyper-parameters? Will the training loss decrease in the training?

[milk-bottle-liyu]

Sorry for not responding in time. I don`t understand why should I change the loss? For now I am working on the image only model, and besides IO I did not change any other things. The loss only drops in the first epoch and did not decrease over dozens epochs. The loss is like:

What can I do to make the loss decrease? I had doubled the train set already.

[milk-bottle-liyu]

Also I change the lr decrease steps, so the lr is normal, I think. I also tried to use the pre-trained res-151 model to initialize the network but it turned out to be the same.

[wangjksjtu]

Sorry, I still don't know the causes from your description :(
It is hard to diagnose from the experience without the code and data.