[FRONTEND][TFLITE] get input tensor information from graph #7400

euntaik · 2021-02-03T09:17:15Z

get input tensor information from graph

merge upstream main

…tend

leandron

In general I think it is an improvement to have a default function to come up with shapes in case we don't provide them, so thanks for the initiative @euntaik.

I'm pointing to some similar logic we have somewhere else in the code base that can be ported here. Please have a look.

I any case, it would be good to come up with some testing also to make sure this doesn't break in future.

leandron · 2021-02-09T22:00:36Z

python/tvm/relay/frontend/tflite.py

@@ -3539,7 +3539,62 @@ def get_tensor_name(subgraph, tensor_idx):
    return subgraph.Tensors(tensor_idx).Name().decode("utf-8")


-def from_tflite(model, shape_dict, dtype_dict):
+def get_tensor_shape(subgraph, tensor_idx):
+    """Get the tensor shape.


minor: the name of the argument doesn't match with the actual argument

the types are not specified

Please review all docstring being introduced here for those items above.

leandron · 2021-02-09T22:06:04Z

python/tvm/relay/frontend/tflite.py

+        if shape_dict:
+            shape = shape_dict[model_input_name] if model_input_name in shape_dict else None
+        else:
+            shape = get_tensor_shape(subgraph, model_input)
+        if dtype_dict:
+            dtype = dtype_dict[model_input_name] if model_input_name in dtype_dict else "float32"
+        else:
+            dtype = get_tensor_type(subgraph, model_input)


We have a similar function, that collect the same information being proposed here in TVMC. I agree we should move what is in there, to unify functionality here.

Can you have a look on the function I'm pointing here (below) and spot why are they so different, and in case you agree on what's the best approach, improve it here and remove it there?

tvm/python/tvm/driver/tvmc/frontends.py

Lines 255 to 278 in 2999d03

@staticmethod

def _decode_type(n):

return TFLiteFrontend._tflite_m[n]

@staticmethod

def _input_type(model):

subgraph_count = model.SubgraphsLength()

assert subgraph_count > 0

shape_dict = {}

dtype_dict = {}

for subgraph_index in range(subgraph_count):

subgraph = model.Subgraphs(subgraph_index)

inputs_count = subgraph.InputsLength()

assert inputs_count >= 1

for input_index in range(inputs_count):

input_ = subgraph.Inputs(input_index)

assert subgraph.TensorsLength() > input_

tensor = subgraph.Tensors(input_)

input_shape = tuple(tensor.ShapeAsNumpy())

tensor_type = tensor.Type()

input_name = tensor.Name().decode("utf8")

shape_dict[input_name] = input_shape

dtype_dict[input_name] = TFLiteFrontend._decode_type(tensor_type)

We have a similar function, that collect the same information being proposed here in TVMC. I agree we should move what is in there, to unify functionality here.

Oh, it was there all along. I think I missed your code since I was loading my models in a separate script to put the relay output into my compile passes.

Can you have a look on the function I'm pointing here (below) and spot why are they so different,

I don't see much difference except that your code accounts for models with more than one subgraph.

and in case you agree on what's the best approach, improve it here and remove it there?

My rationale behind making and putting this code in the tflite.py file was:

use the data in the graph since it is already embedded in it.

place the code inside the frontend code since it is dependent on the frontend.

Cool. I think we both agree that it is better to have the funcionality only in the tflite.py, and remove it from TVMC.

So I suggest we keep the one that accounts for many subgraphs, and move it from TVMC to the official frontend? If you agree, feel free to do it in this PR.

Thanks, I will update the PR.

leandron

Looks better now. A few comments below, mostly on TVMC area.

leandron · 2021-02-10T15:10:01Z

python/tvm/relay/frontend/tflite.py

@@ -3539,7 +3539,45 @@ def get_tensor_name(subgraph, tensor_idx):
    return subgraph.Tensors(tensor_idx).Name().decode("utf-8")


-def from_tflite(model, shape_dict, dtype_dict):
+def _decode_type(n):
+    _tflite_m = {


I see this is duplicated in tvmc/frontends.py - is there any reason why we can't reuse this one there?

I missed it. Fixed.

leandron · 2021-02-10T15:13:01Z

python/tvm/driver/tvmc/frontends.py

-        mod, params = relay.frontend.from_tflite(
-            tflite_model, shape_dict=input_shapes, dtype_dict=dtype_dict
-        )
+        mod, params = relay.frontend.from_tflite(tflite_model)


Since we merged #7366, users are able to provide shapes in tvmc from outside, can you have a look on that one and adjust?

