search

ushasi / A-Simplified-framework-for-Zero-shot-Cross-Modal-Sketch-Data-Retrieval

We deal with the problem of zero-shot cross-modal image retrieval involving color and sketch images through a novel deep representation learning technique.
14 stars 1 forks source link

Question about the loss function #4

Open ghost opened 3 years ago

ghost commented 3 years ago

Dear author, thank you for your excellent job! But when I read your paper and code, I have encountered some problems.

  1. The loss function in your code is not consistent with your paper: Unified_XY_triads.py, line 411, cost = 0.1*loss0 + 0*loss3 + 0.1*loss4 + 1*loss5+ loss6 + loss7 #+loss8 # + 0*loss9 # MODEL 2 doesn't have all loss as your paper claims.
  2. In your paper, Section 3.3,

Since we have multiple losses, our problem becomes a non-convex optimization problem. To solve this, we reduce each loss term individually, keeping the other losses constant. The problem transforms into a convex optimization problem for that loss.

I don't understand. Neural Network is a non-convex function itself. Is a possible to change it to convex function by changing the loss function?

ghost commented 3 years ago

BTW, why you use GCN? graph = sg.StellarGraph(nodes, edges)， line129