[Frontend][TFLite] fix detection_postprocess's non_max_suppression_attrs["force_suppress"]

[czh978]

I found the value should be "True"

[leandron]

Thanks for the fix @czh978.

Do you mind providing some context on why force_supress needs to be set this way? Can you give one example of something that wasn't working before this change, and works after the change?

[czh978]

Thanks for the fix @czh978.

Do you mind providing some context on why force_supress needs to be set this way? Can you give one example of something that wasn't working before this change, and works after the change?

tvm/python/tvm/relay/frontend/tflite.py

Lines 3278 to 3284 in 3d41ac3

	if "use_regular_nms" in custom_options:
	if custom_options["use_regular_nms"]:
	raise tvm.error.OpAttributeUnImplemented(
	"use_regular_nms=True is not yet supported for operator {}.".format(
	"TFLite_Detection_PostProcess"
	)
	)

The value of use_regular_nms is always Flase，becasue use_regular_nms=True is not yet supported in tvm。
https://github.com/tensorflow/tensorflow/blob/47c541330813f575057d7af90ef55985baca87eb/tensorflow/lite/kernels/detection_postprocess.cc#L882-L887
The result will suppress all detections regardless of class_id in tflite，when the value of use_regular_nms is always Flase.
Therefore,the value of force_suppress should be true.
tflite_graph_with_postprocess.zip
There is a model，you can try adjust the code in test_forward.py of tflite,where is in

tvm/tests/python/frontend/tflite/test_forward.py

Lines 4309 to 4336 in 3d41ac3

	def test_detection_postprocess():
	"""Detection PostProcess"""
	tf_model_file = tf_testing.get_workload_official(
	"http://download.tensorflow.org/models/object_detection/"
	"ssd_mobilenet_v2_quantized_300x300_coco_2019_01_03.tar.gz",
	"ssd_mobilenet_v2_quantized_300x300_coco_2019_01_03/tflite_graph.pb",
	)
	converter = tf.lite.TFLiteConverter.from_frozen_graph(
	tf_model_file,
	input_arrays=["raw_outputs/box_encodings", "raw_outputs/class_predictions"],
	output_arrays=[
	"TFLite_Detection_PostProcess",
	"TFLite_Detection_PostProcess:1",
	"TFLite_Detection_PostProcess:2",
	"TFLite_Detection_PostProcess:3",
	],
	input_shapes={
	"raw_outputs/box_encodings": (1, 1917, 4),
	"raw_outputs/class_predictions": (1, 1917, 91),
	},
	)
	converter.allow_custom_ops = True
	converter.inference_type = tf.lite.constants.FLOAT
	tflite_model = converter.convert()
	np.random.seed(0)
	box_encodings = np.random.uniform(size=(1, 1917, 4)).astype("float32")
	class_predictions = np.random.uniform(size=(1, 1917, 91)).astype("float32")
	tflite_output = run_tflite_graph(tflite_model, [box_encodings, class_predictions])

there is my example:

It will get a Mismatched elements.

[leandron]

Is there any publicly available network that we can observe this behaviour and create a test case?

[leandron]

Thanks for adding a test case @czh978. I'm happy for this to be merged once CI is green.

The last thing we need is for the PR description to be improved with a description of your findings that as use_regular_nms is set as False in TVM, we need to have force_supress to be True, to avoid all detections to be suppressed.

Please expand that with the reasoning why this change is needed and point out that it reproduces on networks like ssd_mobilenet_v2 and other examples if you know any.

[czh978]

Thanks for adding a test case @czh978. I'm happy for this to be merged once CI is green.

The last thing we need is for the PR description to be improved with a description of your findings that as use_regular_nms is set as False in TVM, we need to have force_supress to be True, to avoid all detections to be suppressed.

Please expand that with the reasoning why this change is needed and point out that it reproduces on networks like ssd_mobilenet_v2 and other examples if you know any.

I've modified the commit message to meet the requirements of the new RFC, could you review it if you're available?Thanks.

[czh978]

@leandron @AndrewZhaoLuo