[Tensorflow][frontend]Support for tensor frontend new operators

Development
#1

New to this community and still try to learn more.

At the minute, I would like to use TVM to optimise a tensorflow model that uses some operators from tensorflow 2.2.

Use relay.frontend.from_tensorflow to import the model,

mod, params = relay.frontend.from_tensorflow(graph_def,
                                             layout=layout,
                                             shape=shape_dict)

I got the following information:

NotImplementedError: The following operators are not implemented: {'While', 'NonMaxSuppressionV5', 'TensorListStack', 'TensorListReserve'}

I noticed that there are NonMaxSuppressionV2 V3 operators. However, there is no implementation on ‘While’, ‘TensorListStack’, ‘TensorListReserve’.

I wonder if there is a plan to implement these operators. Or any advice on how to implement these operators.

Thanks in advance.

#2

Current TF frontend parser is for tf 1.x. Community has some initial discussions about tf 2.x support but I’m not sure about the status.

#3

Hi Kevin,

Thanks for your reply.

I am trying to adapt the model to tf1, e.g., ‘NonMaxSuppressionV5’ in tf2 to ’ NonMaxSuppressionV3’, which is supported by tensorflow frontend.

The computational graph .pb file seems fine now.

However, when running

mod, params = relay.frontend.from_tensorflow(graph_def,
                                             layout=layout,
                                             shape=shape_dict)

still get three unsupported operators related to the pre-processing the input data.

NotImplementedError: The following operators are not implemented: {'TensorListStack', 'TensorListReserve', 'While'}

Thanks again.

#4

Looks like some op names have changed across tf 1.x and 2.x.

1 Like
#5

Different operators:

[non_max_suppression/NonMaxSuppressionV3: NonMaxSuppressionV3] in tf1 that is supported by the current version of TVM. I looked into the source code of tensorflow frontend and there are two operators NMS, NonMaxSuppressionV2 and V3.

Look forward to seeing V5. :wink:

#6

Hi Kevin,

After a bit break, I got back to the issues.

Now I got rid of the operators from the tensorflow side with no un-supported operators from tensorflow frontend.

However, I got the following issue:

     57         except AttributeError:
     58             raise AttributeError(
---> 59                 "%s has no attribute %s" % (str(type(self)), name))
     60 
     61     def __hash__(self):

AttributeError: <class 'tvm.relay.expr.Call'> has no attribute name_hint

This seems to be similar to this issue raised quite a while ago that mentioned you have fixed it.

I pulled the latest version tvm. If you are interested, I can share the model file with you.

Thanks in advance.

Is NonMaxSupression supported in the Tensorflow frontend?
Is NonMaxSupression supported in the Tensorflow frontend?
#7

@cam145 perhaps you could try to identify which op introduces the name_hint AttributeError, then look into the details(like params, inputs, etc.) that node has.