search

lisa-groundhog / GroundHog

Library for implementing RNNs with Theano
BSD 3-Clause "New" or "Revised" License
598 stars 229 forks source link

Inf log-probs when using score.py #45

Open zhangdongxu opened 8 years ago

zhangdongxu commented 8 years ago

Meet this issue when implement score.py. And all the log-probs are inf. I test this on the training data, where in the training process, most target sentences can be recovered. In my opinion, zero probability should not exist since in the softmax, there is inner product between two vector. Any suggestions?

nouiz commented 8 years ago

Just to tell you that this repo isn't maintained anymore. See the warning on this page:

https://github.com/lisa-groundhog/GroundHog

I'm not sure someone else will reply. There is another project that you can use instead of groundhog that is based on Theano.

On Sat, Jan 23, 2016 at 11:07 PM, zhangdongxu notifications@github.com wrote:

Meet this issue when implement scorepy And all the log-probs are inf I test this on the training data, where in the training process, most target sentences can be recovered In my opinion, zero probability should not exist since in the softmax, there is inner product between two vector Any suggestions?

— Reply to this email directly or view it on GitHub https://github.com/lisa-groundhog/GroundHog/issues/45.

zhangdongxu commented 8 years ago

Thank you. I notice that. But I concerned that has attention-based structure been employed in Block's translation system?

rizar commented 8 years ago

Yes, the implementation in Blocks uses attention.

On 25 January 2016 at 08:41, zhangdongxu notifications@github.com wrote:

Thank you. I notice that. But I concerned that has attention-based structure been employed in Block's translation system?

— Reply to this email directly or view it on GitHub https://github.com/lisa-groundhog/GroundHog/issues/45#issuecomment-174512538 .