lschmidtke
Issue reproducing the uploaded pretrained model
Hello!
I have been trying to reproduce the pretrained checkpoint using the uploaded code (model.py, networks.py, the utils, train.py) to train on human3.6m, but it is not successful. Could you please let me know if you could upload the exact code repository that was used to get the uploaded pretrained checkpoint? Thank you so much in advance!
Also, could you please let me know approximately what the size of the data you trained on was (i.e. approximately how many pairs of images the pretrained model was trained on)? I tried training the model on around 180,000 pairs, but was not able to reproduce the pretrained checkpoint.
Hello,
Specifically, these are the results I am getting:
Could you please let me know how I might fix this, since it does not match the pretrained model? Thank you so much!
Hi! The code should produce the same results of the provided checkpoint. How do the reconstructed images look like? The model was trained on roughly 600k images.
Okay, the reconstructions don't look good. What's your batch size and learning rate? Also, how many epochs are you training for?
Hello! My batch size is 48. Learning rate is 0.0001. I'm training 50 epochs.
On Thu, Mar 9, 2023 at 10:34 AM Luca Schmidtke @.***> wrote:
Okay, the reconstructions don't look good. What's your batch size and learning rate?
— Reply to this email directly, view it on GitHub https://github.com/lschmidtke/shape_templates/issues/8#issuecomment-1462262594, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHRV2UWDIB7EVM7GVJ5ASXTW3H2ADANCNFSM6AAAAAAVLSYBAQ . You are receiving this because you authored the thread.Message ID: @.***>
That seems reasonable to me. I will run the code myself with those parameters and I'll get back to you.
Hello,
Could you please let me know when there are updates? Thank you so much!
Hi,
I encountered the same issue, I suppose it's due to a newer version of one of the packages. I will try and find the cause.