Suggested change

mod, params = relay.frontend.from_tflite(tflite_model)

mod, params = relay.frontend.from_tflite(tflite_model, shape_dict=input_shapes, dtype_dict=dtype_dict)

cc @CircleSpin @hogepodge to help

is the from_tflite() now duplicated? (I just looked quickly, might be wrong)

You are right. Sorry for that.

…tend

euntaik · 2021-02-15T05:54:44Z

Any other comments?

leandron · 2021-02-15T09:22:23Z

Any other comments?

Sorry I forgot to check this again after CI.

leandron · 2021-02-15T09:24:44Z

@mbaret @FrozenGene can you have a look on this one, and merge if you think it is ok?

mbaret

This looks like a good change to me.

* [FRONTEND][TFLITE] get input tensor information from graph * remove bare-except * fix lint * delete empty line * comment change * move some of the tflite frontend code from tvmc to tflite.py * update shape and dtype when user provided them * remove unused var. pass user provided shape_dict * remove duplicate code

euntaik added 7 commits February 3, 2021 17:45

Merge pull request #1 from apache/main

3d4e491

merge upstream main

[FRONTEND][TFLITE] get input tensor information from graph

bd5a180

remove bare-except

d4532d7

fix lint

d58353b

delete empty line

e4621ec

comment change

86c70b7

Merge branch 'main' of https://github.com/apache/tvm into tflite_fron…

7ec10c4

…tend

leandron requested changes Feb 9, 2021

View reviewed changes

euntaik added 3 commits February 10, 2021 20:23

move some of the tflite frontend code from tvmc to tflite.py

c954a95

resolve conflicts

dc91132

update shape and dtype when user provided them

53e044a

leandron requested changes Feb 10, 2021

View reviewed changes

euntaik and others added 3 commits February 11, 2021 00:44

remove unused var. pass user provided shape_dict

2ff9b43

remove duplicate code

b80e798

Merge branch 'main' of https://github.com/apache/tvm into tflite_fron…

bc89d27

…tend

leandron approved these changes Feb 15, 2021

View reviewed changes

mbaret approved these changes Feb 15, 2021

View reviewed changes

mbaret merged commit 6187e1c into apache:main Feb 15, 2021

junrushao mentioned this pull request Nov 1, 2021

Apache TVM v0.8 Release Note Candidate #9416

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[FRONTEND][TFLITE] get input tensor information from graph #7400

[FRONTEND][TFLITE] get input tensor information from graph #7400

euntaik commented Feb 3, 2021

leandron left a comment

leandron Feb 9, 2021

leandron Feb 9, 2021

euntaik Feb 10, 2021 •

edited

Loading

leandron Feb 10, 2021

euntaik Feb 10, 2021

leandron left a comment

leandron Feb 10, 2021

euntaik Feb 10, 2021

leandron Feb 10, 2021

euntaik Feb 10, 2021

leandron Feb 10, 2021

euntaik Feb 10, 2021

euntaik commented Feb 15, 2021

leandron commented Feb 15, 2021

leandron commented Feb 15, 2021 •

edited

Loading

mbaret left a comment

	@staticmethod
	def _decode_type(n):
	return TFLiteFrontend._tflite_m[n]

	@staticmethod
	def _input_type(model):
	subgraph_count = model.SubgraphsLength()
	assert subgraph_count > 0
	shape_dict = {}
	dtype_dict = {}
	for subgraph_index in range(subgraph_count):
	subgraph = model.Subgraphs(subgraph_index)
	inputs_count = subgraph.InputsLength()
	assert inputs_count >= 1
	for input_index in range(inputs_count):
	input_ = subgraph.Inputs(input_index)
	assert subgraph.TensorsLength() > input_
	tensor = subgraph.Tensors(input_)
	input_shape = tuple(tensor.ShapeAsNumpy())
	tensor_type = tensor.Type()
	input_name = tensor.Name().decode("utf8")
	shape_dict[input_name] = input_shape
	dtype_dict[input_name] = TFLiteFrontend._decode_type(tensor_type)

	mod, params = relay.frontend.from_tflite(tflite_model)
	mod, params = relay.frontend.from_tflite(tflite_model, shape_dict=input_shapes, dtype_dict=dtype_dict)

[FRONTEND][TFLITE] get input tensor information from graph #7400

[FRONTEND][TFLITE] get input tensor information from graph #7400

Conversation

euntaik commented Feb 3, 2021

leandron left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

euntaik Feb 10, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

leandron left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

euntaik commented Feb 15, 2021

leandron commented Feb 15, 2021

leandron commented Feb 15, 2021 • edited Loading

mbaret left a comment

Choose a reason for hiding this comment

euntaik Feb 10, 2021 •

edited

Loading

leandron commented Feb 15, 2021 •

edited

Loading