search

LJOVO / TranSalNet

TranSalNet: Towards perceptually relevant visual saliency prediction. Neurocomputing (2022)
https://doi.org/10.1016/j.neucom.2022.04.080
MIT License
51 stars 10 forks source link

The train loss is nan #1

Closed liangchengbeibian closed 1 year ago

liangchengbeibian commented 1 year ago

I retrain fine-tune&train.ipynb with the same dataset and setup, but the train loss is very easy to be nan. Have you encountered this situation? Can you provide a more detailed experimental environment such as random seed and progress of dataset?

LJOVO commented 1 year ago

Hi, I have not encountered this kind of problem before.  Please refer to the following supplementary experimental environment:  CUDA 11   cuDNN 8.0.4 For the SALICON dataset, the stimuli and saliency maps were simply resized from 640x480 to 384x288, and the fixation maps (binary) were generated from the coordinates after resizing. The models were trained without specifying a particular random seed for loading datasets.

liangchengbeibian commented 1 year ago

thank you for your reply, the problem appears in the processing of the label.