evan69
commented
4 years ago I am very interested in your work and I have read your original paper carefully. This is a great work but when I reproduced HSML by pytorch, I encountered some questions.
When I counted the amount of parameters used in this code for 5-shot regression task by tf.trainable_variables, I found that the amount for that(around 170000 params) is about 100 times larger than that of vanilla MAML model(around 1700 params). In your paper, the HSML model achieves much better than vanilla MAML model in both 5-shot and 10-shot experiments, but I don't think it's a fair comparison since there exists a large gap between the models' size. So I wonder if there exists larger MAML model test to show the validation of HSML, or I just have some misunderstandings about it.
In addition, have you conducted the same tsne plot in regression tasks as that in classification tasks, since I have reproduced HSML with pytorch but I cannot achieve great tsne plot to distinguish different functions. I am wondering if there exists some mistakes in my implementation, or it's just hard for HSML to distinguish different functions in regression tasks.
Thank you for answering my question above!
Have you successfully reproduced HSML by pytorch on image classification task? I would appreciate it if you can release this part of your code. Thanks in advance!
I am very interested in your work and I have read your original paper carefully. This is a great work but when I reproduced HSML by pytorch, I encountered some questions.
When I counted the amount of parameters used in this code for 5-shot regression task by tf.trainable_variables, I found that the amount for that(around 170000 params) is about 100 times larger than that of vanilla MAML model(around 1700 params). In your paper, the HSML model achieves much better than vanilla MAML model in both 5-shot and 10-shot experiments, but I don't think it's a fair comparison since there exists a large gap between the models' size. So I wonder if there exists larger MAML model test to show the validation of HSML, or I just have some misunderstandings about it.
In addition, have you conducted the same tsne plot in regression tasks as that in classification tasks, since I have reproduced HSML with pytorch but I cannot achieve great tsne plot to distinguish different functions. I am wondering if there exists some mistakes in my implementation, or it's just hard for HSML to distinguish different functions in regression tasks.
Thank you for answering my question above!