public Function(KerasSymbol[] inputs, KerasSymbol[] output, KerasSymbol[] updates)
{
this.output = output;
this.updates = updates;
this.inputs = inputs;
this.is_train = MxNetBackend.LearningPhase();
}
public NDArrayList Call(KerasSymbol[] inputs)
{
var ret_outputs = new List <NDArray>();
foreach (var x in this.output)
{
var bind_values = MxNetBackend.DfsGetBindValues(x);
NDArrayDict data = new NDArrayDict();
for (int i = 0; i < this.inputs.Length; i++)
{
var arr = bind_values.Where(a => a.Key == this.inputs[i].Name).FirstOrDefault();
if (arr.Value != null)
{
data[this.inputs[i].Name] = arr.Value;
}
}
var args = x.Symbol.ListArguments();
List <DataDesc> data_shapes = new List <DataDesc>();
Dictionary <string, OpGradReq> grad_types = new Dictionary <string, OpGradReq>();
for (int i = 0; i < this.inputs.Length; i++)
{
data_shapes.Add(new DataDesc(this.inputs[i].Name, inputs[i].Shape));
grad_types.Add(this.inputs[i].Name, OpGradReq.Null);
}
var executor = x.Symbol.SimpleBind(mx.Cpu(), grad_req: grad_types, kwargs: data_shapes.ToArray());
var arg_dict = executor.ArgmentDictionary();
foreach (var v in arg_dict)
{
if (data.Contains(v.Key))
{
arg_dict[v.Key] = data[v.Key];
}
}
executor.Forward(this.is_train);
var outputs = executor.Outputs;
ret_outputs.Add(outputs[0]);
}
return(ret_outputs);
}
public KerasContext(Context ctx)
{
scope_ctx = MxNetBackend.GetMxNetContexts(ctx);
MxNetBackend._CURRENT_SCOPE_CTX = scope_ctx;
}
