nntrainer
nntrainer copied to clipboard
NNtrainer is Software Framework for Training Neural Network Models on Devices.
In the current implementation, concat and split layer have to incur a lot of memory copy. However, in many cases, they can be no-op if the memory allocated for their...
Given a model, with input and labels/output nodes name, `graph.initialize()` should only keep nodes that are relevant for generate the output from the output nodes given the input in the...
As time goes by, there are some unused methods that will likely that never going to be used later. (eg) private helper methods that is no longer being used. Shorter...
How to run? build the code Download the Dataset You can download the dataset used from https://github.sec.samsung.net/21-Tizen-ME-SRI-D/On-Device-Learning/tree/master/Project/Source/nntrainer/Applications/SimpleShot/tasks and move to app_path+"/tasks/" After building, run with $./executable model method train_file validation_file...
During an iteration of forwarding/backprop, we know that certain values are not going to be used for long. Such values can be offloaded to the disk and cache later right...
The embedding layer is essentially a hash table where certain elements are selected based on the input. **Current design**: load all the elements to the memory and select those elements,...
layer run-context should perform caller checks before providing access to tensors. For example, accessing a tensor requested with FORWARD_FUNC_LIFESPAN in calcDerivative() must throw.
Below are the listed faulty unittests which were passing because of the unittest bugs and must be resolved soon: - nntrainer_FullyConnectedLayer_TFmatch.DISABLED_forwarding_backwarding_00_p - nntrainer_Conv2DLayer.DISABLED_backwarding_01_p - nntrainer_Conv2DLayer.DISABLED_backwarding_02_p - nntrainer_Conv2DLayer.DISABLED_backwarding_03_p - nntrainer_Conv2DLayer.DISABLED_backwarding_04_p -...