Parichay Kapoor [Tue, 1 Dec 2020 12:07:49 +0000 (21:07 +0900)]
[weight] Updated weights to be vector
Updated weights of layer to be vector than a shared_ptr array
This is for easier management and updating weight internally when
gradient will share the memory
See also #774 #766
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 1 Dec 2020 11:14:57 +0000 (20:14 +0900)]
[manager] Added nntrainer manager for weights
Added manager to manage all the allocated weights
This patch also adds manager to the model and passes manager to the
initialize which allows weights to be added to the manager.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 1 Dec 2020 10:46:38 +0000 (19:46 +0900)]
[weight/var_grad] Make internal variable as shared_ptr
Internal variables in weights/var_grad, namely, the variable and gradient itself
are changed to shared_ptr so that weights can be shared without worrying about
shallow copies.
Also changed the copy constructor to not create new Tensor as copy constructor
of weight will get called and its unnecessary + unintentional overhead.
As weight is just wrapper over tensor, their copy constructors should follow
same behavior as tensor which is to not create new memory.
Added clone as an alternative to create new copy of a given weight.
See also #774 #766
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
jijoong.moon [Tue, 1 Dec 2020 04:31:03 +0000 (13:31 +0900)]
[ CONV2D ] seperate conv2d_gemm and im2col
It is better to split conv2d_gemm and im2col
**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>
Parichay Kapoor [Wed, 2 Dec 2020 05:29:19 +0000 (14:29 +0900)]
[unittest] Enable disabled unittest
Enable fc layer disabled unittest
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 1 Dec 2020 01:48:32 +0000 (10:48 +0900)]
[var_grad] Trainable inferred from gradient
Trainable property of a variable was earlier inferred by storing a trainable variable
Now, trainable will be inferred using gradient.uninitialized()
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Mon, 30 Nov 2020 05:52:20 +0000 (14:52 +0900)]
[tensor] Update tensor operation signature
Update tensor operation signature to return Tensor reference as a retval
than a tensor itself. This avoid creating dummy tensors as a return (which might have been
optimized by the compiler but lets do manually as the input is also a reference).
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Mon, 30 Nov 2020 07:34:09 +0000 (16:34 +0900)]
[CustomLayer] Update readme.md
Add readme.md about how to run and expected output
**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [ ]Passed [ ]Failed [X]Skipped
Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>
Jihoon Lee [Thu, 26 Nov 2020 07:50:53 +0000 (16:50 +0900)]
[Custom] Add actual example
**Changes proposed in this PR:**
- Add an example to create the custom layer to be used with ini
- Add an example to create the custom layer to be used with 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>
Jihoon Lee [Wed, 18 Nov 2020 02:30:40 +0000 (11:30 +0900)]
[Custom] Add an example scaffolding
Add a layer example that depends on the user's custom code
This patch generates scaffolding to the `Application/Custom` folder
**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>
jijoong.moon [Fri, 27 Nov 2020 01:12:45 +0000 (10:12 +0900)]
[ Graph ] remove grad mem buffer for backwarding
This PR includes,
. remove grad memory buffer in n_buffes for graph. We do not need
this because we could use var memory buffer of n_buffers to
backwarding.
. For MNIST, memory consumption is reduced 3.5 to 2.6
**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>
Jihoon Lee [Fri, 27 Nov 2020 08:30:47 +0000 (17:30 +0900)]
[ModelLoader] Use vector<string> when create layer
When creating a layer from an ini, enum based properties were used.
This prevents adding a new properties without changing the api header.
This patch moves to setting layer properties to vector<string>
to enable setting properties without changing the api header, eventually
enabling custom properties in custom layer.
