Trainning details on single-view and multi-view
wangleishan opened this issue · comments
Sorry to bother you again，
After I train the whole network (without the context-aware fusion module) with a single-view image for 250 epochs. What should I do to fix the encoder and decoder and train the rest of the network for 100 epochs ? How to fix the encoder and decoder?
Thank you for your reply，
After referring to this, I tried to modify the code to complete the multi view reconstruction, but failed. Can you provide the modified multi view reconstruction code?
I didn't save that file. I think it is simple to do this.
Thank you for your reply. By the way, when you are training multi-view, when n_view=20, how long does the training time take and how many GPUs are used?
Thank you for your reply. By the way, when you are training multi-view, when n_view=20, how long does the training time take and how many GPUs are used?
I can't remember it clearly. Maybe a single 1080Ti GPU is enough if you fix all networks except the merger.
Thank you for your reply. By the way, when you are training multi-view, when n_view=20, how long does the training time take and how many GPUs are used?
I can't remember it clearly. Maybe a single 1080Ti GPU is enough if you fix all networks except the merger.
Thank you . After I train the whole network with a single-view image for 250 epochs, I modified the code as you said to fix the encode and decoder (The added code is shown in the figure ),But when I set n_view=20 , the batch-size must reduce to 4 or 8 and the training time is so long. My gpu is a single Tesla V100(32G) . Is there something wrong with my code ？
Sorry, I cannot remember it clearly.
Maybe you don't need to train with 20 views? I'm not sure.
Do you remember the approximate time of multi-view training?
No. Maybe 3 days?
Thank you for your prompt reply !