MAML makes weights stuck in some bad equilibrium about maml HOT 3 CLOSED

from maml.

That's strange. I haven't seen anything like that before. Have you tuned the update_lr? (alpha in the paper) It's possible that it might be much too large or much too small.

Can you visualize the motor current for multiple tasks?

Are you using a different sample for the inner and outer objectives? (You should be.)

from maml.

Yes I played around a lot with the learning rates. I found out that it is very sensible to those learning rates, and doesn't learn at all, if its wrong. Currently i have meta_lr 0.0001 and update_lr 0.00001 while having num_updates 3. I recognized, if I make num_updates > 1, I have to make update_lr rather small that it learns anything - is this probably because the updates regarding every single task get to different, when going many steps into one task?!

I think for some reason, now the equilibrium is no problem anymore, don't know why. It takes me about 100 iterations of finetuning with a higher learning rate to come to a good accuracy.

What number of inner gradient steps do you normally use for good results?

The motor currents for the multiple tasks are not so different. Just having their spikes etc. at different locations.

What exactly does the console output of the preloss and postloss mean. Loss before updating the inner gradients and afterwards?!

Are you using a different sample for the inner and outer objectives? (You should be.)
I guess the code below is creating different samples for the inner and outer objective?! a single tasks, b meta training?!

inputa = tf.slice(image_tensor, [0,0,0], [-1,num_classes*FLAGS.update_batch_size, -1])
inputb = tf.slice(image_tensor, [0,num_classes*FLAGS.update_batch_size, 0], [-1,-1,-1])
labela = tf.slice(label_tensor, [0,0,0], [-1,num_classes*FLAGS.update_batch_size, -1])
labelb = tf.slice(label_tensor, [0,num_classes*FLAGS.update_batch_size, 0], [-1,-1,-1])

from maml.