**Semantics Change propesed in this PR**
Ini won't ignore the properties that is not supported since model_loader
would not know if it is supported or not
See also #716
**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>
Parichay Kapoor [Wed, 25 Nov 2020 11:45:09 +0000 (20:45 +0900)]
[bnlayer] bug fix for inference
Batch normalization bug fix for inference mode
when add() was used instead of add_i()
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 25 Nov 2020 11:43:11 +0000 (20:43 +0900)]
[tensor] Support multiply/divide with given output
Support multiply/divide with given output tensor
This reduces temporary allocations for bn layer
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 25 Nov 2020 11:41:54 +0000 (20:41 +0900)]
[pooling] Reduce temporary mem allocations
Reduce temporary memory allocations for pooling
Remove unnecessary temporary memory allocations which can be
replaced with a slice view
Also removed unnecessary setting memory to zero
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 25 Nov 2020 11:40:32 +0000 (20:40 +0900)]
[activaiton] Reduce temporary memory alloc
Reduce temporary memory allocations by activation layer
by using the hidden and ret_derivative class variables
This temporarily increases peak memory but alloc-dealloc is removed
from every iteration
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 25 Nov 2020 11:39:27 +0000 (20:39 +0900)]
[regex] Make regex static const
Make regex static const
Although it is using static string, that memory is always being allocated inside regex
Making is static const only makes it once for the function lifetime
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Mon, 16 Nov 2020 07:31:00 +0000 (16:31 +0900)]
[neuralnet] Skip backwarding for non-trainable layers
This patch skips the backwarding for the non-trainable layers.
Further, the last trainable layer skips calcDerivative as well.
This results in much fewer calculations as well as more importantly, reduced memory.
See also #732
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Mon, 16 Nov 2020 06:36:49 +0000 (15:36 +0900)]
[Layer] Add built-in ops to the context
**Changes proposed in this PR:**
- Add default layers to the global 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>
Jihoon Lee [Mon, 16 Nov 2020 06:59:20 +0000 (15:59 +0900)]
[AppContext] Fix key is case sensitive
In current semantics, type key should be case insensitive however case
was sensitive, 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>
Parichay Kapoor [Fri, 20 Nov 2020 07:04:49 +0000 (16:04 +0900)]
[conv2d] More optimizations for conv2d
This patch provides more optimizations for conv2d
by avoiding more memcopies and operations along with modification
to internal interface of conv2d_gemm operation
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Fri, 20 Nov 2020 05:46:37 +0000 (14:46 +0900)]
[conv2d] Bug fix for regularization loss
Regularization loss for conv2d layer took average over output filters
than adding it up. This patch fixes it.
See also #761
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Fri, 20 Nov 2020 05:37:55 +0000 (14:37 +0900)]
[conv2d] Refactor conv2d layer
Conv2d layer has some issues #761
This patch addresses some of them:
- Weight is now independent of the filter size. Different filter
weights have now been combined. This has resulted in easier addressing of weights
- Above combining of weights also reduced many mem-copies of weights to bring it in a particular shape
- Moved to use getSlice() to get access to some data than creating copies
Now, all layers have fixed number of weights.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Fri, 13 Nov 2020 02:51:12 +0000 (11:51 +0900)]
[AppContext] Register Default ops at the begining
**Changes proposed in this PR:**
- Register default optimizer at the beginning of the load
**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>
Jihoon Lee [Wed, 18 Nov 2020 02:44:25 +0000 (11:44 +0900)]
[Deps] Remove openmp dependency
Openmp is no longer used. It is deleted to reduce memory consumption
**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>
Parichay Kapoor [Mon, 16 Nov 2020 07:00:40 +0000 (16:00 +0900)]
[layers] Split backwarding into smaller functions
Split layer backwarding into smaller functions for optimization purposes
- calcDerivative() - calculate the derivative to be passed to previous layers
this function must be implemented by all derived layers
- calcGradient() - calculate the gradient for the weights of the layer
- applyGradient() - apply the gradients to the weights of the layer
Also, now input->backwarding() throws than just silently not doing anything.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Fri, 13 Nov 2020 06:41:08 +0000 (15:41 +0900)]
[var_grad] Add var_grad for input/output lists
Added var_grad for input/output lists which also combines derivatives
This is the baseclass for the weights
This update will help with graph class
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Wed, 11 Nov 2020 12:30:23 +0000 (21:30 +0900)]
[AppContext] Add registerer,invoke factory methods
**Changes proposed in this PR:**
- Add factory registerer
- Add factory invoker
- Register built-in objects to each layers(postponed)
- Add 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>
jijoong.moon [Fri, 20 Nov 2020 07:39:23 +0000 (16:39 +0900)]
[ GRAPH ] Remove unused function and add doxygen note
In neural network class, there is fucntions which should be moved to
graph.
In this PR, remove member functions which is not used any more and add
doxygen comment in graph header.
