Logo ROOT  

Reference Guide
 
Loading...
Searching...
No Matches
ActivationFunctions.cu
Go to the documentation of this file.
1// @(#)root/tmva/tmva/dnn:$Id$
2// Author: Simon Pfreundschuh 13/07/16
3
4/*************************************************************************
5 * Copyright (C) 2016, Simon Pfreundschuh *
6 * All rights reserved. *
7 * *
8 * For the licensing terms see $ROOTSYS/LICENSE. *
9 * For the list of contributors see $ROOTSYS/README/CREDITS. *
10 *************************************************************************/
11
12 //////////////////////////////////////////////////////////////////
13 // Implementation of the activation functions for the TCuda //
14 // implementation of the low-level interface. //
15 //////////////////////////////////////////////////////////////////
16
17#include "TMVA/DNN/Architectures/Cuda.h"
18#include "TMVA/DNN/Architectures/Cuda/Device.h"
19#include "TMVA/DNN/Functions.h"
20#include "Kernels.cuh"
21
22namespace TMVA
23{
24namespace DNN
25{
26//______________________________________________________________________________
27template<typename AFloat>
28void TCuda<AFloat>::ActivationFunctionForward(Tensor_t & X, EActivationFunction activFunct,
29 const ActivationDescriptor_t /* activationDescr */,
30 const double /* coef */, const AFloat /*alpha */, const AFloat /*beta*/)
31{
32 // scaling and translation is not yet implemented
33 TMVA::DNN::evaluate<TCuda<AFloat>>( X, activFunct);
34}
35//______________________________________________________________________________
36template<typename AFloat>
37void TCuda<AFloat>::ActivationFunctionBackward(Tensor_t & dX, const Tensor_t & /* Y */,
38 const Tensor_t & dY, const Tensor_t & X,
39 EActivationFunction activFunct,
40 const ActivationDescriptor_t /* activationDescr */,
41 const AFloat /* alpha */, const AFloat /* beta */)
42{
43 // scaling and translation not yet implemented
44 // output tensor (Y) could also be used to speed up derivative calculation
45 // compute dx = f'(x)
46 TMVA::DNN::evaluateDerivative<TCuda<AFloat>>(dX, activFunct, X);
47 // Compute element-wise product. dx = f'(x) * dY
48 Hadamard(dX, dY);
49}
50
51//______________________________________________________________________________
52template<typename AFloat>
53void TCuda<AFloat>::IdentityDerivative(TCudaTensor<AFloat> & B,
54 const TCudaTensor<AFloat> & A)
55{
56 dim3 blockDims = TDevice::BlockDims2D();
57 dim3 gridDims = TDevice::GridDims2D(B);
58 cudaStream_t s = A.GetComputeStream();
59 ::TMVA::DNN::Cuda::IdentityDerivative<<<gridDims, blockDims, 0, s>>>(
60 B.GetDataPointer(),
61 (int) B.GetNrows(),
62 (int) B.GetNcols());
63 B.SetComputeStream(s);
64}
65
66//______________________________________________________________________________
67template<typename AFloat>
68void TCuda<AFloat>::Relu(TCudaTensor<AFloat> & A)
69{
70 dim3 blockDims = TDevice::BlockDims2D();
71 dim3 gridDims = TDevice::GridDims2D(A);
72 cudaStream_t s = A.GetComputeStream();
73 ::TMVA::DNN::Cuda::Relu<<<gridDims, blockDims, 0, s>>>(
74 A.GetDataPointer(),
75 (int) A.GetNrows(),
76 (int) A.GetNcols());
77}
78
79//______________________________________________________________________________
80template<typename AFloat>
81void TCuda<AFloat>::ReluDerivative(TCudaTensor<AFloat> & B,
82 const TCudaTensor<AFloat> & A)
83{
84 assert(B.GetNrows() == A.GetNrows() && B.GetNcols() == A.GetNcols());
85 dim3 blockDims = TDevice::BlockDims2D();
86 dim3 gridDims = TDevice::GridDims2D(B);
87 cudaStream_t s = A.GetComputeStream();
88 ::TMVA::DNN::Cuda::ReluDerivative<<<gridDims, blockDims, 0, s>>>(
89 B.GetDataPointer(),
90 A.GetDataPointer(),
91 (int) A.GetNrows(),
92 (int) A.GetNcols());
93 B.SetComputeStream(s);
94}
95
96//______________________________________________________________________________
97template<typename AFloat>
98void TCuda<AFloat>::Sigmoid(TCudaTensor<AFloat> & A)
99{
100 dim3 blockDims = TDevice::BlockDims2D();
