[Tensor Pool] Add expose/persist concept

This patch add expose/persist concept to the tensor pool and manager.

When a tensor is exposed, this means that the tensor is gauranteed to
remain valid max_exec where max_exec is the value passed along allocateTensors(max_exec);

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[ExecOrder] Add exec order control and fix inference

This patch add exec order control and reduce memory overhead of saving
inference result

exec order cotrol has two part

1. Additional exec order. This enable graph to add some exec order.
2. Expose bit. This enables request to be visible after the end of the
execution.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Network Graph] Split max_exec_order

This patch split max_exec_order to three part.

1. graph_exec_end: graph execution end
2. backward_iter_end: end node of backward operation
3. forward_iter_end: end node of the forward operation

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[manager] Separate backwarding -> grad, deriv

This patch separate backwaring to grad and deriv for the requestTensors

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Debug] Add a naive validator to the optimized planner

This patch add a naive validator to the optimized planner. This doubles
the memory consumption. So kept commented but when time comes, this
function can be enabled.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Fix] fix lifespan of recurrent cells

This patch fixes lifespan of recurrent cells

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Fix] change max order to point last

This patch fixes bug that gradient clipping was not applied correctly
when optimization is on.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Fix] BN life span

t_reduced is used in forward, but lifespan was in backwarding. This
patch fixes the issue

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[compat] remove contrib from tf headers

This patch remove contrib path from tflite headers as contrib headers is
being removed.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Fix] Non-const reference

This patch fixes returning reference of local variable.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Android] Delegate option control to android.mk

This patch delegates option control in android mk to meson for debug and
optimized build

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[ Network Graph ] Temporal Fix in getInputGrad

When we turn on the inPlaceOptimization, it will fail if we have
mulitiple inputs: input layer ( can be inplace operatation ), normal
layers. Then, we are not allocate the grad between input layer and
this layer and it cause the error during calcDeriviate which requires
the grad tensor of input layer.

This patch fix this temporally by creating tensor buffer it reqruies.

and it includes modification of mem_check script to generate proper
output.

**Self evaluation:**
1. Build test:  [X]Passed [ ]Failed [ ]Skipped
2. Run test:  [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: jijoong.moon <jijoong.moon@samsung.com>

[grucell] enable multi inout

 - Enable multi inout for grucell
 - Generate grucell layer/model unittest

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[grucell] refactoring grucell

 - Rename grucell variables
 - Uncommented genModelTests

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[Dropout] Disable test strong match

As dropout is statistcally matching strong output of 60%, it turned out that
it is true only sometime, diabling dropout strong match

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[Recurrent] Add input seqeuncing mechanism

As some kind of recurrent model requires input layer it self to be
changed along the sequence, this patch proposes a simple mechanism to
add time stamp suffix for the input layers as well

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[lstmcell] support multi in/out

 - Refactoring lstmcell layer to support multi in/out (3 input / 2 output)
 - Regenerate lstmcell testcase for multi in/out

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[bugfix] zoneout lstmcell regenerate mask

 - Current implementation regenetate the zoneout mask in calcGradient.
    Fix it to reuse the zoneout mask and remove renerating the mask.

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[zoneout lstmcell] support multi in/out

 - Refactoring zoneout lstmcell layer to support multi in/out (3 input / 2output)
 - Regenerate zoneout testcase for multi in/out

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[Update] Multiout model golden test

This patch generate multiout model golden test

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[Output] Return zero grad for empty output

Along this patchline implements returning zero grad for empty output

Particualary in this patch...

 - Remove manager::requestOutput
 - Use requestTensors instead of requestOutput to request outputs

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Clean] Remove getOutputDimensions()

This patch removes initContext::getOutputDimensions().

This function is only used in the test so removed
(It is used in network_graph but soon will be substitued)

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[context] Receive out conn info and use

Init Context receive out conn info (true if conn info exist, false if
not) and use it to determine given output is dangled

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Manager] update requestInputs to use requestTensor

This patch update requestInputs to use requestTensor_ as a bridge.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Context] out_dim -> out spec

This patch prepare migration to tensor spec v2 by substituting out
dimension to out specification inside layer_context

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Tensor] OutputGrad defaults to be zero if not given

This patch creates output grad(incoming) to all zero if the given output
is not trainable. This will make the given output be considered as
constant. If a user wants to check if output is constant-like(having
zero gradient as it's partner), she can easily check if
outputHasGradient

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Const] Make incoming derivative const

As incoming derivative shall not be changed by the layer, this patch
changes runcontext::getIncomingDerivatives and runcontext::getOutputGrad
to return const.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Test] Add test with dangling output

This patch add test with dangling output, this test passes test but
fails optimized run.

Current aim is to make this pass optimized run as well.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[lstm] refactoring lstm layer

 - Refactoring lstm layer to use lstm cell core functions.

