search

shayzweig / InterpoNet

A brain inspired neural network for optical flow dense interpolation
GNU General Public License v3.0
18 stars 6 forks source link

Training code #2

Open fperezgamonal opened 5 years ago

fperezgamonal commented 5 years ago

Hello,

I know it has been a long time since this repository was last updated but I was wondering if you still have access to the training code so we could more fairly compare our method and others with yours by using other matchers as inputs (not only FlowFields).

Additionally, if you would be so kind to confirm if you are using the FlowFields code I found here (link 4), that would be very helpful.

Thanks in advance,

Cheers, Ferran.

shayzweig commented 5 years ago

Hi Ferran, It has been a long time since I touched this piece of code. About the flowfields code, I am pretty sure it it is indeed the right one. About the training code, unfortunately I don't have an organized version of this code, I can give you access to a private repo that has all of my experimentation piled up but to be honest, it is a mess, and it will take time to organize which unfortunately I don't currently have. LMK if you want access. Best, Shay

shayzweig commented 5 years ago

BTW, I think your best approach would be to take the inference model and just build the training code around it.

fperezgamonal commented 5 years ago

Hello Shay,

Thanks for your quick reply. I think your idea of building the training code around the inference one is great though, if it is at all possible, I would like to take a look at your experimentation in case that may speed-up the process as we have a deadline in less than a month (to see if we get anything worth sharing in our tests).

Lastly, if you would be so kind to clarify one doubt about the training schedules in InterpoNet. In section C. Early Stopping you said that:

If I am not mistaken, you mean that in total, you do 5,000 4 = 20,000 iterations for FlyingChairs, 1,0004 for Sintel, and 400 * 4 = 1,600 for KITTI, is that right? Compared to training a U-Net like architecture, these numbers seem rather small but may totally plausible since you use a simpler network, different loss terms and validate on Sintel (which in my experiments also helps generalisation significantly, at around 2x the convergence "rate").

Thanks again for your time and help