101 dim3 gridDims = TDevice::GridDims2D(A);
102 cudaStream_t s = A.GetComputeStream();
103 ::TMVA::DNN::Cuda::Sigmoid<<<gridDims, blockDims, 0, s>>>(
104 A.GetDataPointer(),
105 (int) A.GetNrows(),
106 (int) A.GetNcols());
107}
108
109//______________________________________________________________________________
110template<typename AFloat>
111void TCuda<AFloat>::SigmoidDerivative(TCudaTensor<AFloat> & B,
112 const TCudaTensor<AFloat> & A)
113{
114 assert(B.GetNrows() == A.GetNrows() && B.GetNcols() == A.GetNcols());
115 dim3 blockDims = TDevice::BlockDims2D();
116 dim3 gridDims = TDevice::GridDims2D(B);
117 cudaStream_t s = A.GetComputeStream();
118 ::TMVA::DNN::Cuda::SigmoidDerivative<<<gridDims, blockDims, 0, s>>>(
119 B.GetDataPointer(),
120 A.GetDataPointer(),
121 (int) A.GetNrows(),
122 (int) A.GetNcols());
123 B.SetComputeStream(s);
124}
125
126//______________________________________________________________________________
127template<typename AFloat>
128void TCuda<AFloat>::Tanh(TCudaTensor<AFloat> & A)
129{
130 dim3 blockDims = TDevice::BlockDims2D();
131 dim3 gridDims = TDevice::GridDims2D(A);
132 cudaStream_t s = A.GetComputeStream();
133 ::TMVA::DNN::Cuda::Tanh<<<gridDims, blockDims, 0, s>>>(
134 A.GetDataPointer(),
135 (int) A.GetNrows(),
136 (int) A.GetNcols());
137}
138
139//______________________________________________________________________________
140template<typename AFloat>
141void TCuda<AFloat>::TanhDerivative(TCudaTensor<AFloat> & B,
142 const TCudaTensor<AFloat> & A)
143{
144 assert(B.GetNrows() == A.GetNrows() && B.GetNcols() == A.GetNcols());
145 dim3 blockDims = TDevice::BlockDims2D();
146 dim3 gridDims = TDevice::GridDims2D(B);
147 cudaStream_t s = A.GetComputeStream();
148 ::TMVA::DNN::Cuda::TanhDerivative<<<gridDims, blockDims, 0, s>>>(
149 B.GetDataPointer(),
150 A.GetDataPointer(),
151 (int) A.GetNrows(),
152 (int) A.GetNcols());
153 B.SetComputeStream(s);
154}
155
156//______________________________________________________________________________
157template<typename AFloat>
158void TCuda<AFloat>::SymmetricRelu(TCudaTensor<AFloat> & A)
159{
160 dim3 blockDims = TDevice::BlockDims2D();
161 dim3 gridDims = TDevice::GridDims2D(A);
162 cudaStream_t s = A.GetComputeStream();
163 ::TMVA::DNN::Cuda::SymmetricRelu<<<gridDims, blockDims, 0, s>>>(
164 A.GetDataPointer(),
165 (int) A.GetNrows(),
166 (int) A.GetNcols());
167}
168
169//______________________________________________________________________________
170template<typename AFloat>
171void TCuda<AFloat>::SymmetricReluDerivative(TCudaTensor<AFloat> & B,
172 const TCudaTensor<AFloat> & A)
173{
174 assert(B.GetNrows() == A.GetNrows() && B.GetNcols() == A.GetNcols());
175 dim3 blockDims = TDevice::BlockDims2D();
176 dim3 gridDims = TDevice::GridDims2D(B);
177 cudaStream_t s = A.GetComputeStream();
178 ::TMVA::DNN::Cuda::SymmetricReluDerivative<<<gridDims, blockDims, 0, s>>>(
179 B.GetDataPointer(),
180 A.GetDataPointer(),
181 (int) A.GetNrows(),
182 (int) A.GetNcols());
183 B.SetComputeStream(s);
184}
185
186//______________________________________________________________________________
187template<typename AFloat>
188void TCuda<AFloat>::SoftSign(TCudaTensor<AFloat> & A)
189{
190 dim3 blockDims = TDevice::BlockDims2D();
191 dim3 gridDims = TDevice::GridDims2D(A);
192 cudaStream_t s = A.GetComputeStream();
193 ::TMVA::DNN::Cuda::SoftSign<<<gridDims, blockDims, 0, s>>>(
194 A.GetDataPointer(),
195 (int) A.GetNrows(),
196 (int) A.GetNcols());
197}
198
199//______________________________________________________________________________
200template<typename AFloat>
201void TCuda<AFloat>::SoftSignDerivative(TCudaTensor<AFloat> & B,
202 const TCudaTensor<AFloat> & A)
203{
204 assert(B.GetNrows() == A.GetNrows() && B.GetNcols() == A.GetNcols());
205 dim3 blockDims = TDevice::BlockDims2D();
206 dim3 gridDims = TDevice::GridDims2D(B);
207 cudaStream_t s = A.GetComputeStream();
