nntrainer

nntrainer copied to clipboard

Add identity layer

1

Now we are in need of identity layer(do nothing flow everything). This is specifically to solve the issue addressed here https://github.com/nnstreamer/nntrainer/commit/e76a186c8b4a7bea1d8d30ffd9f208cfd2ec78e8#r62007856 ``` A -> outputs ->(a0, a1) B -> outputs...

zhoonit

Support zoneout LSTM

2

Support zoneout LSTM - https://arxiv.org/abs/1606.01305

kparichay

Tizen 7.0 Items

4

1. Checkout delayed items from #1033 2. related to #1516, add support for `ml_train_set_compile_properties()`, `ml_train_set_run_properties()` officially to overcome var_args limitation. documentation 1. ini inference default behavior https://github.com/nnstreamer/nntrainer/pull/1711#discussion_r755199205

zhoonit

Learning Rate Scheduler support

6

Support for an independent learning rate scheduler interface is needed. - [x] create learning rate scheduler interface - must be easy to support custom scheduler #1776 - [x] support existing...

kparichay

Support sharing labels for loss layers

5

Sharing a given label with the loss layers of the model is not fully supported. - when only a single label is given, it is shared with all the labels...

kparichay

Make layer plugin directly work with ml api.

2

layer plugin developer has some burden to create and use the layer right away. let's assume there is a custom layer ```cpp class MyLayer : public nntrainer::Layer { /** define...

zhoonit

Expose custom optimizer to the public.

1

docs and example how to use custom optimizer to the public

zhoonit

Please add a test case of ML-Inference-Single::NNTrainer

2

https://github.com/nnstreamer/api/blob/95faa11b553861a42cedb4cae441f87a9ec353ce/c/src/ml-api-inference-single.c#L833-L853 ```C if (nnfw == ML_NNFW_TYPE_NNTR_INF) { if (!in_tensors_info || !out_tensors_info) { if (!in_tensors_info) { ml_tensors_info_h in_info; status = ml_tensors_info_create (&in_info); ////////// 1. if (status != ML_ERROR_NONE) { goto error;...

myungjoo

Broken invariant: loss layer always requires label

1

For the model which is using reparameterization trick (eg, variational auto encoder) does not take label but calculates loss. loss should not determined by `requireLabel()` but solely by `outconnection ==...

zhoonit

Enable contiguous check on add_i

1

For now checking the tensor which is contiguous or not is disabled since the add_strided does not support broadcast. Enable this one after support broadcast operation and then replace add_i...

lhs8928