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();
}
}
}
private static void _train_multi_device(Symbol symbol, List <Context> ctx, List <string> arg_names,
List <string> param_names, List <string> aux_names, Dictionary <string, NDArray> arg_params,
Dictionary <string, NDArray> aux_params, int begin_epoch, int end_epoch, int?epoch_size, Optimizer optimizer,
IDataIter train_data, IDataIter eval_data, EvalMetric eval_metric, List <Action> epoch_end_callback,
List <Action <BatchEndParam> > batch_end_callback, KVStore kvstore, bool update_on_kvstore, ILog logger, List <int> work_load_list,
Monitor monitor, Action eval_batch_end_callback, SymbolGenerate sym_gen)
{
if (logger == null)
{
logger = LogManager.GetLogger("");
}
var executor_manager = new DataParallelExecutorManager(symbol: symbol,
sym_gen: sym_gen,
ctx: ctx,
train_data: train_data,
param_names: param_names,
arg_names: arg_names,
aux_names: aux_names,
work_load_list: work_load_list,
logger: logger);
if (monitor != null)
{
executor_manager.install_monitor(monitor);
}
executor_manager.set_params(arg_params, aux_params);
Action <int, NDArray, NDArray> updater = null;
if (!update_on_kvstore)
{
updater = Optimizer.get_updater(optimizer);
}
if (kvstore != null)
{
_initialize_kvstore(kvstore: kvstore,
param_arrays: executor_manager.param_arrays,
arg_params: arg_params,
param_names: executor_manager.param_names,
update_on_kvstore: update_on_kvstore);
}
if (update_on_kvstore)
{
kvstore.set_optimizer(optimizer);
}
//Now start training
for (int epoch = 0; epoch < end_epoch - begin_epoch; epoch++)
{
// Training phase
Stopwatch toc = new Stopwatch();
toc.Start();
eval_metric.Reset();
var nbatch = 0;
// Iterate over training data.
while (true)
{
var do_reset = true;
foreach (var data_batch in train_data)
{
executor_manager.load_data_batch(data_batch);
monitor?.Tic();
executor_manager.Forward(is_train: true);
executor_manager.Backward();
if (update_on_kvstore)
{
_update_params_on_kvstore(
executor_manager.param_arrays,
executor_manager.grad_arrays,
kvstore);
}
else
{
_update_params(executor_manager.param_arrays,
executor_manager.grad_arrays,
updater: updater,
num_device: ctx.Count,
kvstore: kvstore);
}
monitor?.toc_print();
// evaluate at end, so we can lazy copy
executor_manager.update_metric(eval_metric, data_batch.label);
nbatch += 1;
//batch callback (for print purpose)
if (batch_end_callback != null)
{
var batch_end_params = new BatchEndParam(epoch: epoch,
nbatch: nbatch,
eval_metric: eval_metric,
locals: Thread.CurrentThread.CurrentCulture);
foreach (var call in batch_end_callback)
{
call(batch_end_params);
}
}
if (epoch_size != null && nbatch >= epoch_size)
{
do_reset = false;
break;
}
}
if (do_reset)
{
logger.Info($"Epoch[{epoch}] Resetting Data Iterator");
train_data.Reset();
}
if (epoch_size == null || nbatch >= epoch_size)
{
break;
}
}
logger.Info($"Epoch[{epoch}] Time cost={(toc.ElapsedMilliseconds/1000):.000}");
}
}