static void Evaluate(Func<Tuple<INetwork, IFactory>> GetNetwork, bool verbose)
{
var StartTimingLayer = new TimingLayer() {StartCounters = new string[] { "Prediction-Time" } };
var StopTimingLayer = new TimingLayer() {StopCounters = new string[] { "Prediction-Time" } };
IInputLayer ReaderLayer = null;
var NetworkAndFactory = GetNetwork();
var Network = NetworkAndFactory.Item1;
var Factory = NetworkAndFactory.Item2;
{
var p = Network;
while (!(p.GetSource() is EncryptLayer)) p = p.GetSource();
StartTimingLayer.Source = p.GetSource();
var b = p as BaseLayer;
b.Source = StartTimingLayer;
// find the reader
while (p.GetSource() != null) p = p.GetSource();
ReaderLayer = p as IInputLayer;
// stop the timing counters after computing the entire network
StopTimingLayer.Source = Network;
Network = StopTimingLayer;
p = Network;
while (p != null)
{
p.Factory = Factory;
if (p is BaseLayer bas) bas.Verbose = verbose;
p = p.GetSource();
}
Network.PrepareNetwork();
}
int errs = 0;
for (int i = 0; i < 10000; i++)
{
using (var m = Network.GetNext())
{
var l = ReaderLayer.Labels[0];
int pred = 0;
Utils.ProcessInEnv(env =>
{
var dec = m.Decrypt(env);
for (int j = 0; j < 10; j++)
if (dec[j, 0] > dec[pred, 0]) pred = j;
if (pred != l) errs++;
}, Factory);
Console.WriteLine("errs {0}/{1} accuracy {2:0.000}% {3} prediction {4} label {5}", errs, i + 1, 100 - (100.0 * errs / (i + 1)), TimingLayer.GetStats(), pred, l);
}
}
Network.DisposeNetwork();
}
public static void Main(string[] args)
{
var options = new Options();
var parsed = Parser.Default.ParseArguments <Options>(args).WithParsed(x => options = x);
if (parsed.Tag == ParserResultType.NotParsed)
{
Environment.Exit(-2);
}
WeightsReader wr = new WeightsReader("CifarWeight.csv", "CifarBias.csv");
// Model has accuracy of 76.5
// Current parameters (scale) provide accuracy of 76.31% and uses 78.55 + 1 bits in message length
// It has a latency of 740 seconds on the reference machine (Azure B8ms server at rest)
Console.WriteLine("Generating encryption keys {0}", DateTime.Now);
IFactory factory = null;
if (options.Encrypt)
{
factory = new EncryptedSealBfvFactory(new ulong[] { 957181001729, 957181034497 }, 16384, DecompositionBitCount: 60, GaloisDecompositionBitCount: 60, SmallModulusCount: 8);
}
else
{
factory = new RawFactory(16 * 1024);
}
Console.WriteLine("Encryption keys ready {0}", DateTime.Now);
int numberOfRecords = 10000;
bool verbose = options.Verbose;
string fileName = "cifar-test.tsv";
var readerLayer = new LLConvReader
{
FileName = fileName,
SparseFormat = false,
InputShape = new int[] { 3, 32, 32 },
KernelShape = new int[] { 3, 8, 8 },
Upperpadding = new int[] { 0, 1, 1 },
Lowerpadding = new int[] { 0, 1, 1 },
Stride = new int[] { 1000, 2, 2 },
NormalizationFactor = 1.0 / 256.0,
Scale = 8,
Verbose = verbose
};
var EncryptLayer = new EncryptLayer()
{
Source = readerLayer, Factory = factory
};
var StartTimingLayer = new TimingLayer()
{
Source = EncryptLayer, StartCounters = new string[] { "Inference-Time" }
};
var ConvLayer1 = new LLPoolLayer()
{
Source = StartTimingLayer,
InputShape = new int[] { 3, 32, 32 },
KernelShape = new int[] { 3, 8, 8 },
Upperpadding = new int[] { 0, 1, 1 },
Lowerpadding = new int[] { 0, 1, 1 },
Stride = new int[] { 1000, 2, 2 },
MapCount = new int[] { 83, 1, 1 },
WeightsScale = 256.0,
Weights = (double[])wr.Weights[0],
Bias = (double[])wr.Biases[0],
Verbose = verbose
};
