private static void Main()
{
const int imageSize = 28;
int[] layers = { 128, 64, 10 };
const int batchSize = 100;
const int maxEpoch = 10;
const float learningRate = 0.1f;
const float weightDecay = 1e-2f;
var trainIter = new MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/train-images-idx3-ubyte")
.SetParam("label", "./mnist_data/train-labels-idx1-ubyte")
.SetParam("batch_size", batchSize)
.SetParam("flat", 1)
.CreateDataIter();
var valIter = new MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/t10k-images-idx3-ubyte")
.SetParam("label", "./mnist_data/t10k-labels-idx1-ubyte")
.SetParam("batch_size", batchSize)
.SetParam("flat", 1)
.CreateDataIter();
var net = Mlp(layers);
Context ctx = Context.Cpu(); // Use CPU for training
var args = new SortedDictionary <string, NDArray>();
args["X"] = new NDArray(new Shape(batchSize, imageSize * imageSize), ctx);
args["label"] = new NDArray(new Shape(batchSize), ctx);
// Let MXNet infer shapes other parameters such as weights
net.InferArgsMap(ctx, args, args);
// Initialize all parameters with uniform distribution U(-0.01, 0.01)
var initializer = new Uniform(0.01f);
foreach (var arg in args)
{
// arg.first is parameter name, and arg.second is the value
initializer.Operator(arg.Key, arg.Value);
}
// Create sgd optimizer
var opt = OptimizerRegistry.Find("sgd");
opt.SetParam("rescale_grad", 1.0 / batchSize)
.SetParam("lr", learningRate)
.SetParam("wd", weightDecay);
// Create executor by binding parameters to the model
using (var exec = net.SimpleBind(ctx, args))
{
var argNames = net.ListArguments();
// Start training
var sw = new Stopwatch();
for (var iter = 0; iter < maxEpoch; ++iter)
{
var samples = 0;
trainIter.Reset();
sw.Restart();
while (trainIter.Next())
{
samples += batchSize;
var dataBatch = trainIter.GetDataBatch();
// Set data and label
dataBatch.Data.CopyTo(args["X"]);
dataBatch.Label.CopyTo(args["label"]);
// Compute gradients
exec.Forward(true);
exec.Backward();
// Update parameters
for (var i = 0; i < argNames.Count; ++i)
{
if (argNames[i] == "X" || argNames[i] == "label")
{
continue;
}
opt.Update(i, exec.ArgmentArrays[i], exec.GradientArrays[i]);
}
}
sw.Stop();
var accuracy = new Accuracy();
valIter.Reset();
while (valIter.Next())
{
var dataBatch = valIter.GetDataBatch();
dataBatch.Data.CopyTo(args["X"]);
dataBatch.Label.CopyTo(args["label"]);
// Forward pass is enough as no gradient is needed when evaluating
exec.Forward(false);
accuracy.Update(dataBatch.Label, exec.Outputs[0]);
}
var duration = sw.ElapsedMilliseconds / 1000.0;
Logging.LG($"Epoch: {iter} {samples / duration} samples/sec Accuracy: {accuracy.Get()}");
}
}
MXNet.MXNotifyShutdown();
}
private static void Main()
{
/*basic config*/
const int batchSize = 256;
const int maxEpo = 100;
const float learningRate = 1e-4f;
const float weightDecay = 1e-4f;
/*context and net symbol*/
var ctx = Context.Gpu();
var net = AlexnetSymbol(2);
/*args_map and aux_map is used for parameters' saving*/
var argsMap = new Dictionary <string, NDArray>();
var auxMap = new Dictionary <string, NDArray>();
/*we should tell mxnet the shape of data and label*/
argsMap["data"] = new NDArray(new Shape(batchSize, 3, 256, 256), ctx);
argsMap["label"] = new NDArray(new Shape(batchSize), ctx);
/*with data and label, executor can be generated varmatically*/
using (var exec = net.SimpleBind(ctx, argsMap))
{
var argNames = net.ListArguments();
var auxiliaryDictionary = exec.AuxiliaryDictionary();
var argmentDictionary = exec.ArgmentDictionary();
/*if fine tune from some pre-trained model, we should load the parameters*/
// NDArray.Load("./model/alex_params_3", nullptr, &args_map);
/*else, we should use initializer Xavier to init the params*/
var xavier = new Xavier(RandType.Gaussian, FactorType.In, 2.34f);
foreach (var arg in argmentDictionary)
{
/*be careful here, the arg's name must has some specific ends or starts for
* initializer to call*/
xavier.Operator(arg.Key, arg.Value);
}
/*print out to check the shape of the net*/
foreach (var s in net.ListArguments())
{
Logging.LG(s);
var sb = new StringBuilder();
var k = argmentDictionary[s].GetShape();
foreach (var i in k)
{
sb.Append($"{i} ");
}
Logging.LG(sb.ToString());
}
/*these binary files should be generated using im2rc tools, which can be found
* in mxnet/bin*/
var trainIter = new MXDataIter("ImageRecordIter")
.SetParam("path_imglist", "./data/train.lst")
.SetParam("path_imgrec", "./data/train.rec")
.SetParam("data_shape", new Shape(3, 256, 256))
.SetParam("batch_size", batchSize)
.SetParam("shuffle", 1)
.CreateDataIter();
var valIter = new MXDataIter("ImageRecordIter")
