Questions about training phase

[LiangbinXie]

Thanks for your excellent work~, and I have some questions about the training phase, hoping for your response.

1. It seems that there is something wrong with the file Vimeo7_dataset.py, the data structure of self.paths_GT is dict, so I think you need to add one line to make it works.
   
2. I am very curious about your runtime during training phase, when I trained the model from the stratch on 4 Nvidia 2080TI GPUs, it spent nearly 2 minutes for every 100 iterations, which is longer than EDVR.

[Mukosame]

Hi @LiangbinXie , sorry for the late reply cz I'm occupied with my finals recently. Thanks for bringing up these questions and nice comments!

1. sorry I didn't find this loc in the Vimeo_dataset.py. Would you like to point it out for me?

2, I only have 2GPUs and the data is stored on HDD. My training time is also ~2mins when I/O is fine, and sometimes can grow longer when the disk is busy, Since I haven't trained EDVR on the same machine, I don't know if it is a problem. Anyway, please let me know if you have good ideas to accelerate the training speed. I'd be more than happy to merge them into this repo :)

[LiangbinXie]

Thanks for your reply~

1. The specific number of rows is 66 where you definite the variables self.paths_GT.
   

2. At first, I think it will accelerate the training speed if I change the way of read image by using lmdb, but it seems not working. I'm sorry that I don't have a good idea to accelerate the training speed yet. Maybe ConvLSTM is the main reason.

[XY-boy]

Same question，self.path_GT is dict，can't be indexed by 【index】