vanstorm9/yugioh-one-shot-learning

design questions

simonbuehler opened this issue · 1 comments

hi,

inspiring work! i'm trying it to adopt it for pokemon cards but some questions popped up, maybe you can enlighten me ;)

  • is there a reason not to uses the color images but clone the grayscale version on all color channels ?
  • is there a reason that not the resnet101 (224,224) dimension is used but 244x244?
  • are the performance gains to inline the generation of the augmented images and not generate /read them from disk?
  • did you test Prototypical networks ?
    --
    edit:
  • where does the Blur pooling mechanism hiappen?

greets!

and another one:

what could cause the TripletLoss to be often be zero ?