.SetParam("path_imglist", "./data/val.lst")
.SetParam("path_imgrec", "./data/val.rec")
.SetParam("data_shape", new Shape(3, 256, 256))
.SetParam("batch_size", batchSize)
.CreateDataIter();
var opt = OptimizerRegistry.Find("ccsgd");
opt.SetParam("momentum", 0.9)
.SetParam("rescale_grad", 1.0 / batchSize)
.SetParam("clip_gradient", 10)
.SetParam("lr", learningRate)
.SetParam("wd", weightDecay);
var accuracyTrain = new Accuracy();
var accuracyVal = new Accuracy();
var loglossVal = new LogLoss();
for (var iter = 0; iter < maxEpo; ++iter)
{
Logging.LG($"Train Epoch: {iter}");
/*reset the metric every epoch*/
accuracyTrain.Reset();
/*reset the data iter every epoch*/
trainIter.Reset();
while (trainIter.Next())
{
var batch = trainIter.GetDataBatch();
Logging.LG($"{trainIter.GetDataBatch().Index.Length}");
/*use copyto to feed new data and label to the executor*/
batch.Data.CopyTo(argmentDictionary["data"]);
batch.Label.CopyTo(argmentDictionary["label"]);
exec.Forward(true);
exec.Backward();
for (var i = 0; i < argNames.Count; ++i)
{
if (argNames[i] == "data" || argNames[i] == "label")
{
continue;
}
opt.Update(i, exec.ArgmentArrays[i], exec.GradientArrays[i]);
}
NDArray.WaitAll();
accuracyTrain.Update(batch.Label, exec.Outputs[0]);
}
Logging.LG($"ITER: {iter} Train Accuracy: {accuracyTrain.Get()}");
Logging.LG($"Val Epoch: {iter}");
accuracyVal.Reset();
valIter.Reset();
loglossVal.Reset();
while (valIter.Next())
{
var batch = valIter.GetDataBatch();
Logging.LG($"{valIter.GetDataBatch().Index.Length}");
batch.Data.CopyTo(argmentDictionary["data"]);
batch.Label.CopyTo(argmentDictionary["label"]);
exec.Forward(false);
NDArray.WaitAll();
accuracyVal.Update(batch.Label, exec.Outputs[0]);
loglossVal.Update(batch.Label, exec.Outputs[0]);
}
Logging.LG($"ITER: {iter} Val Accuracy: {accuracyVal.Get()}");
Logging.LG($"ITER: {iter} Val LogLoss: {loglossVal.Get()}");
/*save the parameters*/
var savePathParam = $"./model/alex_param_{iter}";
var saveArgs = argmentDictionary;
/*we do not want to save the data and label*/
if (saveArgs.ContainsKey("data"))
{
saveArgs.Remove("data");
}
if (saveArgs.ContainsKey("label"))
{
saveArgs.Remove("label");
}
/*the alexnet does not get any aux array, so we do not need to save
* aux_map*/
Logging.LG($"ITER: {iter} Saving to...{savePathParam}");
NDArray.Save(savePathParam, saveArgs);
}
/*don't foget to release the executor*/
}
MXNet.MXNotifyShutdown();
}
private static void Main(string[] args)
{
//var minScore = float.Parse(args[0], NumberStyles.Float, null);
var minScore = 0.9f;
const int imageSize = 28;
var layers = new[] { 128, 64, 10 };
const int batchSize = 100;
const int maxEpoch = 10;
const float learningRate = 0.1f;
const float weightDecay = 1e-2f;
var trainIter = new MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/train-images-idx3-ubyte")
.SetParam("label", "./mnist_data/train-labels-idx1-ubyte")
.SetParam("batch_size", batchSize)
.SetParam("flat", 1)
.CreateDataIter();
var valIter = new MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/t10k-images-idx3-ubyte")
.SetParam("label", "./mnist_data/t10k-labels-idx1-ubyte")
.SetParam("batch_size", batchSize)
.SetParam("flat", 1)
.CreateDataIter();
var net = Mlp(layers);
var ctx = Context.Cpu(); // Use GPU for training
var dictionary = new Dictionary <string, NDArray>();
dictionary["X"] = new NDArray(new Shape(batchSize, imageSize * imageSize), ctx);
dictionary["label"] = new NDArray(new Shape(batchSize), ctx);
// Let MXNet infer shapes of other parameters such as weights
net.InferArgsMap(ctx, dictionary, dictionary);
// Initialize all parameters with uniform distribution U(-0.01, 0.01)
var initializer = new Uniform(0.01f);
foreach (var arg in dictionary)
{
// arg.first is parameter name, and arg.second is the value
initializer.Operator(arg.Key, arg.Value);
}
// Create sgd optimizer
var opt = OptimizerRegistry.Find("sgd");
opt.SetParam("rescale_grad", 1.0 / batchSize)
.SetParam("lr", learningRate)
.SetParam("wd", weightDecay);
var lrSch = new UniquePtr <LRScheduler>(new FactorScheduler(5000, 0.1f));
opt.SetLearningRateScheduler(lrSch);
// Create executor by binding parameters to the model
using (var exec = net.SimpleBind(ctx, dictionary))
{
var argNames = net.ListArguments();
float score = 0;
// Start training
var sw = new Stopwatch();
for (var iter = 0; iter < maxEpoch; ++iter)
{
var samples = 0;
trainIter.Reset();
sw.Restart();
while (trainIter.Next())
{
samples += batchSize;
var dataBatch = trainIter.GetDataBatch();
// Data provided by DataIter are stored in memory, should be copied to GPU first.
