What is tensorflow concrete function outputs correspond to structured_outputs?

Question

I trained my customized ssd_mobilenet_v2 using TensorFlow2 Object Detection API.
After training completed, I used exporter_main_v2.py to export a saved_model of my customized model.

If I load saved_model by TensorFlow2, it seem there are two kind of output format.

import tensorflow as tf
saved_model = tf.saved_model.load("saved_model")
detect_fn = saved_model["serving_default"]
print(detect_fn.outputs)
'''
[<tf.Tensor 'Identity:0' shape=(1, 100) dtype=float32>,
 <tf.Tensor 'Identity_1:0' shape=(1, 100, 4) dtype=float32>,
 <tf.Tensor 'Identity_2:0' shape=(1, 100) dtype=float32>,
 <tf.Tensor 'Identity_3:0' shape=(1, 100, 7) dtype=float32>,
 <tf.Tensor 'Identity_4:0' shape=(1, 100) dtype=float32>,
 <tf.Tensor 'Identity_5:0' shape=(1,) dtype=float32>,
 <tf.Tensor 'Identity_6:0' shape=(1, 1917, 4) dtype=float32>,
 <tf.Tensor 'Identity_7:0' shape=(1, 1917, 7) dtype=float32>]
'''
print(detect_fn.structured_outputs)
'''
{'detection_classes': TensorSpec(shape=(1, 100), dtype=tf.float32, name='detection_classes'),
 'detection_scores': TensorSpec(shape=(1, 100), dtype=tf.float32, name='detection_scores'),
 'detection_multiclass_scores': TensorSpec(shape=(1, 100, 7), dtype=tf.float32, name='detection_multiclass_scores'),
 'num_detections': TensorSpec(shape=(1,), dtype=tf.float32, name='num_detections'),
 'raw_detection_boxes': TensorSpec(shape=(1, 1917, 4), dtype=tf.float32, name='raw_detection_boxes'),
 'detection_boxes': TensorSpec(shape=(1, 100, 4), dtype=tf.float32, name='detection_boxes'),
 'detection_anchor_indices': TensorSpec(shape=(1, 100), dtype=tf.float32, name='detection_anchor_indices'),
 'raw_detection_scores': TensorSpec(shape=(1, 1917, 7), dtype=tf.float32, name='raw_detection_scores')}
'''

Then, I try to convert this saved_model to onnx format using tf2onnx.
However, the outputs of onnxruntime was a list.
By the shape of result in the list, I think that the sequence is same as detect_fn.outputs

import numpy as np
import onnxruntime as rt

sess = rt.InferenceSession("model.onnx")
input_name = sess.get_inputs()[0].name
pred_onx = sess.run(None, {input_name: np.zeros((1,300,300,3), dtype=np.uint8)})
print(pred_onx) # a list
print([i.shape for i in pred_onx])
'''
[(1, 100),
 (1, 100, 4),
 (1, 100),
 (1, 100, 7),
 (1, 100),
 (1,),
 (1, 1917, 4),
 (1, 1917, 7)]
'''

Because there is some shape of result which is same as others, so it become hard to recognized.
Is there any document talk about this relationship that I can refer?

Answer 1

After I looked closely into the values in outputs. I found the mapping relationship below.

def result_mapper(outputs):
    result_dict = dict()
    result_dict["num_detections"] = outputs[5]
    result_dict["raw_detection_scores"] = outputs[7]
    result_dict["raw_detection_boxes"] = outputs[6]
    result_dict["detection_multiclass_scores"] = outputs[3]
    result_dict["detection_boxes"] = outputs[1]
    result_dict["detection_scores"] = outputs[4]
    result_dict["detection_classes"] = outputs[2]
    result_dict["detection_anchor_indices"] = outputs[0]
    return result_dict

Answer 2

Had the same question and after some hours of stepping through the debugger found they are... in sorted order. The method determining iteration order is here.

In the case of dict instances, the sequence consists of the values, sorted by key to ensure deterministic behavior.