[Question] What is the iteration for in the training

[TianyangChen357]

Hello Boulch,

Thanks for your nice work!

I was utilizing the pre-trained model you made for the semantic3D dataset and train a new model on my dataset.

I found # iteration impacted a lot of the model performance. I expect to know what it means but did not find an explanation in your paper. Could you give me a brief explanation about what it is?

Thanks so much for your help! Tianyang

[aboulch]

Hello, At training, the iter parameter is acually the number of batch in one epoch. It is true that is has a great impact: the number of training step on you model is define by num_steps = iter x epochs.

[TianyangChen357]

Thanks for your kind reply！！

Is here any reason that your design do not take all blocks in one epoch?

With many thanks, Tianyang

On Wed, Sep 2, 2020, 9:15 AM Alexandre Boulch [email protected] wrote:

Hello, At training, the iter parameter is acually the number of batch in one epoch. It is true that is has a great impact: the number of training step on you model is define by num_steps = iter x epochs.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/aboulch/ConvPoint/issues/32#issuecomment-685728770, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW7TOYO6VFSEPUE5RMTSDZAQNANCNFSM4PX2DTLA .

[aboulch]

Hello, At training blocks are not predefined but randomly sampled.