208 ::TMVA::DNN::Cuda::SoftSignDerivative<<<gridDims, blockDims, 0, s>>>(
209 B.GetDataPointer(),
210 A.GetDataPointer(),
211 (int) A.GetNrows(),
212 (int) A.GetNcols());
213 B.SetComputeStream(s);
214}
215
216//______________________________________________________________________________
217template<typename AFloat>
218void TCuda<AFloat>::Gauss(TCudaTensor<AFloat> & A)
219{
220 dim3 blockDims = TDevice::BlockDims2D();
221 dim3 gridDims = TDevice::GridDims2D(A);
222 cudaStream_t s = A.GetComputeStream();
223 ::TMVA::DNN::Cuda::Gauss<<<gridDims, blockDims, 0, s>>>(
224 A.GetDataPointer(),
225 (int) A.GetNrows(),
226 (int) A.GetNcols());
227}
228
229//______________________________________________________________________________
230template<typename AFloat>
231void TCuda<AFloat>::GaussDerivative(TCudaTensor<AFloat> & B,
232 const TCudaTensor<AFloat> & A)
233{
234 assert(B.GetNrows() == A.GetNrows() && B.GetNcols() == A.GetNcols());
235 dim3 blockDims = TDevice::BlockDims2D();
236 dim3 gridDims = TDevice::GridDims2D(B);
237 cudaStream_t s = A.GetComputeStream();
238 ::TMVA::DNN::Cuda::GaussDerivative<<<gridDims, blockDims, 0, s>>>(
239 B.GetDataPointer(),
240 A.GetDataPointer(),
241 (int) A.GetNrows(),
242 (int) A.GetNcols());
243 B.SetComputeStream(s);
244}
245
246} // namespace DNN
247} // namespace TMVA
X
#define X(type, name)
TRangeDynCast
ROOT::Detail::TRangeCast< T, true > TRangeDynCast
TRangeDynCast is an adapter class that allows the typed iteration through a TCollection.
Definition TCollection.h:358
TMVA::DNN::TCudaTensor
TCudaTensor Class.
Definition CudaTensor.h:84
TMVA::DNN::TCuda::SoftSignDerivative
static void SoftSignDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:201
TMVA::DNN::TCuda::SymmetricReluDerivative
static void SymmetricReluDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:171
TMVA::DNN::TCuda::IdentityDerivative
static void IdentityDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:53
TMVA::DNN::TCuda::ActivationFunctionForward
static void ActivationFunctionForward(Tensor_t &X, EActivationFunction activFunct, const ActivationDescriptor_t activationDescr, const double coef=0.0, const AFloat alpha=1, const AFloat beta=0)
Definition ActivationFunctions.cu:28
TMVA::DNN::TCuda::SoftSign
static void SoftSign(Tensor_t &B)
Definition ActivationFunctions.cu:188
TMVA::DNN::TCuda::Gauss
static void Gauss(Tensor_t &B)
Definition ActivationFunctions.cu:218
TMVA::DNN::TCuda::Sigmoid
static void Sigmoid(Tensor_t &B)
TMVA::DNN::TCuda::Tanh
static void Tanh(Tensor_t &B)
Definition ActivationFunctions.cu:128
TMVA::DNN::TCuda::ActivationFunctionBackward
static void ActivationFunctionBackward(Tensor_t &dX, const Tensor_t &Y, const Tensor_t &dY, const Tensor_t &X, EActivationFunction activFunct, const ActivationDescriptor_t activationDescr, const AFloat alpha=1, const AFloat beta=0)
Computes the gradient of the activation function.
Definition ActivationFunctions.cu:37
TMVA::DNN::TCuda::ReluDerivative
static void ReluDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:81
TMVA::DNN::TCuda::GaussDerivative
static void GaussDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:231
TMVA::DNN::TCuda::Relu
static void Relu(Tensor_t &B)
Definition ActivationFunctions.cu:68
TMVA::DNN::TCuda::SymmetricRelu
static void SymmetricRelu(Tensor_t &B)
Definition ActivationFunctions.cu:158
TMVA::DNN::TCuda::SigmoidDerivative
static void SigmoidDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:111
TMVA::DNN::TCuda::TanhDerivative
static void TanhDerivative(Tensor_t &B, const Tensor_t &A)
Definition ActivationFunctions.cu:141
TMVA::DNN::TDevice::BlockDims2D
static dim3 BlockDims2D()
Definition Device.h:55
TMVA::DNN::TDevice::GridDims2D
static dim3 GridDims2D(int nrows, int ncols)
Definition Device.h:74
TMVA::DNN::EActivationFunction
EActivationFunction
Enum that represents layer activation functions.
Definition Functions.h:32
TMVA
create variable transformations
Definition GeneticMinimizer.h:22