I have been trying to find out that whether or not there is any problems with gradient calculation of LanguageModel and LanguageModelCriterion since gradCheckLM and gradCheck fail in "
"test_language_model.lua" couple of times, especially gradCheckLM. I did my own calculations and results are same as Karpathy but still gradcheck fails and don't know what is the problem. I was wondering does anybody know where is the bug/problem if there is any?
Thanks
PS: To be clear, the current version of the code fails the gradient check. Clone the code and run test_language_model.lua couple of times to see the fail cases.
Hi @karpathy and All,
I have been trying to find out that whether or not there is any problems with gradient calculation of LanguageModel and LanguageModelCriterion since gradCheckLM and gradCheck fail in " "test_language_model.lua" couple of times, especially gradCheckLM. I did my own calculations and results are same as Karpathy but still gradcheck fails and don't know what is the problem. I was wondering does anybody know where is the bug/problem if there is any?
Thanks PS: To be clear, the current version of the code fails the gradient check. Clone the code and run test_language_model.lua couple of times to see the fail cases.