[Optimized] Making more macros to be executed before the execution. Testing time is reduced from 1min to 5sec.
[Optimized] define-impl-op macro enabled reducing the time due to (compile nil body). More Operations (e.g.: MaxMin, Matmul) could be reimplemented with this macro. This indicates that in the future cl-waffe2 can be used like define-by-run while Composites can be compiled statically.
Disassembled Backward InstructionSeq has become much more smaller!!! In-place mutation in backward also works well!!!
[Refactor] cl-waffe2/vm.nodes
[Deleted] define-static-nodedefine-composite-function. Integrated into define-op and defmodel-as macros respectively.
[Added] define-impl-opdefmodel-as (See documentations)
CPUTensor is now a subclass of LispTensor
[Documentation] Updated docstrings on the nodes/generic-tensor package.
Changes
[Optimize] cl-waffe2/vm cl-waffe2/vm.generic-tensor
!softmax is now working 2x times faster compared to the old one
proceed
is working with cl-waffe2 VM.MoveTensor(SAVE_FOR_BACKWARD)
intoAbstractNode
so that more rebundant nodes are prunedIR: Block
in the backward computation nodes. Ref: https://arxiv.org/pdf/1810.07951.pdf https://arxiv.org/pdf/1810.11530.pdf(compile nil body)
. More Operations (e.g.: MaxMin, Matmul) could be reimplemented with this macro. This indicates that in the future cl-waffe2 can be used like define-by-run while Composites can be compiled statically.[Refactor] cl-waffe2/vm.nodes
define-static-node
define-composite-function
. Integrated intodefine-op
anddefmodel-as
macros respectively.define-impl-op
defmodel-as
(See documentations)