Skip to content

Commit

Permalink
Speeding up TNNetDeMaxPool.
Browse files Browse the repository at this point in the history
  • Loading branch information
@joaopauloschuler
joaopauloschuler committed Dec 6, 2020
1 parent 076b9e2 commit b7a707e
Showing 1 changed file with 84 additions and 26 deletions.
110 changes: 84 additions & 26 deletions neural/neuralnetwork.pas
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1106,10 +1106,11 @@ TNNetAvgChannel = class(TNNetAvgPool)
{ TNNetDeMaxPool }
TNNetDeMaxPool = class(TNNetMaxPool)
private
FSpacing: integer;
procedure SetPrevLayer(pPrevLayer: TNNetLayer); override;
function CalcOutputSize(pInputSize: integer) : integer; override;
public
constructor Create(pPoolSize: integer); overload;
constructor Create(pPoolSize: integer; pSpacing: integer = 0); overload;
procedure Compute(); override;
procedure Backpropagate(); override;
procedure ComputePreviousLayerError(); override;
Expand Down Expand Up @@ -5812,9 +5813,11 @@ function TNNetDeMaxPool.CalcOutputSize(pInputSize: integer): integer;
Result := pInputSize * FPoolSize;
end;

constructor TNNetDeMaxPool.Create(pPoolSize: integer);
constructor TNNetDeMaxPool.Create(pPoolSize: integer; pSpacing: integer = 0);
begin
inherited Create(pPoolSize);
FSpacing := pSpacing;
FStruct[7] := FSpacing;
end;

procedure TNNetDeMaxPool.Compute();
Expand All @@ -5824,25 +5827,50 @@ procedure TNNetDeMaxPool.Compute();
OutX, OutY: integer;
CurrValue: TNeuralFloat;
StartTime: double;
PrevLayerRawPos, OutputRawPos: integer;
begin
StartTime := Now();
Output.Fill(-1000000);
Output.Fill(0);
MaxX := FPrevLayer.Output.SizeX - 1;
MaxY := FPrevLayer.Output.SizeY - 1;
MaxD := FPrevLayer.Output.Depth - 1;

for CntD := 0 to MaxD do
if FSpacing = 1 then
begin
for CntX := 0 to MaxX do
begin
OutX := CntX*FPoolSize;
for CntY := 0 to MaxY do
begin
CurrValue := FPrevLayer.Output[CntX,CntY,CntD];
for OutX := CntX*FPoolSize to CntX*FPoolSize + FPoolSize - 1 do
OutY := CntY*FPoolSize;
PrevLayerRawPos := FPrevLayer.FOutput.GetRawPos(CntX,CntY,0);
OutputRawPos := FOutput.GetRawPos(OutX,OutY,0);
for CntD := 0 to MaxD do
begin
for OutY := CntY*FPoolSize to CntY*FPoolSize + FPoolSize - 1 do
//CurrValue := FPrevLayer.FOutput[CntX,CntY,CntD];
//FOutput[OutX,OutY,CntD] := CurrValue;
FOutput.FData[OutputRawPos] := FPrevLayer.FOutput.FData[PrevLayerRawPos];
Inc(PrevLayerRawPos);
Inc(OutputRawPos);
end;
end;
end;
end
else
begin
for CntD := 0 to MaxD do
begin
for CntX := 0 to MaxX do
begin
for CntY := 0 to MaxY do
begin
CurrValue := FPrevLayer.Output[CntX,CntY,CntD];
for OutX := CntX*FPoolSize to CntX*FPoolSize + FPoolSize - 1 do
begin
Output[OutX,OutY,CntD] := CurrValue;
for OutY := CntY*FPoolSize to CntY*FPoolSize + FPoolSize - 1 do
begin
Output[OutX,OutY,CntD] := CurrValue;
end;
end;
end;
end;
Expand Down Expand Up @@ -5873,32 +5901,62 @@ procedure TNNetDeMaxPool.ComputePreviousLayerError();
RawPos, PrevRawPos: integer;
PrevPosX, PrevPosY: integer;
floatPoolSize: TNeuralFloat;
OutX, OutY: integer;
begin
MaxX := Output.SizeX - 1;
MaxY := Output.SizeY - 1;
MaxD := Output.Depth - 1;

floatPoolSize := FPoolSize;
FOutputError.Divi(floatPoolSize);

