foolyc / meta-sgd Goto Github PK

Thank you for your paper and project！I find that the only difference between Meta-SGD and MAML in source code is the update_lr is defined by tf.Variable(0.001, "updatelr") rather than defined by meta_update_lr. Is that correct？

hi,foolyc.
I have read your Meta-SGD code recently, that's a good project.
But I have some questions about the code and paper.
In the paper, update_lr and meta_lr seem like same item, and used element wise product with network parameters.
The line here seems not conform to the paper?
And set self.meta_lr and self.update same tf.variable?

I run code for :python main.py --datasource=omniglot --metatrain_iterations=40000 --meta_batch_size=32 --update_batch_size=1 --update_lr=0.4 --num_updates=1 --logdir=logs/omniglot5way/

and got this error in data_generator.py :
sampled_character_folders = random.sample(folders, self.num_classes)
File "/usr/lib/python2.7/random.py", line 321, in sample
raise ValueError("sample larger than population")
would you please help me?

Thank you for making this comparison between MAML and MetaSGD.

Could you share the final performance results for the two models? The training history is a bit unclear

Thanks.