dataBatch.Data.CopyTo(dictionary["X"]);
dataBatch.Label.CopyTo(dictionary["label"]);
// CopyTo is imperative, need to wait for it to complete.
NDArray.WaitAll();
// Compute gradients
exec.Forward(true);
exec.Backward();
// Update parameters
for (var i = 0; i < argNames.Count; ++i)
{
if (argNames[i] == "X" || argNames[i] == "label")
{
continue;
}
var weight = exec.ArgmentArrays[i];
var grad = exec.GradientArrays[i];
opt.Update(i, weight, grad);
}
}
sw.Stop();
var acc = new Accuracy();
valIter.Reset();
while (valIter.Next())
{
var dataBatch = valIter.GetDataBatch();
dataBatch.Data.CopyTo(dictionary["X"]);
dataBatch.Label.CopyTo(dictionary["label"]);
NDArray.WaitAll();
// Only forward pass is enough as no gradient is needed when evaluating
exec.Forward(false);
acc.Update(dataBatch.Label, exec.Outputs[0]);
}
var duration = sw.ElapsedMilliseconds / 1000.0;
var message = $"Epoch: {iter} {samples / duration} samples/sec Accuracy: {acc.Get()}";
Logging.LG(message);
score = acc.Get();
}
MXNet.MXNotifyShutdown();
var ret = score >= minScore ? 0 : 1;
Console.WriteLine($"{ret}");
}
}
private static void Main()
{
/*setup basic configs*/
const int W = 28;
const int H = 28;
const int batchSize = 128;
const int maxEpoch = 100;
const float learningRate = 1e-4f;
const float weightDecay = 1e-4f;
var contest = Context.Gpu();
var lenet = LenetSymbol();
var argsMap = new SortedDictionary <string, NDArray>();
argsMap["data"] = new NDArray(new Shape(batchSize, 1, W, H), contest);
argsMap["data_label"] = new NDArray(new Shape(batchSize), contest);
lenet.InferArgsMap(contest, argsMap, argsMap);
argsMap["fc1_w"] = new NDArray(new Shape(500, 4 * 4 * 50), contest);
NDArray.SampleGaussian(0, 1, argsMap["fc1_w"]);
argsMap["fc2_b"] = new NDArray(new Shape(10), contest);
argsMap["fc2_b"].Set(0);
var trainIter = new MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/train-images-idx3-ubyte")
.SetParam("label", "./mnist_data/train-labels-idx1-ubyte")
.SetParam("batch_size", batchSize)
.SetParam("shuffle", 1)
.SetParam("flat", 0)
.CreateDataIter();
var valIter = new MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/t10k-images-idx3-ubyte")
.SetParam("label", "./mnist_data/t10k-labels-idx1-ubyte")
.CreateDataIter();
var opt = OptimizerRegistry.Find("ccsgd");
opt.SetParam("momentum", 0.9)
.SetParam("rescale_grad", 1.0)
.SetParam("clip_gradient", 10)
.SetParam("lr", learningRate)
.SetParam("wd", weightDecay);
using (var exec = lenet.SimpleBind(contest, argsMap))
{
var argNames = lenet.ListArguments();
// Create metrics
var trainAccuracy = new Accuracy();
var valAccuracy = new Accuracy();
var sw = new Stopwatch();
for (var iter = 0; iter < maxEpoch; ++iter)
{
var samples = 0;
trainIter.Reset();
trainAccuracy.Reset();
sw.Restart();
while (trainIter.Next())
{
samples += batchSize;
var dataBatch = trainIter.GetDataBatch();
dataBatch.Data.CopyTo(argsMap["data"]);
dataBatch.Label.CopyTo(argsMap["data_label"]);
NDArray.WaitAll();
// Compute gradients
exec.Forward(true);
exec.Backward();
// Update parameters
for (var i = 0; i < argNames.Count; ++i)
{
if (argNames[i] == "data" || argNames[i] == "data_label")
{
continue;
}
