search

xukechun / Efficient_goal-oriented_push-grasping_synergy

[RAL & IROS 2021] Efficient learning of goal-oriented push-grasping synergy in clutter
60 stars 10 forks source link

How to continue training at stage 2 #12

Open andreipit opened 2 years ago

andreipit commented 2 years ago

After step 1 we have the "...grasp_only.pth" network. And we pass it to the 2nd stage as the "--snapshot_file" argument. Step 2 creates "...push_only.pth". Let's say the training process was interrupted. It's unclear how to continue learning at stage 2. We can't use the "--snapshot_file" argument anymore. So how do we input "...push_only.pth"?

Kamalnl92 commented 2 years ago

You are right you cannot use it. What you have now ".push_only.pth" is the weight values of the model connections. if you want to continue training a model you need to save checkpoints. That is not implemented in this source code. have a look at this from PyTorch: https://pytorch.org/tutorials/recipes/recipes/saving_and_loading_a_general_checkpoint.html

Cheers, Kamal

andreipit commented 2 years ago

Thank you for the answer, Kamal! And to make things clear: 1) At stage 3 "Train the grasp net": --snapshot_file is "...push_only.pth" (from stage 2)? 2) At stage 3 "Train the push net": --snapshot_file is "...push_only.pth" (from stage 3 "Train the grasp net")? 3) At stage "Test": --snapshot_file is what?)