var VectorizeLayer2 = new LLVectorizeLayer()
{
Source = ConvLayer1,
Verbose = verbose
};
var ActivationLayer3 = new SquareActivation()
{
Source = VectorizeLayer2,
Verbose = verbose
};
var ConvEngine = new ConvolutionEngine()
{
InputShape = new int[] { 83, 14, 14 },
KernelShape = new int[] { 83, 10, 10 },
Upperpadding = new int[] { 0, 4, 4 },
Lowerpadding = new int[] { 0, 4, 4 },
Stride = new int[] { 83, 2, 2 },
MapCount = new int[] { 112, 1, 1 }
};
var DenseLayer4 = new LLDenseLayer
{
Source = ActivationLayer3,
WeightsScale = 512.0,
Weights = ConvEngine.GetDenseWeights((double[])wr.Weights[1]),
Bias = ConvEngine.GetDenseBias((double[])wr.Biases[1]),
InputFormat = EVectorFormat.dense,
ForceDenseFormat = true,
Verbose = verbose
};
var ActivationLayer5 = new SquareActivation()
{
Source = DenseLayer4,
Verbose = verbose
};
var DenseLayer6 = new LLDenseLayer()
{
Source = ActivationLayer5,
Weights = (double[])wr.Weights[2],
Bias = (double[])wr.Biases[2],
WeightsScale = 512.0,
InputFormat = EVectorFormat.dense,
Verbose = verbose
};
var StopTimingLayer = new TimingLayer()
{
Source = DenseLayer6, StopCounters = new string[] { "Inference-Time" }
};
var network = StopTimingLayer;
Console.WriteLine("Preparing");
network.PrepareNetwork();
int count = 0;
int errs = 0;
int batchSize = 1;
while (count < numberOfRecords)
{
using (var m = network.GetNext())
{
Utils.ProcessInEnv(env =>
{
var decrypted = m.Decrypt(env);
int pred = 0;
for (int j = 1; j < decrypted.RowCount; j++)
{
if (decrypted[j, 0] > decrypted[pred, 0])
{
pred = j;
}
}
if (pred != readerLayer.Labels[0])
{
errs++;
}
count++;
if (count % batchSize == 0)
{
Console.Write("errs {0}/{1} accuracy {2:0.000}% prediction {3} label {4} {5}ms", errs, count, 100 - (100.0 * errs / (count)), pred, readerLayer.Labels[0], TimingLayer.GetStats());
if (options.Encrypt)
{
Console.WriteLine();
}
else
{
Console.WriteLine(" 2^{0} largest-value", Math.Log(RawMatrix.Max) / Math.Log(2));
}
}
}, factory);
}
}
Console.WriteLine("errs {0}/{1} accuracy {2:0.000}%", errs, count, 100 - (100.0 * errs / (count)));
network.DisposeNetwork();
if (!options.Encrypt)
{
Console.WriteLine("Max computed value 2^{1}", Math.Log(RawMatrix.Max) / Math.Log(2));
}
}
static void Main(string[] args)
{
string fileName = "MNIST-28x28-test.txt";
int batchSize = 8192;
int numberOfRecords = 10000;
var Factory = new EncryptedSealBfvFactory(new ulong[] { 2148728833, 2148794369, 2149810177 }, 16384, DecompositionBitCount: 60, GaloisDecompositionBitCount: 60, SmallModulusCount: 7);
int weightscale = 32;
var ReaderLayer = new BatchReader
{
FileName = fileName,
SparseFormat = true,
MaxSlots = batchSize,
NormalizationFactor = 1.0 / 256.0,
Scale = 16.0
};
var EncryptedLayer = new EncryptLayer()
{
Source = ReaderLayer, Factory = Factory
};
var StartTimingLayer = new TimingLayer()
{
Source = EncryptedLayer, StartCounters = new string[] { "Batch-Time" }
};
var ConvLayer1 = new PoolLayer()
{
Source = StartTimingLayer,
InputShape = new int[] { 28, 28 },
KernelShape = new int[] { 5, 5 },
Upperpadding = new int[] { 1, 1 },
Stride = new int[] { 2, 2 },
MapCount = new int[] { 5, 1 },
WeightsScale = weightscale,
Weights = Weights.Weights_0
};
//var ActivationLayer2 = new SquareActivation() { Source = ConvLayer1 };
var ActivationLayer2 = new AppxReLUActivation()
{
Source = ConvLayer1
};
var DenseLayer3 = new PoolLayer()
{
Source = ActivationLayer2,
