TensorFlow Lite conversion of frozen graph error

[Dayananda-V]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04.2
• TensorFlow installed from (source or binary): Source
• TensorFlow version (or github SHA if from source): 2f9cc84

Provide the text output from tflite_convert
While converting faster_rcnn_nas from tensoflow trained model to tensorflow lite found some error.

Unsupported Control Flow Ops:
1. Enter
2. Exit
3. Merge
4. Switch
5. LoopCond

Unsupported Lite Ops:
1. NonMaxSuppressionV2
2. TensorArrayGatherV3
3. TensorArrayReadV3
4. TensorArrayScatterV3
5. TensorArraySizeV3
6. TensorArrayV3
7. TensorArrayWriteV3

Unsupported Quantize Ops:
1. Div
2. Exp
3. Cast
4. Fill
5. Range
6. Size
7. Tile
8. ZerosLike
9. Unpack 

Also, please include a link to a GraphDef or the model if possible.

Any other info / logs

Include any logs or source code that would be helpful to diagnose the problem. If including tracebacks, please include the full traceback. Large logs and files should be attached.
faster_rcnn_nas_lite_error.txt

[amitsrivastava78]

@jdduke This seems to be a problem faced by alot of people in the past, is there any plans of implementing the data flow ops in the tflite ? Also would like to know your take on implementing it along with flow_in's which is tricky thing to do.
If you are ok would like to raise on RFC for this and take it forward for implementation.

Regards
Amit

[jdduke]

@miaout17 can you de-dupe against our tracking control flow issue? We'll be improving the documentation around control flow-related conversion in the near future (an area we're actively working to improve over the next several quarters).

[alanchiao]

Would it be helpful at all for us to add an error message in TOCO where it points directly to a FAQ or Github issue (for control flow / LSTMs) when there is an Unsupported Control Flow Op?

[miaout17]

Let's use #28485 for tracking TensorFlow Lite control flow support. A RFC already existed.
@alanchiao sounds a good idea.

[ANSHUMAN87]

Unpack PR is already raised & approved: #27881

[Dayananda-V]

Range op quantization support found under #27103

[amitsrivastava78]

Quantization support for Exp is raised as PR now.

Regards
Amit

[garikana]

The PR for Quantization support for Exp was closed. Any update on when/if this will be include in future?

Thanks,
Alex

[tilakrayal]

@Dayananda-V,
The related issue #28485 was closed & where it was mentioned that With a new converter now the default. TFLite should support Control Flow V2 ops. If you have a model with control flow ops, please try to enable control flow v2.
to enable Control flow v2: tf.enable_control_flow_v2() and also the PR #27881 was merged for Quantization support added for Unpack Op. Thank you!

[github-actions]

This issue was closed because it has been inactive for 7 days since being marked as stale. Please reopen if you'd like to work on this further.