Issues exporting ssdlite_mobilenet_v2 to tensorflow-lite

[uelordi01]

System information

• What is the top-level directory of the model you are using:

  • ssdlite_mobilenet_v2_coco_2018_05_09 pretrained model

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux ubuntu 16.04

• TensorFlow installed from binary (CPU):

• TensorFlow version (use command below):

  • 1.5

Describe the problem

Hi:

I am trying to export ssdlite_Mobilenet_v2 model to tf-lite: I downloaded from this path I tried with these two ways:

1.- following this tutorial with export_tflite_ssd_graph.py, using the checkpoint

• I use export_tflite_ssd_graph.py like this:

  
  export CHECKPOINT_PATH=/home/VICOMTECH/uelordi/projects/tflite_models/ssdlite_mobilenet_v2_coco_2018_05_09/model.ckpt
  
  export OUTPUT_DIR=/home/VICOMTECH/uelordi/projects/tflite_models/ssdlite_mobilenet_v2_coco_2018_05_09/tflite
  
  python /home/VICOMTECH/uelordi/SDK/tensorflow1/models/models/research/object_detection  /export_tflite_ssd_graph.py \
  --pipeline_config_path=$CONFIG_FILE \
  --trained_checkpoint_prefix=$CHECKPOINT_PATH \
  --output_directory=$OUTPUT_DIR \
  --add_postprocessing_op=true

  • then I created toco script:

    toco \
    --input_file=tflite_graph.pb \
    --output_file=latest_ssdlite_mobilenetv2.tflite \
    --input_format=TENSORFLOW_GRAPHDEF \
    --input_shapes=1,300,300,3 \
    --output_format=TFLITE \
    --input_arrays=normalized_input_image_tensor \
    --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3'  \
    --inference_type=FLOAT \
    --mean_values=128 \
    --std_values=128 \
    --change_concat_input_ranges=false \

    and I got this errors:

Some of the operators in the model are not supported by the standard TensorFlow Lite runtime. If you have a custom implementation for them you can disable this error with --allow_custom_ops. Here is a list of operators for which you will need custom implementations: DIV, Squeeze, TFLite_Detection_PostProcess.

So I added allow-custom-ops, and when I use the tf-lite interperter in android:

d.tfLite = new Interpreter(loadModelFile(assetManager, modelFilename));

And as I supposed I have custom operation error.

Internal error: Cannot create interpreter: Didn't find custom op for name 'DIV' with version 1 Didn't find custom op for name 'Squeeze' with version 1 Registration failed.

2.- using toco with the pretrained model.

my toco script is:

toco \
--input_file=frozen_inference_graph.pb \
--output_file=ssd_lite_v2.tflite \
  --input_format=TENSORFLOW_GRAPHDEF \
--input_shapes=1,300,300,3 \
  --output_format=TFLITE \
--input_arrays=normalized_input_image_tensor \
--output_arrays='detection_boxes,detection_scores,detection_classes,num_detections' \
--inference_type=FLOAT \
--mean_values=128 \
--std_values=128 \
--change_concat_input_ranges=false \

As I have some errors with the operators I added allow-custom-ops, and when I use the tf-lite interperter in android:

d.tfLite = new Interpreter(loadModelFile(assetManager, modelFilename));

But as I supposed I have custom operation error.

Internal error: Cannot create interpreter: Didn't find custom op for name 'DIV' with version 1 Didn't find custom op for name 'Squeeze' with version 1 Registration failed.

So my questions are:

• What makes different export_tflite_ssd_graph.py (with the checkpoint) from a common toco script with pretrained model? It creates custom operations? complementary postprocessing operations? In case of toco pretrained problems:
• This means that I have to create my custom operation of DIV and Squeeze operations?

[aifollower]

I have a similar problem! 2018-08-07 13:21:07.345473: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1053] Converting unsupported operation: SquaredDifference 2018-08-07 13:21:07.361241: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1053] Converting unsupported operation: TFLite_Detection_PostProcess

[kate-kate]

same problem in Android. I've tried first variant but have an error like "Cannot create interpreter: Didn't find custom op for name 'TFLite_Detection_PostProcess'". Then I paid attention that there were some notifications during toco converting, like "Converting unsupported operation: TFLite_Detection_PostProcess". But tflite file was created anyway. Then I've tried to generate tflite without any warnings, and I just can't. I also have tried your second variant, but it causes error " Output array not found: detection_boxes". Well, at least it doesn't generate invalid file P.S. I used ssd_mobilenet_v1_coco pretrained model

[achowdhery]

If the tflite file was created, it is valid. Does the file run in the app? FYI, the warning converting unsupported operation is there because the postprocessing op is a custom TFLite op (not the regular Tensorflow op)

[kate-kate]

@achowdhery the file does not run due to " Cannot create interpreter: Didn't find custom op for name 'TFLite_Detection_PostProcess'" exception

[achowdhery]

@kate-kate I think this is a version issue where you dont have the correct version of Tensorflow which has the op available. Please provide details on how you installed Tensorflow. Is this from within ML Kit?

[kate-kate]

@achowdhery I installed tensorflow from source just like it is described in the tensorflow site Here is tf_env_collect.sh result, maybe it will help

== cat /etc/issue ===============================================
Darwin MacBook-Pro-Ekaterina.local 17.7.0 Darwin Kernel Version 17.7.0: Thu Jun 21 22:53:14 PDT 2018; root:xnu-4570.71.2~1/RELEASE_X86_64 x86_64
Mac OS X 10.13.6

== are we in docker =============================================
No

== compiler =====================================================
Apple LLVM version 9.1.0 (clang-902.0.39.2)
Target: x86_64-apple-darwin17.7.0
Thread model: posix
InstalledDir: /Library/Developer/CommandLineTools/usr/bin

== uname -a =====================================================
Darwin MacBook-Pro-Ekaterina.local 17.7.0 Darwin Kernel Version 17.7.0: Thu Jun 21 22:53:14 PDT 2018; root:xnu-4570.71.2~1/RELEASE_X86_64 x86_64

== check pips ===================================================
numpy                              1.14.5   
numpydoc                           0.7.0    
protobuf                           3.6.0    
tensorflow                         1.10.0   

== check for virtualenv =========================================
False

== tensorflow import ============================================
tf.VERSION = 1.10.0
tf.GIT_VERSION = v1.10.0-rc1-19-g656e7a2b34
tf.COMPILER_VERSION = v1.10.0-rc1-19-g656e7a2b34
Sanity check: array([1], dtype=int32)

== env ==========================================================
LD_LIBRARY_PATH is unset
DYLD_LIBRARY_PATH is unset

== nvidia-smi ===================================================
tf_env_collect.sh: line 105: nvidia-smi: command not found

== cuda libs  ===================================================

[patrick-ucr]

I am having a similar issue here but my tflite model works on Android demo only that it returns many false positive detections.

I am not sure if this toco log suggests the problem tensorflow/lite/toco/import_tensorflow.cc:1332] Converting unsupported operation: TFLite_Detection_PostProcess

[vivekpanchal]

@achowdhery any update regarding this issue ? i am facing the same problem when converting a custom trained model to tflite .

[akshay512]

I am facing the same problem did anyone fix?

[pritasam]

I am having same issues any updates ??