Convert my saved model to TFlite

[ankit]

I am building a custom object detector and I am following the object-detection template . As an output I am getting the saved model and the frozen inference graph further i want to convert it into a TFlite file and for that I am using the toco converter and the following commands:
!python /floyd/home/export_tflite_ssd_graph.py \
--pipeline_config_path /floyd/home/object_detection/samples/configs/ssd_mobilenet_v2_coco.config \
--trained_checkpoint_prefix /floyd/home/trained_models/ssdlite_mobilenet_v2_coco_2018_05_09/model.ckpt-100 \
--output_directory /floyd/home/ \
add_postprocessing_op True \
--max_detections 10

but I am getting an error:

File "/floyd/home/export_tflite_ssd_graph.py", line 143, in <module>
    tf.app.run(main)
  File "/usr/local/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 126, in run
    _sys.exit(main(argv))
  File "/floyd/home/export_tflite_ssd_graph.py", line 139, in main
    FLAGS.max_classes_per_detection, use_regular_nms=FLAGS.use_regular_nms)
  File "/floyd/home/object_detection/export_tflite_ssd_graph_lib.py", line 277, in export_tflite_graph
    if pipeline_config.model.ssd.feature_extractor.HasField('fpn'):
ValueError: Unknown field fpn.

Can you help me resolve this .
Is there any other way to convert my model into TFlite ?

[redeipirati]

Hi @ankit,

The config file you are using doesn't provide the FPN field in the Feature Extractor. Commenting lines 277-278 of this file should fix it.

Hope that helps.

[ankit]

There is an error of shapes not matching in lhs and rhs. Is there a tutorial or a job in which output tflite has been generated?

[redeipirati]

Hi @ankit,

There is an error of shapes not matching in lhs and rhs.

It looks like a filter mismatch in featureExtractor/MobilenetV2/layer_19_2_Conv2d_5_3x3_s2_128/weights. I'm not sure about the cause of this. My recommendation is to open a ticket on this GH repo: TensorFlow Models.

Is there a tutorial or a job in which output tflite has been generated?

Unfortunately, there isn't any tutorial available about TFLite at the moment. I'll take this as a Feature Request and add this to our backlog.

Hope that helps.

[robmarkcole]

The following article would be a good starting point:

Inferences from a TF Lite model — Transfer Learning on a Pre-trained Model