CNNTrainLang issueshttps://git.rwth-aachen.de/monticore/EmbeddedMontiArc/languages/CNNTrainLang/-/issues2019-11-04T16:27:33+01:00https://git.rwth-aachen.de/monticore/EmbeddedMontiArc/languages/CNNTrainLang/-/issues/3Shuffle Parameter2019-11-04T16:27:33+01:00Evgeny KusmenkoShuffle ParameterAt the moment shuffling is hard coded (set to True by the generator); please create an appropriate parameter in the train languageAt the moment shuffling is hard coded (set to True by the generator); please create an appropriate parameter in the train languagehttps://git.rwth-aachen.de/monticore/EmbeddedMontiArc/languages/CNNTrainLang/-/issues/2Loss function2019-02-16T02:12:51+01:00Evgeny KusmenkoLoss functionloss function is set to a default value currently based on the output layer
Please add a loss parameter to training language.
therefore, create a table with
- default loss functions for different output layers (e.g. softmax -> cross ent...loss function is set to a default value currently based on the output layer
Please add a loss parameter to training language.
therefore, create a table with
- default loss functions for different output layers (e.g. softmax -> cross entropy, fc -> mse, etc.)
- forbidden (or allowed) combinations (xentropy cannnot be used if the output is not a valid probability, etc)
extend the generator
lets support the loss functions listed here:
https://heartbeat.fritz.ai/5-regression-loss-functions-all-machine-learners-should-know-4fb140e9d4b0
@sebastian.nickels do you need any other loss definitions?Eyüp HarputluCarlos Alfredo Yeverino RodriguezEyüp Harputlu2019-01-16https://git.rwth-aachen.de/monticore/EmbeddedMontiArc/languages/CNNTrainLang/-/issues/1Language is backend dependent2018-08-01T11:12:09+02:00Evgeny KusmenkoLanguage is backend dependent@svetlana.pavlitskaya @carlos.alfredo.yeverino
The generator is backend dependent now, i.e. the language contains an mxnet generator.
Since we want to extend the backend support to more backends, the train generator should not reside ...@svetlana.pavlitskaya @carlos.alfredo.yeverino
The generator is backend dependent now, i.e. the language contains an mxnet generator.
Since we want to extend the backend support to more backends, the train generator should not reside in the TrainLang project any more.
Please keep an abstract generator interface in CNNTrainLang and move the MXNetTrain generator to CNNArch2MxNetSvetlana PavlitskayaSvetlana Pavlitskaya