How to convert these models to TF-Lite models and run it

[rayeh5]

I can't find useful information to convert these models to TF-Lite and run it at Linux environment. Could someone give some hints? Thanks!

[pooyaww]

• Train your desired model
• Freez it by the provided python script ang generate a pb file.
• Use TensorFlow Lite Converter to convert .pb file to .tflite file
• Use it in TF-lite ! check the documentation

[rayeh5]

• Train your desired model
• Freez it by the provided python script ang generate a pb file.
• Use TensorFlow Lite Converter to convert .pb file to .tflite file
• Use it in TF-lite ! check the documentation

Thanks for the reply. I think that I don'e explain my question well.

I try to convert the Pretrained_models/CNN/CNN_S.pb via the following commands:

toco
--input_file=./CNN_S.pb
--input_format=TENSORFLOW_GRAPHDEF
--output_format=TFLITE
--output_file=./CNN_S.tflite
--inference_type=FLOAT
--input_shape=1,49,10,1
--input_array=x
--input_data_type=FLOAT
--output_arrays=final_fc

The following are the error messages: From the error messages, it seems that:

1. TF-Lite doesn't support operation: "DecodeWav", "AudioSpectrogram" and "Mfcc";
2. I don't know how to assign the "output_arrays", where can find this information?

---

Error messages:

N$ ./my_toco_float.sh 2018-11-30 14:23:35.412174: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1233] Converting unsupported operation: DecodeWav 2018-11-30 14:23:35.412606: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1233] Converting unsupported operation: AudioSpectrogram 2018-11-30 14:23:35.412876: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1233] Converting unsupported operation: Mfcc 2018-11-30 14:23:35.414064: F tensorflow/contrib/lite/toco/tooling_util.cc:711] Check failed: model.HasArray(output_array) Output array not found: final_fc ./my_toco_float.sh: line 10: 3577 Aborted (core dumped) toco --input_file=./CNN_S.pb --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --output_file=./CNN_S.tflite --inference_type=FLOAT --input_shape=1,49,10,1 --input_array=x --input_data_type=FLOAT --output_arrays=final_fc

---

Thanks for your time.

[pooyaww]

@rayeh5 Obviously it is an issue with toco which currently does not support all operations. The trained models here include some modules to extract PCM data from wave files, finding spectrum and computing MFCC. You can find all supported operation at https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/g3doc/tf_ops_compatibility.md . As tensorflow has a good back-end for Audio recognition, I think you can open in their repo and asking for a support regarding these operations.

[marjanemd]

What are the input array and output arrays here? I am using tensor board and the .pbtxt format to check the input and output names for the 'DNN', the output is correct, but it is showing that the input is fingerprint_input and dropout_prob, and when using that it gives me the error of :

ValueError: Invalid tensors 'fingerprint_input,dropout_prob' were found,

Any hints will be really appreciated?