for CntY := 0 to MaxY do
if FSpacing = 1 then
begin
MaxX := FPrevLayer.FOutput.SizeX - 1;
MaxY := FPrevLayer.FOutput.SizeY - 1;
for CntX := 0 to MaxX do
begin
RawPos := FOutput.GetRawPos(CntX, CntY, 0);
for CntD := 0 to MaxD do
OutX := CntX*FPoolSize;
for CntY := 0 to MaxY do
begin
OutY := CntY*FPoolSize;
PrevRawPos := FPrevLayer.FOutputError.GetRawPos(CntX,CntY,0);
RawPos := FOutputError.GetRawPos(OutX,OutY,0);
for CntD := 0 to MaxD do
begin
{$IFDEF FPC}
FPrevLayer.FOutputError.FData[PrevRawPos] += FOutputError.FData[RawPos];
{$ELSE}
FPrevLayer.FOutputError.FData[PrevRawPos] :=
FPrevLayer.FOutputError.FData[PrevRawPos] + FOutputError.FData[RawPos];
{$ENDIF}
Inc(PrevRawPos);
Inc(RawPos);
end;
end;
end;
end
else
begin
MaxX := FOutput.SizeX - 1;
MaxY := FOutput.SizeY - 1;
FOutputError.Divi(floatPoolSize);
for CntY := 0 to MaxY do
begin
PrevPosY := CntY div FPoolSize;
for CntX := 0 to MaxX do
begin
PrevPosX := CntX div FPoolSize;
PrevPosY := CntY div FPoolSize;
PrevRawPos := FPrevLayer.FOutputError.GetRawPos(PrevPosX, PrevPosY, CntD);

{$IFDEF FPC}
FPrevLayer.FOutputError.FData[PrevRawPos] += FOutPutError.FData[RawPos];
{$ELSE}
FPrevLayer.FOutputError.FData[PrevRawPos] :=
FPrevLayer.FOutputError.FData[PrevRawPos] + FOutPutError.FData[RawPos];
{$ENDIF}
Inc(RawPos);
RawPos := FOutput.GetRawPos(CntX, CntY, 0);
PrevRawPos := FPrevLayer.FOutputError.GetRawPos(PrevPosX, PrevPosY, 0);
for CntD := 0 to MaxD do
begin
{$IFDEF FPC}
FPrevLayer.FOutputError.FData[PrevRawPos] += FOutPutError.FData[RawPos];
{$ELSE}
FPrevLayer.FOutputError.FData[PrevRawPos] :=
FPrevLayer.FOutputError.FData[PrevRawPos] + FOutPutError.FData[RawPos];
{$ENDIF}
Inc(RawPos);
Inc(PrevRawPos);
end;
end;
end;
end;
Expand Down Expand Up @@ -8253,7 +8311,7 @@ function TNNet.CreateLayer(strData: string): TNNetLayer;
'TNNetDeLocalConnectReLU' : Result := TNNetDeLocalConnectReLU.Create(St[0], St[1], St[4]);
'TNNetDeconvolution' : Result := TNNetDeconvolution.Create(St[0], St[1], St[4]);
'TNNetDeconvolutionReLU' : Result := TNNetDeconvolutionReLU.Create(St[0], St[1], St[4]);
'TNNetDeMaxPool' : Result := TNNetDeMaxPool.Create(St[0]);
'TNNetDeMaxPool' : Result := TNNetDeMaxPool.Create(St[0], St[7]);
'TNNetDeAvgPool' : Result := TNNetDeAvgPool.Create(St[0]);
'TNNetUpsample' : Result := TNNetUpsample.Create();
'TNNetLayerMaxNormalization': Result := TNNetLayerMaxNormalization.Create();
Expand Down Expand Up @@ -8328,7 +8386,7 @@ function TNNet.CreateLayer(strData: string): TNNetLayer;
if S[0] = 'TNNetDeLocalConnectReLU' then Result := TNNetDeLocalConnectReLU.Create(St[0], St[1], St[4]) else
if S[0] = 'TNNetDeconvolution' then Result := TNNetDeconvolution.Create(St[0], St[1], St[4]) else
if S[0] = 'TNNetDeconvolutionReLU' then Result := TNNetDeconvolutionReLU.Create(St[0], St[1], St[4]) else
if S[0] = 'TNNetDeMaxPool' then Result := TNNetDeMaxPool.Create(St[0]) else
if S[0] = 'TNNetDeMaxPool' then Result := TNNetDeMaxPool.Create(St[0], St[7]) else
if S[0] = 'TNNetDeAvgPool' then Result := TNNetDeAvgPool.Create(St[0]) else
if S[0] = 'TNNetUpsample' then Result := TNNetUpsample.Create() else
if S[0] = 'TNNetLayerMaxNormalization' then Result := TNNetLayerMaxNormalization.Create() else
Expand Down

0 comments on commit b7a707e

Please sign in to comment.