Find issue in implementation

Tensor-Reloaded / Convergence

Artificial Neural Network training convergence boosting by smart data ordering

Academic Free License v3.0

0 stars 0 forks source link

Find issue in implementation #3

Closed simi2525 closed 4 years ago

simi2525 commented 5 years ago

Currently the implementation will aggressively over-fit for the following parameters: "python main.py --batch_size=256 --interval=1 --lr=0.1 --model="VGG('VGG16')" --aux_batch_size=50000 --grad_clip=0.45 --descending --epoch_skip=10"

I can't really get why. Maybe someone can take a look and see what simple oversight made this happen.