In this week, I tried to find out input/output size of each layers. Following is the example code from Ekaterina.

  const tflite::Interpreter* interpreter = ...; // Initialize interpreter
  for (int op_index : interpreter->execution_plan()) {
    const auto* op_and_reg = interpreter->node_and_registration(op_index);
    if (op_and_reg->second.builtin_code == kTfLiteBuiltinConv2d) {
      // Parse operation basing on its type.
      // See other types in tensorflow/lite/builtin_ops.h
    }
    for (const auto input_idx :
         tflite::TfLiteIntArrayView(op_and_reg->first.inputs)) {
      // Access operation input using its index.
    }
    for (const auto output_idx :
         tflite::TfLiteIntArrayView(op_and_reg->first.outputs)) {
      // Access operation output using its index.
    }
  }

For starters, I wanted to change it to python version of code. I tried densenet at first which has no quantization. TFLite models can be downloaded here.

And Below is the first part of Densenet’s Architecture.

Densenet Archtecture

I coded the following by referring tensorflow’s interpreter.py code.

import tensorflow as tf
SAVED_MODEL_PATH = "./MODLES/densenet.tflite"

interpreter = tf.lite.Interpreter(model_path=SAVED_MODEL_PATH)

for op_index in range(interpreter._interpreter.NumNodes()):
    op_and_reg = interpreter2._get_op_details(op_index)
    # to do.. 

_get_op_details function is experimental method, but anyway it returns all of layers sequentially. So I thought with this code, I could know the kind of each (like Conv2d, Maxpooling etc) and the input/output size of the layers. But It was not what I expected.

_get_ops_details() and get_tensor_details()

_get_ops_details method let me know the order of each layer, but its input and output have numbers that I do not know what it means.

BTW, get_tensor_details() function returns all of the tensors and It looks quite understandable numbers. But It doesn’t retuns sequentially.

---

So In this week… What I would like to know is..

To Check List

• What is kTFLiteBuiltinConv2d? Does it just mean ‘this is Conv2D’?
• The Difference between tensor and ops(?)