**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>
jijoong.moon [Fri, 20 Nov 2020 06:28:17 +0000 (15:28 +0900)]
[ ANDROID ] Enable graph for andoid build
Fix Android.mk to support graph
**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>
jijoong.moon [Fri, 20 Nov 2020 05:22:03 +0000 (14:22 +0900)]
[ GRAPH ] Split initilization & Assign Memory
Split ini and mem assignment
**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>
jijoong.moon [Fri, 20 Nov 2020 02:35:18 +0000 (11:35 +0900)]
[ NNSTREAMER ] Fix NNStreamer Filter for graph
Describe a commit content (Until 80 colums per line) in detail ASAP.
**Changes proposed in this PR:**
- Added TOC generator for README.md
Resolves:
**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>
jijoong.moon [Fri, 20 Nov 2020 00:58:43 +0000 (09:58 +0900)]
[ NNSTREAMER FILTER ] Fix nnstreamer filter to support graph
Describe a commit content (Until 80 colums per line) in detail ASAP.
**Changes proposed in this PR:**
- Added TOC generator for README.md
Resolves:
**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>
jijoong.moon [Fri, 20 Nov 2020 00:51:04 +0000 (09:51 +0900)]
[ Fix ] istrequal to check length of string
fix istreaqual
**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>
jijoong.moon [Thu, 19 Nov 2020 09:53:31 +0000 (18:53 +0900)]
[ GRAPH ] Support Backbone Network
This PR includes :
. Modification of Network Graph to enable Backbone network support
. Fixes for unittest
**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>
jijoong.moon [Thu, 19 Nov 2020 00:43:17 +0000 (09:43 +0900)]
[ GRAPH ] Add Compiled Variable
- In order to make sure to run initialize after success of compile(),
compiled variable is used.
- additional unittest cases are fixed.
**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>
jijoong.moon [Wed, 18 Nov 2020 07:20:40 +0000 (16:20 +0900)]
[ UNITTEST ] Fix unitest & Applications to support NetworkGraph
Unittest and Applications need to be fixed to support NetworkGraph.
**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>
jijoong.moon [Tue, 17 Nov 2020 12:19:27 +0000 (21:19 +0900)]
[ UNIT TEST ] Enable Unit test for graph
Because of graph implementation, Unittest must be changed.
In this PR, Enabling Unit Test & Bug Fixes are included.
**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>
jijoong.moon [Mon, 16 Nov 2020 05:39:57 +0000 (14:39 +0900)]
[ Graph ] Add Graph into Network
Describe a commit content (Until 80 colums per line) in detail ASAP.
**Changes proposed in this PR:**
- Added TOC generator for README.md
Resolves:
**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>
jijoong.moon [Fri, 13 Nov 2020 02:13:51 +0000 (11:13 +0900)]
[ Graph ] Modify backwarding/forwarding to use graph data
In this PR, Using graph data is eanbled.
. Each layer use grap data instead of input/hidden tensors.
. Not any more capy between layers
**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>
jijoong.moon [Mon, 9 Nov 2020 13:03:39 +0000 (22:03 +0900)]
[ GRAPH ] Initialize graph
This PR includes,
. Set the Network Buffer
. Initialize Layer
. Calculate Dimension
. Assign Network Buffer at each layer properly
**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>
jijoong.moon [Mon, 9 Nov 2020 01:10:01 +0000 (10:10 +0900)]
[ LAYER ] Expose Output Layer
It is not enough to support various connection in network if we
support output layer only in implicitly. This PR includes eanbling
output layer explicitly.
**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>
jijoong.moon [Mon, 9 Nov 2020 01:08:06 +0000 (10:08 +0900)]
[ Example ] Mini Resnet
This PR includes,
. Addintion of Mini Resnet Example to test
. Skip connection
. Ouput Layer
. Addition Layer
**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>
jijoong.moon [Wed, 21 Oct 2020 02:42:22 +0000 (11:42 +0900)]
[ Graph ] Neural Network Graph
In this PR, NetworkGraph Class is introduced and compile method is
implemented.
During the compile process, it takes layers vector which is created by
model_loader and create required layers like Activation, Flatten,
Concat or Addition Layers. Also it modify inputs relatively and
Calculate the order of calculation. So far, only serial processing is
supported.