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[zoneout lstmcell] refactoring zoneout lstmcell layer

 - Refactoring zoneout lstmcell layer to use lstmcore functions.
 - Preserve lstm_cell_state tensor for calcGradient.
 - Remove lstmcell core layer

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[lstmcell] refactoring lstmcell layer

 - Refactoring LSTMCell layer to use LSTM core functions.

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[lstmcell core] prepare refactoring lstmcell core layer

 - LSTM cell core layer will be refactoring from layer based to function based.
   This commit prepare the core functions which is not used right now.

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[Recurrent] Support connection for as_sequence

This patch enables recurrent realizer to consume support connection
residing in as_sequence parameter

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[RecurrentRealizer] Modify to have connection

This patch modifies recurrent realizer to have connetion

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Realizer] Change input connection semantics

Input connection semantic changed to be more intuitive by mapping
connection one to one

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Realizer] Change slice realizer to get connection

This patch change slice realizer to get connection

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Semantics] InitContext NumOutput is hint

This patch changes init context num output to be a hint how many outputs
has to be allocated. Layers may or may not depend on this information.
What Layer must gaurantee is that number of acutal outputs must be
bigger than the init context num output requested, this fixes dangling,
unused variable issue

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Spec] Change nntrainer install dir

This patch change nntrainer install dir to /usr/prefix/*,
see also https://github.com/nnstreamer/nnstreamer/issues/3560

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Identity] Add identity layer test

This patch add identity layer test

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Identity] Implement and connect identity layer

THis patch implement and connect identity layer

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Identity] Add identity header

This patch add identity layer header.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Graph] set trainable only if trainable

This patch fixes bug that needsCalcGradient is set true when it
shouldn't be

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[layer] MAE layer bug fix

MAE layer bug fix:
- forwarding fixed to propagate input unchanged
- backwarding updated to scale with size of the tensor

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[optimizer] no request vars for non-trainable weights

optimizer is requesting variables for non-trainable weights, but it
should not.
This patch provides the corresponding fix.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[tensor] Tensor pool request bug fix

If the max execution order for the tensor pool was less than the
largest execution order for a tensor in the tensor pool, then the max
execution order was being included in the execution order of that
tensor.
This patch resolves this issue for end order, as well as for the start
order.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[Tensor] Add tensor cat

This patch add tensor cat method and it's test

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[tensor] Add tensor::split

This patch add tensor split for wider use

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Trivial] Tensor save changes /ofstream/ostream

There is no reason to limit saving to ofstream. Thus it is changed to
ostream

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[API] Add layer visitor for the model

This patch add visitor for the model api

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[gru] enable bias_hh, reset_after

 - Enable bias_hh in gru.
 - Enable reset_after in gru, grucell. If reset_after is set to true,
   apply reset gate after matrix multiplication.

close #1768

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[loss] optimize mse part 3

optimize mse layer to use scaled l2norm instead of manually calculating
the l2norm.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[loss] optimize mse part 2

optimize mse backwarding calculation by merging multiple and divide into
a single call.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[loss] optimize mse part1

reduce memory usage for mse forwarding.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[grucell] enable 2 bias

 - Enable bias_hh in grucell

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[lstm, lstmcell, zoneout] enable 2 bias

 - Enable bias_hh in lstm, lstmcell, zoneout lstmcell layers

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[rnn] enable 2 bias

 - Enable bias_hh in rnn

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[rnncell] enable 2 bias

 - Make a integrate bias property. It decide whether integrate 2 bias to 1
   or not. It will be used in rnn variant for now.
 - Added bias_hh in rnncell

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[test] enable zoneout lstm test

enable zoneout lstm test

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] bug fix for array initializer

Bug fix for array initializer storing the indices of the requested
weight so that accessing weights not requested always causes error.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[Connection] Enhance parsing logic

This patch enhances parsing logic of connection especially related to
indexing

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Cc: Parichay Kapoor <pk.kapoor@samsung.com>
Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Trivial] Complement connection doxygen

Add missing parameter explanation in doxygen

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[node/trivial] Change input layers to input conns

This patch chagne input layers local variable name to input connections to
make it consistent

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[manager] Remove activation input exec order from backwarding

This patch removes activation input exec order from backwarding as input
of the activation layer is not used in the backwarding.

This leads to change in the unittests as we cannot check all the
outputs, especially close to the end of the model. So, with optimization
enabled, only the output layer's forwarding is checked.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Disable bias enabled for fc and conv

This patch enables disable bias for fully connected and convolution layer.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Add disable bias property for layer impl

Add disable bias property for the layer impl class.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[test] Disable zoneout lstm unittests

Disable zoneout lstms unittests till they are fixed.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[test] recoder to read from input from file

support recorder to read data from file than always generating random
data.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[model] Add clip gradient norm property

Add clip gradient norm property for the the model as a simple interface
to set for all the layers.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Embedding fix

Undo earlier changes due to bug in them.
The changes will be proposed again with more unittests.
The changes have been commented for now.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] zoneout lstm bug fix

zoneout lstm bug fix for handling the scenario where mask rate is set to
zero for either hidden or cell.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[realizer] bug fix for activation realizer

Bug fix for activaiton realizer to find the realizer where the layer is
looping to itself with name change.
TODO: also do for flatten and other realizers which modified the graph
in similar fashion.

