Class unify

[again4you]

AS-IS

• 2 TensorsInfo for each Tensor information
• 2 TensorsData for each Tensor Buffer
• SingleShot class

class diagram

code

byte[] in_buffer = new byte[3 * 224 * 224 * 1];
byte[] out_buffer;
string model_path = ResourcePath + "models/mobilenet_v1_1.0_224_quant.tflite";

TensorsInfo in_info;
TensorsInfo out_info;
TensorsData in_data;
TensorsData out_data;

/* Set input & output TensorsInfo */
in_info = new TensorsInfo();
in_info.AddTensorInfo(TensorType.UInt8, new int[4] { 3, 224, 224, 1 });

out_info = new TensorsInfo();
out_info.AddTensorInfo(TensorType.UInt8, new int[4] { 1001, 1, 1, 1 });

/* Create single inference engine */
SingleShot single = new SingleShot(model_path, in_info, out_info);

/* Set input data */
in_data = in_info.GetTensorsData();
in_data.SetTensorData(0, in_buffer);

/* Single shot invoke */
out_data = single.Invoke(in_data);

/* Get output data from TensorsData */
out_buffer = out_data.GetTensorData(0);

/* Release Single inference instance */
single.Dispose();

/* clean up */
in_data.Dispose();
out_data.Dispose();
in_info.Dispose();
out_info.Dispose();

[again4you]

To-Be (First draft)

• Newly add Tensors Class
• Keep TensorsInfo and TensorsData classes and its related method for backword compatibility
• Pros
  • Shorter than before and easy to maintain
  • Support previous APIs
• Cons
  • New Tensors class ?!
  • SingleShot.Invoke() returns different classes.

Class diagram

code

var in_buffer = new byte[3 * 224 * 224 * 1];
string model_path = ResourcePath + "models/mobilenet_v1_1.0_224_quant.tflite";

/* Set Input Tensors */
Tensors inTensors = new Tensors();
inTensors.AddTensor(TensorType.UInt8, new int[4] { 3, 224, 224, 1 });

/* Set Output Tensors */
Tensors outTensors = new Tensors();
outTensors.AddTensor(TensorType.UInt8, new int[4] { 1001, 1, 1, 1 });

/* Create single inference engine */
SingleShot single = new SingleShot(_modelPath, inTensors, outTensors);

/* Set input data */
inTensors.SetData(0, in_buffer);

/* Single shot invoke */
outTensors = single.Invoke(inTensors);

/* Release Single Inference instance */
single.Dispose();

/* clean up */
inTensors.Dispose();
outTensors.Dispose();

[again4you]

To-Be (Second draft)

• TensorsData class has a TensorsInfo class
• Pros
  • No class are newly added.
• Cons
  • TensorsData has lots of method for TensorsInfo
  • SingleShot class should check the TensorsInfo validataion of TensorsData class

Class Diagram

code

var in_buffer = new byte[3 * 224 * 224 * 1];
string model_path = ResourcePath + "models/mobilenet_v1_1.0_224_quant.tflite";

/* Set input TensorsData info & data */
/* With TensorsInfo class */
in_info = new TensorsInfo();
in_info.AddTensorInfo(TensorType.UInt8, new int[4] { 3, 224, 224, 1 });

TensorsData in_data = new TensorsData (in_info)
in_data.SetData(0, in_buffer);

/* Set output TensorsData info */
out_info = new TensorsInfo();

/* Without TensorsInfo class */
TensorsData out_data = TensorsInfo();
out_data.AddTensorInfo(TensorType.UInt8, new int[4] { 1001, 1, 1, 1 });

/* Create single inference engine */
SingleShot single = new SingleShot(_modelPath, in_data, out_data);

/* Single shot invoke */
out_data = single.Invoke(in_data);

/* Release Single Inference instance */
single.Dispose();

/* clean up */
in_data.Dispose();
out_data.Dispose();
in_info.Dispose();