**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>
Parichay Kapoor [Fri, 13 Nov 2020 13:01:20 +0000 (22:01 +0900)]
[fc,tensor,adam] Change signature for sum
Added new signature for sum
Also removed unnecessary calculation from fc
Optimzed adam calculation to reduce extra memory requirement
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Fri, 13 Nov 2020 12:17:56 +0000 (21:17 +0900)]
[tensor] Reduce overall memory overhead
Reduce overall tensor memory overhead
- standardization and normalization now take in-place
- input and label tensor in train are now allocated only once externally and reused for each epoch
that being used in each epoch
- add_i does not allocate new memory now
- removed support for normalization and standardization for conv layer
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 11 Nov 2020 10:49:18 +0000 (19:49 +0900)]
[ccapi] Direct methods to create layers
Added syntactic sugar constructors for layers and loss
This allows creating various layers and losses directly with c++ API
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 11 Nov 2020 07:45:05 +0000 (16:45 +0900)]
[ccapi] Syntactic sugar constructors
Added syntactic sugar constructors for optimizers
which allow making optimizer closer to existing API
and more readable.
See also #734
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Thu, 12 Nov 2020 11:51:53 +0000 (20:51 +0900)]
[tensor] Update tensor signature for apply
This patch updates tensor signature to include output for apply
Correspondingly, activation layer and loss layer have been updated.
More changes in the patch
- Convolution layer has been updated to reuse its output derivative.
- added a getBatchSlice() which provides a slice of the original tensor by batch
and avoids creating a copy everytime
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Thu, 12 Nov 2020 05:23:24 +0000 (14:23 +0900)]
[tensor] Update tensor op signature for dot
Update tensor op signature for dot
Further, fc layer has been updated based on this signature
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Wed, 11 Nov 2020 09:43:35 +0000 (18:43 +0900)]
[Optimizer] Add enum type factory
This patch adds integer factory to align with capi
**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>
Jihoon Lee [Wed, 11 Nov 2020 02:59:20 +0000 (11:59 +0900)]
[Model] Apply appcontext
Apply appcontext to NeuralNetwork. From this patch `chdir()` hack is not
needed :)
**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>
Parichay Kapoor [Wed, 11 Nov 2020 12:14:25 +0000 (21:14 +0900)]
[model] Handle loss layer to be added from user
Handle loss layer to be added from the user which is created with c++ API
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Mon, 9 Nov 2020 03:16:48 +0000 (12:16 +0900)]
[AppContext] Add AppContext
This patch add basic app context with setting current working directory
**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>
Parichay Kapoor [Tue, 10 Nov 2020 04:57:47 +0000 (13:57 +0900)]
[docs] Update docs about backbone features
Update documentation for the ini with the newly added backbone features
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 10 Nov 2020 02:23:27 +0000 (11:23 +0900)]
[backbone/ini] Support subgraph with ini backbone
Added support for subgraph of a given ini backbone
Added corresponding unittests as well
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Mon, 9 Nov 2020 06:21:29 +0000 (15:21 +0900)]
[backbone] Add ini backbone properties
Added ini backbone properties :
- scaleSize - scale the size of the model
- preload - load the weights of this backbone model before adding to the model
preload has issues which requires the format of the model file to update
wait for #361
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Tue, 10 Nov 2020 01:03:00 +0000 (10:03 +0900)]
[Layer] Change layer type to string
This patch changes layer type to string
**Changes proposed in this PR:**
- Add Layer::getType() and Layer::type
- Add `istrequal` to the `parse_util` for case insensitive compare
- Test changes accordingly
**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>
Mete Ozay [Wed, 11 Nov 2020 12:05:19 +0000 (12:05 +0000)]
Update README.md
- Update links to examples
- Update Getting Started and Running Examples
- Fix grammar and typo.
Signed-off-by: Mete Ozay <meteozay@gmail.com>
Parichay Kapoor [Wed, 4 Nov 2020 04:00:11 +0000 (13:00 +0900)]
[nnstreamer/backbone] Update to support more backbone
Update nnstreamer backbone layer to support more backbones than just tflite
Wait for this PR to merge till the bug fix on nnstreamer
https://github.com/nnstreamer/nnstreamer/pull/2850 is merged and reflected
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 10 Nov 2020 12:23:15 +0000 (21:23 +0900)]
[cifar] Update cifar application with backbone
Update cifar application using databuffer with generator to
use tflite backbone.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Thu, 5 Nov 2020 10:48:24 +0000 (19:48 +0900)]
[Optimizer] Change enum type to string
Change optimizer type to string
**Changes proposed in this PR:**
- Bullet points are okay, too
**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>
Jihoon Lee [Fri, 6 Nov 2020 05:55:36 +0000 (14:55 +0900)]
[Fix] Throw when read/save fails
Currently, read save didn't throw when failing on those operation.
