[GitHub] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-330895488 @thirdwing CI has issue with some Python test, not related to this pull. This is an automated message from the Apache Git Service. To respond to the message, please log on 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] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-326354740 @thirdwing I'm out for 10 days, I'll try to put my hands a computer by then but it may take a few days. This is an automated message from the Apache Git Service. To respond to the message, please log on 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] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-325228128 Improve the harmonization with model.FeedForward: same fixed.params, arg.params and aux.params input arguments. Remove redundancies between model.buckets and model.train.buckets. model.buckets is as performant on single symbol model as model.Feedforward. If using BatchNorm, single symbol works fine but there's still an issue at inference if training on list of symbols/bucketing. Comment out the lstm in testthat to pass CI. Integrate CPU compatible RNN construction with raw lstm and gru cells in rnn.graph. No support for masking. Still need to test efficiency of inference and potentially adapt the inference functions. Unsure if wouldn't be preferable to assume a batch.size X seq.length input dimension to interator as it's the format expectged by symnol.RNN cell. Or add a shape detector to handle it automatically as in the python API. This is an automated message from the Apache Git Service. To respond to the message, please log on 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] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-325228128 Improve the harmonization with model.FeedForward: same fixed.params, arg.params and aux.params input arguments. Remove redundancies between model.buckets and model.train.buckets. model.buckets is as performant on single symbol model as model.Feedforward. If using BatchNorm, single symbol works fine but the still is issue if training on list of symbols. Comment out the lstm in testthat to pass CI. Integrate CPU compatible RNN construction with raw lstm and gru cells in rnn.graph. No support for masking. Still need to test efficiency of inference and potentially adapt the inference functions. Unsure if wouldn't be preferable to assume a batch.size X seq.length input dimension to interator as it's the format expectged by symnol.RNN cell. Or add a shape detector to handle it automatically as in the python API. This is an automated message from the Apache Git Service. To respond to the message, please log on 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] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-323931760 rnn.model.R is compatible with regular iterator and non RNN models. However, the need to bind at each batch results in a slower training time of about 15-20%. Also, aux.params doesn't behave as expected, so shouldn't use BatchNorm in a model trained with mx.rnn.buckets - yet to figure how to correct this. Example of application with [one-to-one](https://github.com/jeremiedb/mxnet_R_bucketing/blob/master/README_one_to_one.md) and [seq-to-one](https://github.com/jeremiedb/mxnet_R_bucketing/blob/master/README.md) updated. Several backend functionnalities were brought up explicitly in the mx.rnn.buckets. I think it makes it more transparent to track treatment of the arg.params. This is an automated message from the Apache Git Service. To respond to the message, please log on 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] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-322632234 @thirdwing `source()` and `library()` calls removed. Functions `mx.model.train.rnn.buckets` and `mx.rnn.buckets` merged into `model.rnn.R` in order to better align with `model.R`. Sorry for multiple commits - I struggled a bit with rebasing the submodules. This is an automated message from the Apache Git Service. To respond to the message, please log on 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] jeremiedb commented on issue #7476: R-package RNN refactor
jeremiedb commented on issue #7476: R-package RNN refactor URL: https://github.com/apache/incubator-mxnet/pull/7476#issuecomment-322632234 @thirdwing `source()` and `library()` calls removed. Functions `mx.model.train.rnn.buckets` and `mx.rnn.buckets` merged into `model.rnn.R` in order to better align with `model.R`. This is an automated message from the Apache Git Service. To respond to the message, please log on 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