[GitHub] [incubator-mxnet] anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function
anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function URL: https://github.com/apache/incubator-mxnet/pull/17298#issuecomment-574802286 @haojin2 if I randomized the input data in the original test code the losses would would have different values during each run (SDML loss imposes a distribution over the relative distances of data points in a minibatch) - so I would not be able to compare the output against precomputed loss values any more - thus the original unit test procedure cannot be reused. That's why I added a test that fits a toy model to some toy data instead. The current test was running in ~50 ms on my machine on CPU. Would love to hear your thoughts on how to improve on this. This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function
anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function URL: https://github.com/apache/incubator-mxnet/pull/17298#issuecomment-574458020 It looks a little tricky to port this into the 'fit' and 'score' paradigm since this is a retrieval specific loss function which uses the other elements in a batch as implicit negative samples - and I'm not sure how cleanly it fits into the Module API for this kind of test. Specially since the loss computation needs to know the shape of the minibatch which doesn't seem to be possible in the symbol API. The loss only guarantees that associated pairs will be closer in the chosen metric space after learning as compared to the non-associated pairs. Maybe I can write something equivalent using the gluon API, to train a small network and ensure it learns the right associations. I'll come up with a proposal shortly. This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function
anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function URL: https://github.com/apache/incubator-mxnet/pull/17298#issuecomment-574458020 It looks a little tricky to port this into the 'fit' and 'score' paradigm since this is a retrieval specific loss function which uses the other elements in a batch as implicit negative samples - and I'm not sure how cleanly it fits into the Module API for this kind of test. The loss only guarantees that associated pairs will be closer in the chosen metric space after learning as compared to the non-associated pairs. Maybe I can write something equivalent using the gluon API, to train a small network and ensure it learns the right associations. I'll come up with a proposal shortly. This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function
anjishnu edited a comment on issue #17298: [MXNET-1438] Adding SDML loss function URL: https://github.com/apache/incubator-mxnet/pull/17298#issuecomment-574415884 @haojin2 Sure will address the sanity cases. Can you give an example of a unit test that is appropriately randomized so I can base it on that? This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