From this patch, if an object fails to read, it throws an exception
to preven false positive on test result
**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>
Jihoon Lee [Wed, 4 Nov 2020 11:31:03 +0000 (20:31 +0900)]
[Test] Add conv2d Integrated test
This patch add conv2d integrated test
Model test passes, layer test failes due to bug fix
**Changes proposed in this PR:**
- add conv2d integrated test/generator
- add transLayer for channelfirst <-> channel last conversion
**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>
Jihoon Lee [Wed, 4 Nov 2020 11:25:26 +0000 (20:25 +0900)]
[Conv] Fix convlayer backwarding
There was missed calculation doing conv2d backwarding.
This patch fixes the issue.
With this patch, some test fails. I'll take a look and fix accordingly
within this commit
**Self evaluation:**
1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [ ]Passed [X]Failed [ ]Skipped
Signed-off-by: Jihoon Lee <jhoon.it.lee@samsung.com>
jijoong.moon [Mon, 9 Nov 2020 13:01:40 +0000 (22:01 +0900)]
[ POOLING ] Fix global pooling dimensions
For the global pooling, we need to set the output dimension on last
dimension, width.
**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>
Parichay Kapoor [Fri, 6 Nov 2020 04:41:32 +0000 (13:41 +0900)]
[backbone] Documentation for backbone
Added documentation for backbone in ini-configuration
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
jijoong.moon [Thu, 5 Nov 2020 12:39:20 +0000 (21:39 +0900)]
[ Layer ] Add Output Layer to support multiple output
Currently there is no way to support multiple output.
However it is not neccessary to be explicit support.
Instead if user use "output_layers = layername0, layername1", then nntrainer
automatically add output layer. ( comparing ouput dimension and input
layer's dimension, we could decide it is addition or split operation. )
Also, if certain layer is specify as an input layer from mutiple
layers, then it should be added implicitly.
TODO: support split operation
**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>
hyeonseok lee [Fri, 6 Nov 2020 06:21:18 +0000 (15:21 +0900)]
[README] Modify the image url
Modify the url from absolute path to relative path
Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>
Parichay Kapoor [Thu, 5 Nov 2020 04:55:13 +0000 (13:55 +0900)]
[bn layer] Support non-trainable mode
Support non-trainable mode for bn layer when the parameters
for this layer are not updated.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Mon, 2 Nov 2020 11:21:09 +0000 (20:21 +0900)]
[Test/Py] Add translayer
Add `translayer` to encapsulate keras layer to use nntrainer layout
This can be used for `genInput.py` 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>
jijoong.moon [Fri, 6 Nov 2020 02:59:07 +0000 (11:59 +0900)]
[ Util ] Add how to generate cifar 10/100 bmp images
Add document explain how to generate cifar 10/100 bmp images
**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>
Parichay Kapoor [Fri, 6 Nov 2020 02:28:09 +0000 (11:28 +0900)]
[README] Bugfix for image display
Add bugfix for readme where images were not being shown
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 4 Nov 2020 02:27:53 +0000 (11:27 +0900)]
[debian/dist] Updated debian packaging of files
Updated debian packaging to limit the files which are packaged and exposed.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 4 Nov 2020 02:25:58 +0000 (11:25 +0900)]
[restructure] Restructure the core files
This patch restructures the internal files
include and src folders are replaced with more relevant and clustered folders
headers and souces now live together
Also the headers exposed in the packaging are severely limited than exposing
all the headers. Updated for Android and Tizen packaging as well.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Fri, 7 Aug 2020 08:00:19 +0000 (17:00 +0900)]
[Delegate] Add delegate support header
Add delegate support header
This supports settings backend and device
Added some properties but they are experimental
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Thu, 29 Oct 2020 06:38:39 +0000 (15:38 +0900)]
Update application to use backbone
Update transfer learning application to use backbone
Now feature extractor has been removed from the application
and dependency on tflite is also removed
However, the caching of features from feature extractor is not yet supported.