opt.Update(i, exec.ArgmentArrays[i], exec.GradientArrays[i]);
}
// Update metric
trainAccuracy.Update(dataBatch.Label, exec.Outputs[0]);
}
// one epoch of training is finished
sw.Stop();
var duration = sw.ElapsedMilliseconds / 1000.0;
Logging.LG($"Epoch[{iter}] {samples / duration} samples/sec Train-Accuracy={trainAccuracy.Get()}");
valIter.Reset();
valAccuracy.Reset();
var accuracy = new Accuracy();
valIter.Reset();
while (valIter.Next())
{
var dataBatch = valIter.GetDataBatch();
dataBatch.Data.CopyTo(argsMap["data"]);
dataBatch.Label.CopyTo(argsMap["data_label"]);
NDArray.WaitAll();
// Only forward pass is enough as no gradient is needed when evaluating
exec.Forward(false);
NDArray.WaitAll();
accuracy.Update(dataBatch.Label, exec.Outputs[0]);
valAccuracy.Update(dataBatch.Label, exec.Outputs[0]);
}
Logging.LG($"Epoch[{iter}] Val-Accuracy={valAccuracy.Get()}");
}
}
MXNet.MXNotifyShutdown();
}
private static void Main()
{
const uint batchSize = 50;
const uint maxEpoch = 100;
const float learningRate = 1e-4f;
const float weightDecay = 1e-4f;
var googlenet = GoogleNetSymbol(101 + 1); // +1 is BACKGROUND_Google
var argsMap = new Dictionary <string, NDArray>();
var auxMap = new Dictionary <string, NDArray>();
// change device type if you want to use GPU
var context = Context.Cpu();
argsMap["data"] = new NDArray(new Shape(batchSize, 3, 256, 256), context);
argsMap["data_label"] = new NDArray(new Shape(batchSize), context);
googlenet.InferArgsMap(Context.Cpu(), argsMap, argsMap);
var trainIter = new MXDataIter("ImageRecordIter")
.SetParam("path_imglist", "train.lst")
.SetParam("path_imgrec", "train.rec")
.SetParam("data_shape", new Shape(3, 256, 256))
.SetParam("batch_size", batchSize)
.SetParam("shuffle", 1)
.CreateDataIter();
var valIter = new MXDataIter("ImageRecordIter")
.SetParam("path_imglist", "val.lst")
.SetParam("path_imgrec", "val.rec")
.SetParam("data_shape", new Shape(3, 256, 256))
.SetParam("batch_size", batchSize)
.CreateDataIter();
var opt = OptimizerRegistry.Find("ccsgd");
opt.SetParam("momentum", 0.9)
.SetParam("rescale_grad", 1.0 / batchSize)
.SetParam("clip_gradient", 10)
.SetParam("lr", learningRate)
.SetParam("wd", weightDecay);
using (var exec = googlenet.SimpleBind(Context.Cpu(), argsMap))
{
var argNames = googlenet.ListArguments();
for (var iter = 0; iter < maxEpoch; ++iter)
{
Logging.LG($"Epoch: {iter}");
trainIter.Reset();
while (trainIter.Next())
{
var dataBatch = trainIter.GetDataBatch();
dataBatch.Data.CopyTo(argsMap["data"]);
dataBatch.Label.CopyTo(argsMap["data_label"]);
NDArray.WaitAll();
exec.Forward(true);
exec.Backward();
for (var i = 0; i < argNames.Count; ++i)
{
if (argNames[i] == "data" || argNames[i] == "data_label")
{
continue;
}
var weight = exec.ArgmentArrays[i];
var grad = exec.GradientArrays[i];
opt.Update(i, weight, grad);
}
}
var acu = new Accuracy();
valIter.Reset();
while (valIter.Next())
{
var dataBatch = valIter.GetDataBatch();
dataBatch.Data.CopyTo(argsMap["data"]);
dataBatch.Label.CopyTo(argsMap["data_label"]);
NDArray.WaitAll();
exec.Forward(false);
NDArray.WaitAll();
acu.Update(dataBatch.Label, exec.Outputs[0]);
}
Logging.LG($"Accuracy: {acu.Get()}");
}
}
MXNet.MXNotifyShutdown();
}