InputShape = new int[] { 5 * 13 * 13 },
KernelShape = new int[] { 5 * 13 * 13 },
Stride = new int[] { 1000 },
MapCount = new int[] { 100 },
Weights = Transpose(Weights.Weights_1, 5 * 13 * 13, 100),
Bias = Weights.Biases_2,
WeightsScale = weightscale * weightscale
};
//var ActivationLayer4 = new SquareActivation() { Source = DenseLayer3 };
var ActivationLayer4 = new AppxReLUActivation()
{
Source = DenseLayer3
};
var DenseLayer5 = new PoolLayer()
{
Source = ActivationLayer4,
InputShape = new int[] { 100 },
KernelShape = new int[] { 100 },
Stride = new int[] { 1000 },
MapCount = new int[] { 10 },
Weights = Weights.Weights_3,
Bias = Weights.Biases_3,
WeightsScale = weightscale
};
var StopTimingLayer = new TimingLayer()
{
Source = DenseLayer5, StopCounters = new string[] { "Batch-Time" }
};
var network = StopTimingLayer;
OperationsCount.Reset();
Console.WriteLine("Preparing");
network.PrepareNetwork();
OperationsCount.Print();
OperationsCount.Reset();
for (var p = (INetwork)network; p != null; p = p.Source)
{
if (p is BaseLayer b)
{
b.Verbose = true;
}
}
int count = 0;
int errs = 0;
while (count < numberOfRecords)
{
using (var m = network.GetNext())
Utils.ProcessInEnv(env =>
{
var decrypted = m.Decrypt(env);
for (int i = 0; i < decrypted.RowCount; i++)
{
int pred = 0;
for (int j = 1; j < decrypted.ColumnCount; j++)
{
if (decrypted[i, j] > decrypted[i, pred])
{
pred = j;
}
}
if (pred != ReaderLayer.Labels[i])
{
errs++;
}
count++;
if (count % 100 == 0)
{
Console.WriteLine("errs {0}/{1} accuracy {2:0.000}% prediction {3} label {4}", errs, count, 100 - (100.0 * errs / (count)), pred, ReaderLayer.Labels[i]);
}
}
Console.WriteLine("Batch size {0} {1}", batchSize, TimingLayer.GetStats());
}, Factory);
}
Console.WriteLine("errs {0}/{1} accuracy {2:0.000}%", errs, count, 100 - (100.0 * errs / (count)));
network.DisposeNetwork();
}
static void Main(string[] args)
{
var ini = new IniReader(@"cal.model.ini", 4096, 102);
ini.Normalize(@"cal.AffineNormalizer.txt");
var start = DateTime.Now;
var Factory = new EncryptedSealBfvFactory(new ulong[] { 4300801 }, 4096, DecompositionBitCount: 60, GaloisDecompositionBitCount: 60, SmallModulusCount: 2);
double weightscale = 256;
string FileName = "cal_deep_test.tsv";
if (!File.Exists(FileName))
{
Console.WriteLine("ERROR: Can't find data file {0}", FileName);
Console.WriteLine("Please use DataPreprocess to obtain the Caltech-101 dataset");
return;
}
var readerLayer = new LLSingleLineReader()
{
FileName = FileName,
SparseFormat = true,
NormalizationFactor = 1.0,
Scale = 256,
};
var encryptLayer = new EncryptLayer()
{
Source = readerLayer, Factory = Factory
};
var denseLayer = new LLDenseLayer()
{
Source = encryptLayer,
Weights = ini.Weights,
Bias = ini.Bias,
WeightsScale = weightscale,
InputFormat = EVectorFormat.dense
};
var network = denseLayer;
network.PrepareNetwork();
int errs = 0;
var N = 1020;
IMatrix m = null;
Utils.ProcessInEnv(env =>
{
for (int i = 0; i < N; i++)
{
Utils.Time("Prediction+Encryption", () => m = network.GetNext());
var dec = m.Decrypt(env);
m.Dispose();
var l = readerLayer.Labels[0];
int pred = 0;
for (int j = 0; j < 101; j++)
{
if (dec[j, 0] > dec[pred, 0])
{
pred = j;
}
}
if (pred != l)
{
errs++;
}
Console.WriteLine("errs {0}/{1} accuracy {2:0.000}% {3} prediction {4} label {5}", errs, i + 1, 100 - (100.0 * errs / (i + 1)), TimingLayer.GetStats(), pred, l);
}
}, Factory);
network.DisposeNetwork();
}