This results in the application to be slow to run full 1000 epochs on gbs
Till caching is supported, application test from gbs build is removed
Also fixed android packaging for nntrainer with tflite and KNN application
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 3 Nov 2020 10:03:42 +0000 (19:03 +0900)]
[backbone] Added native support for tflite backbone
Added native support for tflite backbone
The unittests are verified for tflite backbone as tflite backbone
is preferred over nnstreamer backbone.
Interfacing directly with tflite API allows directly using tensor memory
for forwarding, avoid memcpy incurred with nnstreamer backbone.
TfLite takes input shape of format NHWC.
However, nntrainer takes the shape NCHW.
This will be resolved later with a transpose operator later.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
jijoong.moon [Thu, 29 Oct 2020 07:10:55 +0000 (16:10 +0900)]
[ Layer ] Add input_layers keyword
This PR includes enabling 'input_layers' keyword.
With this keyword, we could specify layer's input tensor.
. Added skip in parse_util to remove space and '[',']' in string and
split with ',' delimiter.
**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>
Parichay Kapoor [Thu, 29 Oct 2020 06:35:14 +0000 (15:35 +0900)]
[backbone] unittest for external backbone
Added support for external backbone
Also added a small add.tflite file for unittests
trainable is not supported with external backbones for now
Added tizen packaging and other unittest fixes
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Thu, 29 Oct 2020 06:32:42 +0000 (15:32 +0900)]
[backbone] Support for external backbone
Added support for external backbone with nnstreamer
This is enabled for both ubuntu and Tizen
This patch creates a nnstreamer layer to support running different kind of model files
with nnstreamer single c-api
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
hyeonseok lee [Wed, 4 Nov 2020 06:54:48 +0000 (15:54 +0900)]
Add github account on README.md file
Add my github account on README.md file
Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>
Jihoon Lee [Wed, 28 Oct 2020 11:42:37 +0000 (20:42 +0900)]
[IntegratedTest] Add batch normalization test
Add test that using batchnormalization with a minor structural change on
the tester
**Changes proposed in this PR for tester:**
- Implement reorder logic for bn layer in `recorder.py`
- Add color to some debug prints
- Let layer update inside actual `backwarding`
- save weights including non_trainable but save gradients only for the
trainables
**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>
Jihoon Lee [Wed, 28 Oct 2020 06:44:10 +0000 (15:44 +0900)]
[Test/Util] Update ini / debug info
**Changes proposed in this PR:**
- Add some helperfunction to iniTestWrapper
- Add color and adam debug info to recorder.py
**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>
Parichay Kapoor [Tue, 3 Nov 2020 04:33:05 +0000 (13:33 +0900)]
[concat] Move validity checks to init
Move the dimension checks for validity of the various inputs
for concat layer are moved to initialize().
They are kept back in forwarding() under DEBUG conditional.
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 3 Nov 2020 13:07:44 +0000 (22:07 +0900)]
[tensorfilter] Bug fix of tranpose
nntrainer takes data of NCHW
however, nnstreamer's tensor_converter produces data of format NHWC.
Existing implementation transposed the dimensions to match to NCHW format
However, the data was not transposed.
The unittest passed because the channel was just 1 which does not require
transpose on data side.
This patch removes the transpose of dimension from the nntrainer tensor_filter
and the transpose is done properly in the pipeline with nnstreamer's tensor_transform
v2:
Update transfer learning application which uses the pipeline.
v3:
Updated customshortcut application
Resolves #695
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Mete Ozay [Tue, 3 Nov 2020 07:53:36 +0000 (07:53 +0000)]
Troubleshooting installation process with observed errors and working solutions
Signed-off-by: Mete Ozay <meteozay@gmail.com>
Troubleshooting installation process with observed errors and working solutions
Jihoon Lee [Wed, 28 Oct 2020 05:23:47 +0000 (14:23 +0900)]
[Fix/test] Fix adam iteration start to 1
Fix bug that epoch_idx is not saved for continue_train
Fix iteration to start from 0 in model 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>
jijoong.moon [Tue, 27 Oct 2020 04:51:00 +0000 (13:51 +0900)]
[ Layer ] Add Concat Layer
This PR includes Concatenate layer.
