training on a large number of concepts

[Nikiforos Pittaras]

I want to run a network on the 1K imagenet concepts + a few more thousand concepts.

Anyone have an tips/material on the training strategy? Should I train all concepts alltogether or train a few sets at a time and "fine-tune" the network on additional concepts later?

Thanks.

[npit]

I have attempted training a few times.

I have 4K classes, with ~ 4 million training images, ~ 200K validation images.

I have set test batch size and test_ter s.t. batchSize x test_iter = number of validation images.

I have used various batch sizes and learning rates.

Every time, after a few thousand iterations the test accuracy gets stuck at a value close to 0.1 %.

The training loss fluctuates around random guess levels, i.e. around 8.2