Hey, I really like your code and wanted to use it but when i add the coco data handling from the predecessor repos and train from scratch on coco it does not work. The training curve looks good and no errors or problems however results are really bad and not in the right direction at all. Since there are no errors and hints I am a bit stuck. I checked the code against lasseha and matterport and looks all good. Do you know what i can look into? Have you ever tried training on coco from scratch? I would be really happy if you could help since i have limited time to make it work and I really like this repo. If necessary I can share my code somehow I made some changes. One bug I found is on line: 221 in data_generator.py. Seems that normalization was moved to the outside of the loop by mistake. Also mean subtraction was done twice in the training of the network. I would appreciate any help.
Hey, I really like your code and wanted to use it but when i add the coco data handling from the predecessor repos and train from scratch on coco it does not work. The training curve looks good and no errors or problems however results are really bad and not in the right direction at all. Since there are no errors and hints I am a bit stuck. I checked the code against lasseha and matterport and looks all good. Do you know what i can look into? Have you ever tried training on coco from scratch? I would be really happy if you could help since i have limited time to make it work and I really like this repo. If necessary I can share my code somehow I made some changes. One bug I found is on line: 221 in data_generator.py. Seems that normalization was moved to the outside of the loop by mistake. Also mean subtraction was done twice in the training of the network. I would appreciate any help.