Fix HDF5 1.10.0

minor fix

• fix HDF5 1.10.0 error (H5Fcreate, H5Dcreate2, etc)

support

• upgrade Ubuntu 14.04 (trusty) to 16.04 (xenial) on Travis
• add Circle CI for Ubuntu 16.04 and 18.04

Move global configuration to grain.config and implement grain.chain.dropout

• dropout https://shigekikarita.github.io/grain/grain.chain.dropout.html
• global https://shigekikarita.github.io/grain/grain.config.html

v0.0.10

upgrade mir-algorithm to v2.0.1 #10

add squeeze/unsqueeze

v0.0.9

add squeeze/unsqueeze

Support autograd view (a.k.a. reshape) and MNIST using CNN example

• new function/chain for topology operation https://github.com/ShigekiKarita/grain/blob/v0.0.8/source/grain/functions/topology.d
• new example using convolutional neural networks https://github.com/ShigekiKarita/grain/blob/v0.0.8/example/mnist_cnn.d

DMD support (only CPU backend)

v0.0.7

update serializer

HDF5 Serialization and N-Dimentional Convolution/Cross-Correlation implemented

• HDF5 Serialization https://github.com/ShigekiKarita/grain/tree/v0.0.6/source/grain/serializer.d
• Convolution/Cross-Correlation
  • CPU implementation supports arbitrary number of convolution dim but slow
  • CUDNN implementation supports many optimized algorithms (winograd/fft/precumted gemm etc) but limited dim and shapes

Stateful Adaptive Optimizers (Adam, AdaDelta, AdaGrad)

v0.0.5

refactor optim

RNN support

• see example/char_rnn.d for tiny RNN example

add autograd opBinary("+", "-", "*", "/") and fast math functions in LLVM/CUDA

• Variable.opBinary https://shigekikarita.github.io/grain/grain.autograd.Variable.html#examples
• Intrinsic math functions https://shigekikarita.github.io/grain/grain.functions.unary.unaryFunc.html
  • LLVM http://mir.dlang.io/mir_math_common.html
  • CUDA https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#intrinsic-functions