**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>
Jihoon Lee [Fri, 23 Oct 2020 06:36:39 +0000 (15:36 +0900)]
[Test] Refactor recorder.py
This commit mainly patches KerasRecorder to be more flexible
**Changes proposed in this PR:**
- Pass file, label info at `KerasRecorder.run` phase instead of __init__
to leave room to reuse the model
- Allow initiation with SequentialModel for usuability
- Deal with cross_sigmoid, cross_softmax
- Move some functions out of class
**V2**
Since `KerasRecorder` class was highly coupled to a certain model and
made it hard to make some variation out of it, (e.g. using "mse" instead
of "cross" will need to make a whole new class and it is much
error-prone.
This patch move the class implementation to several functions.
This will be used with `functools.partial` so to easily generate loss,
optimizer variation.
**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/Refactor] Restructure data format
Restructure golden data to reduce redundant data and only check updated
weight thus making code more readable
**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>
Jihoon Lee [Wed, 28 Oct 2020 11:07:35 +0000 (20:07 +0900)]
[Fix/bn] Fix batchnormalize layer
Fix batchnormalizationLayer that epsilon is included toward saving
moving variacne that could lead to long term inaccuracy
**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>
Jihoon Lee [Wed, 28 Oct 2020 04:56:11 +0000 (13:56 +0900)]
[Fix/Optimizer] Fix decay_rate application
There was a bug that decay_rate were always applied when it is set to
default value. Becuase `decay_steps != -1` was evaluated to true.
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>
Parichay Kapoor [Fri, 23 Oct 2020 06:00:51 +0000 (15:00 +0900)]
[backbone] Add trainable feature to backbone
This patch adds supporting training feature to backbone
Corresponding unittests are also added
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
jijoong.moon [Tue, 27 Oct 2020 11:01:52 +0000 (20:01 +0900)]
[ Layer ] Multiple Input Dimensions
Current implementaion only take one input. In order to take multiple
input, input_dim / output_dim should be vector type.
This PR includes this fixes except about addition layer which requires
actual multiple input. This will be done consequtive PR.
**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>
Parichay Kapoor [Fri, 23 Oct 2020 03:06:04 +0000 (12:06 +0900)]
[backbone] Add unittests for backbone
Added unittests for backbone where model constructed
with and without backbone are matched to be equivalent.
Corresponding bug fixes are also added.
See also #660
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Thu, 22 Oct 2020 08:39:10 +0000 (17:39 +0900)]
[model] Remove redundant checks
Remove redundant check of adding layer to ensure that
name is unique. This is already ensure with ensureName()
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Thu, 22 Oct 2020 08:28:26 +0000 (17:28 +0900)]
[model] Add support of ini based backbone to the model
This patch adds support of ini based backbone to the model neural network
From the point of view of ini file, backbone is treated as a layer itself.
This allows a graph of layers to be represented as a layer itself in the
ini file.
With this design, backbone must be specified as a layer with property backbone
as shown below with a sample pseudo-ini:
```ini
[Block1]
backbone: base_block.ini
[PoolLayer]
type: pooling2d
[Block2]
backbone: base_block.ini
```
ModelLoader loads the layer configuration from the backbone independently
and then extends the existing graph in the main model with this newly created
graph from the backbone ini.
The names of all layers which are inserted from the backbone to a model are
prefixed with the name of the backbone for easier management and for the user
to identify/manage the layers from a backbone.
The patch allows nested backbones and multiple backbones in a model description.
Unittests for this backbone support will follow in the next patch.
See also #660
**Self evaluation:**
1. Build test: [x]Passed [ ]Failed [ ]Skipped
2. Run test: [x]Passed [ ]Failed [ ]Skipped
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Jihoon Lee [Thu, 22 Oct 2020 05:14:42 +0000 (14:14 +0900)]
[Layer] Fix 'stream << ' delegation
There was a bug that `std::cout << layer` is calling undefined function.
This patche fixes the issue and add some regression 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>
Dongju Chae [Thu, 29 Oct 2020 02:36:22 +0000 (11:36 +0900)]
[README.md] Add hall-of-fame section to README.md
This patch adds hall-of-fame section to README.md
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Jihoon Lee [Wed, 21 Oct 2020 10:20:05 +0000 (19:20 +0900)]
[Test] Add ParamTest scaffolding for model tests
Add Parameterized test scaffolding for integrated tests with minor
changes
**Additional Changes proposed in this PR:**
- Add `TensorDim::TensorDim(const std::string &shape)
- Move `Tensor::epsilon` to public
**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>