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Fixes for mol attention layer

- gradient accumulation
- support only 1 output in case state update is not used
- bug fix in backwarding
- activation to work out of place

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[realizer] recurrent realizer bug fix in concat

concat to not loop over all inputs multiple times.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[hotfix] Enable zoneout lstmcell unittest

 - Fix broken zoneout lstmcell unittest

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[layer] Bug fix for embedding layer

This patch adds bug fix for the embedding layer related to the index of
the data.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[Recurrent] Support connection for recurrents

This patch support connection for recurrent input, outputs

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Recurrent] recurrent using zero connection

This patch implements multi inout recurrent using zero connection

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Recurrent] Support multiple sequence

This patch updates recurrent realizer to suupporting multiple sequence with layer name not boolean
with `as_sequence` property

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Test] Update multiout support specification

This patch updates multiout support specification.

Patch planned: 1. support return sequence with layer name
               2. support multiple input, output layers
               3. support multiple input, output connections

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Clean] Prepare remap realizer for further changes

This patch add comments and terms(idx -> time_idx) inside remap realizer
to prepare further changes

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[unittest] zoneout lstmcell unittest

 - unittest for zoneout lstmcell layer

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[zoneout lstmcell] Implement zoneout lstm cell

 - Zoneout lstmcell is based on the paper and the github repo
   which is mentioned in paper.
 - Todo: Zoneout at inference time is not implemented yet.

refer: https://arxiv.org/pdf/1606.01305.pdf
       https://github.com/teganmaharaj/zoneout

Self evaluation:

    Build test: [X]Passed [ ]Failed [ ]Skipped
    Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[lstmcell] Refactoring the lstmcell

 - Refactoring the lstmcell to lstm core layer.
   This lstm core layer will be used in zoneout lstmcell layer
 - lstm core layer is designed to have 3 inputs, 2 outputs
   like other framework.

Self evaluation:

Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

[Fix] Add dstate to mol attention batch

This patch fixes that dstate does not have the batch size updated

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[graph] Extend lifespan of model outputs

The lifespan of model outputs is extended to the max forward exec order
so that the outputs remain valid till the full forward is completed and
they can be checked in the training or even inference.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] fix for mol attention layer

Bug fix for mol attention layer as getOutputDerivatives is not available
in calcGradient, so the usage is replaced with a temporary tensor.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Support multiple outputs for mol attention layer

Support multiple outputs for mol attention layer along with the
unitests.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Support filter masking in mol attention

Add support for filter based masking in mol attention.
Add corresponding unittest.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[Fix] Setting batch after finalize/initialize

This patch delays setting batch after finalize/initialize.

Setting batch must be done after runcontext has made, the reason is that
we have semantics that finalize should be independent of batch size, but
if batch size is set before output_dims batch must be set according to
input_dims batch, which is not desirable.

**V2**
Fix hidden bugs related to this

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Test] Add multiout tests

This patch add multiout tests

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[Fix] fix inconsistency when saving a model

When saving a model, if added order is different from sorted order,
loading a model with binary breaks some times

eg)
a -> b
  \
   v
   c

and layer was added a, b, c
it is being sorted to a, c, b

When this is saved, it is saved a, c, b.
When loading the graph from saved file it is loaded a, b, c.

This patch fixes this inconsistency happens when saving a model.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[graph] Use connection info for outputs

This patch enable connection infromation to inbound connection starting
from output while deprecating some unused methods

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[graph] implement setOutputConnections

THis patch implement setOutputConnections. Now, every connection has
defined place to be, we can pin point where the connection has to go.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[tidy] Remove getInputLayers usage

This patch removes getInputLayers usage as connection is now defined by
name + index, getInputLayers poses ambiguity.

**Changes proposed in this PR:**
- Also, connection now has hash function defined.

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[nn] Apply input connection to compile/initialize

**Changes proposed in this PR:**
- input connection is used for previous realizer/graph.initialize
- temporary string cast operator deleted

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[nn] InputLayer->InputConnection

This patch substitue input layer to input connection while removing
props::InputLayer

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[props] Extract connection

This patch extract connection from common_properties to have more room
to handle connections freely

**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>

[layer] perform layer context check on first forward

This patch enables layer context check on the first forward itself,
which revealed a bug in forward which was earlier being shown in the
calcGradient in mol attention layer.
Added the corresponding fix for mol attention layer.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Add MoL layer unittest

Add MoL layer golden test.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

[layer] Add full support of MoL attention layer

Add MoL attention layer backwarding, exportTo, setbatch member
functions.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>