private static void UpdateParams(
IList <List <NdArray> > paramArrays,
IList <List <NdArray> > gradArrays,
Action <int, NdArray, NdArray> updater,
int numDevice,
KvStore kvstore = null)
{
for (int index = 0; index < paramArrays.Count; index++)
{
var argList = paramArrays[index];
var gradList = gradArrays[index];
if (gradList[0] == null)
{
continue;
}
if (kvstore != null)
{
//push gradient, priority is negative index
kvstore.Push(index, gradList, priority: -index);
//pull back the weights
kvstore.Pull(index, argList, priority: -index);
}
for (int k = 0; k < argList.Count; k++)
{
var w = argList[k];
var g = gradList[k];
updater(index * numDevice + k, g, w);
}
}
}
private static void InitializeKvstore(KvStore kvstore,
IList <List <NdArray> > paramArrays,
Dictionary <string, NdArray> argParams,
IList <string> paramNames,
bool updateOnKvstore)
{
for (int idx = 0; idx < paramArrays.Count; idx++)
{
var paramOnDevs = paramArrays[idx];
kvstore.Init(idx, argParams[paramNames[idx]]);
if (updateOnKvstore)
{
kvstore.Pull(idx, paramOnDevs, priority: -idx);
}
}
}
private static void UpdateParamsOnKvstore(
IList <List <NdArray> > paramArrays,
IList <List <NdArray> > gradArrays,
KvStore kvstore)
{
for (int index = 0; index < paramArrays.Count; index++)
{
var argList = paramArrays[index];
var gradList = gradArrays[index];
if (gradList[0] == null)
{
continue;
}
//push gradient, priority is negative index
kvstore.Push(index, gradList, priority: -index);
//pull back the weights
kvstore.Pull(index, argList, priority: -index);
}
}
private static Tuple <KvStore, bool> CreateKvstore(
string kvstore, int count, Dictionary <string, NdArray> argParams)
{
KvStore kv;
if (kvstore == null)
{
kv = null;
}
else
{
if (count == 1 && !kvstore.Contains("dist"))
{
kv = null;
}
else
{
if (kvstore == "local")
{
//automatically select a proper local
var maxSize = argParams.Select(s => Util.Prod(s.Value.GetShape())).Max();
if (maxSize < 1024 * 1024 * 16)
{
kvstore = "local_update_cpu";
}
else
{
kvstore = "local_allreduce_cpu";
}
}
kv = new KvStore(kvstore);
}
}
bool updateOnKvstore = !(kv == null || kv.Type.Contains("local_allreduce"));
return(Tuple.Create(kv, updateOnKvstore));
}
private static void TrainMultiDevice(Symbol symbol,
IList <Context> ctx,
IList <string> argNames,
IList <string> paramNames,
IList <string> auxNames,
Dictionary <string, NdArray> argParams,
Dictionary <string, NdArray> auxParams,
int beginEpoch,
int endEpoch,
int?epochSize,
Optimizer optimizer,
IDataIter trainData,
IDataIter evalData,
EvalMetric evalMetric,
IList <EpochEndDelegate> epochEndCallback,
IList <BatchEndDelegate> batchEndCallback,
KvStore kvstore, bool updateOnKvstore,
ILog logger,
IList <int> workLoadList,
Monitor monitor,
IList <BatchEndDelegate> evalBatchEndCallback,
SymbolGenerate symGen)
{
if (logger == null)
{
logger = LogManager.GetLogger("");
}
var executorManager = new DataParallelExecutorManager(symbol: symbol,
symGen: symGen,
ctx: ctx,
trainData: trainData,
paramNames: paramNames,
argNames: argNames,
auxNames: auxNames,
workLoadList: workLoadList,
logger: logger);
if (monitor != null)
{
executorManager.InstallMonitor(monitor);
}
executorManager.SetParams(argParams, auxParams);
Action <int, NdArray, NdArray> updater = null;
if (!updateOnKvstore)
{
updater = Optimizer.GetUpdater(optimizer);
}
if (kvstore != null)
{
InitializeKvstore(kvstore: kvstore,
paramArrays: executorManager.ParamArrays,
argParams: argParams,
paramNames: executorManager.ParamNames,
updateOnKvstore: updateOnKvstore);
}
if (updateOnKvstore)
{
kvstore?.SetOptimizer(optimizer);
}
//Now start training
for (int epoch = 0; epoch < endEpoch - beginEpoch; epoch++)
{
// Training phase
Stopwatch toc = new Stopwatch();
toc.Start();
evalMetric.Reset();
var nbatch = 0;
// Iterate over training data.
while (true)
{
var doReset = true;
foreach (var dataBatch in trainData)
{
executorManager.LoadDataBatch(dataBatch);
monitor?.Tic();
executorManager.Forward(isTrain: true);
executorManager.Backward();
if (updateOnKvstore)
{
UpdateParamsOnKvstore(
executorManager.ParamArrays,
executorManager.GradArrays,
kvstore);
}
else
{
UpdateParams(executorManager.ParamArrays,
executorManager.GradArrays,
updater: updater,
numDevice: ctx.Count,
kvstore: kvstore);
}
monitor?.TocPrint();
// evaluate at end, so we can lazy copy
executorManager.UpdateMetric(evalMetric, dataBatch.Label);
nbatch += 1;
//batch callback (for print purpose)
if (batchEndCallback != null)
{
var batchEndParams = new BatchEndParam(epoch: epoch,
nbatch: nbatch,
evalMetric: evalMetric,
locals: Thread.CurrentThread.CurrentCulture);
foreach (var call in batchEndCallback)
{
call(batchEndParams);
}
}
if (epochSize != null && nbatch >= epochSize)
{
doReset = false;
break;
}
}
if (doReset)
{
logger.Info($"Epoch[{epoch}] Resetting Data Iterator");
trainData.Reset();
}
if (epochSize == null || nbatch >= epochSize)
{
break;
}
}
logger.Info($"Epoch[{epoch}] Time cost={(toc.ElapsedMilliseconds/1000):.000}");
if (epochEndCallback != null || epoch + 1 == endEpoch)
{
executorManager.copy_to(argParams, auxParams);
}
if (epochEndCallback != null)
{
EpochEndParam epochEndParam = new EpochEndParam(epoch, symbol, argParams, auxParams);
foreach (var callitem in epochEndCallback)
{
callitem(epochEndParam);
}
}
// evaluation
if (evalData != null)
{
evalMetric.Reset();
evalData.Reset();
int i = 0;
foreach (var eval_batch in evalData)
{
executorManager.LoadDataBatch(eval_batch);
executorManager.Forward(isTrain: false);
executorManager.UpdateMetric(evalMetric, eval_batch.Label);
if (evalBatchEndCallback != null)
{
var batchEndParams = new BatchEndParam(epoch: epoch,
nbatch: i,
evalMetric: evalMetric,
locals: Thread.CurrentThread.CurrentCulture);
foreach (var call in evalBatchEndCallback)
{
call(batchEndParams);
}
}
i++;
}
var nameValue = evalMetric.get_name_value();
foreach (var item in nameValue)
{
logger.Info($"Epoch[{epoch}] Validation-{item.Name}={item.Value:0.000}");
}
evalData.Reset();
}
}
}