Tensorflow TocoConverter 给出 toco_from_protos 错误

Question

I basically used freeze_graph to freeze the model and then tried to use TocoConverter to convert the model to tflite, however it gives me the error:

RuntimeError: TOCO failed see console for info.
b'/bin/sh: 1: toco_from_protos: not found\n'
None 

Below is my code:

graph_def_file = './frozen_model2.pb'
input_arrays = ['IteratorGetNext']
output_arrays = ['model/fc_result/prediction/BiasAdd']

converter = tf.contrib.lite.TocoConverter.from_frozen_graph(
graph_def_file, input_arrays, output_arrays)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

Any help would be much appreciated!!!

Answer 1

The error b'/bin/sh: 1: toco_from_protos: not found\\n' indicates that the shell command toco_from_protos is not found. I'm guessing toco --help doesn't work for you either.

I built Tensorflow 1.9.0 from source so I assume I must have screwed up a step where shell commands like toco and toco_from_protos get set up.

Anyway, my traceback was this:

  File "/home/casey/anaconda3/envs/mnist/lib/python3.6/site-packages/tensorflow/contrib/lite/python/lite.py", line 330, in convert
    dump_graphviz_video=self.dump_graphviz_video)
  File "/home/casey/anaconda3/envs/mnist/lib/python3.6/site-packages/tensorflow/contrib/lite/python/convert.py", line 263, in toco_convert
    input_data.SerializeToString())
  File "/home/casey/anaconda3/envs/mnist/lib/python3.6/site-packages/tensorflow/contrib/lite/python/convert.py", line 107, in toco_convert_protos
    (stdout, stderr))
RuntimeError: TOCO failed see console for info.
b'/bin/sh: 1: toco_from_protos: not found\n'
None

If you follow the traceback to line 107 in [python path]/site-packages/tensorflow/contrib/lite/python/convert.py , you'll see there is some logic in the definition for toco_convert_protos() to run toco_from_protos as a shell command:

# TODO(aselle): When toco does not use fatal errors for failure, we can
# switch this on.
if not _toco_from_proto_bin:
    return _toco_python.TocoConvert(
        model_flags_str, toco_flags_str, input_data_str)

with _tempfile.NamedTemporaryFile() as fp_toco, \
           _tempfile.NamedTemporaryFile() as fp_model, \
           _tempfile.NamedTemporaryFile() as fp_input, \
           _tempfile.NamedTemporaryFile() as fp_output:
    fp_model.write(model_flags_str)
    fp_toco.write(toco_flags_str)
    fp_input.write(input_data_str)
    fp_model.flush()
    fp_toco.flush()
    fp_input.flush()

    cmd = [
        _toco_from_proto_bin, fp_model.name, fp_toco.name, fp_input.name,
        fp_output.name
    ]
    cmdline = " ".join(cmd)
    proc = _subprocess.Popen(
        cmdline,
        shell=True,
        stdout=_subprocess.PIPE,
        stderr=_subprocess.STDOUT,
        close_fds=True)
    stdout, stderr = proc.communicate()
    exitcode = proc.returncode
    if exitcode == 0:
      stuff = fp_output.read()
      return stuff
    else:
      raise RuntimeError("TOCO failed see console for info.\n%s\n%s\n" %
                         (stdout, stderr))

If your toco and toco_from_protos shell commands aren't working then obviously this step will fail.

I made this change as a workaround in [python path]/site-packages/tensorflow/contrib/lite/python/convert.py , to force the first case to return:

  if _toco_from_proto_bin:
  # if not _toco_from_proto_bin: # comment out this line to force return
    return _toco_python.TocoConvert(
        model_flags_str, toco_flags_str, input_data_str)

Janky, but it got the job done for me. The proper solution would probably be to reinstall a version of Tensorflow that has working toco or redo the bazel build if you are building from source.

Answer 2

Make sure that the toco scripts are in your search path. Check PATH variable.

It seems that they are installed in /home/user/.local/bin so add this to PATH variable.

Answer 3

You have two options

1. Download and build tensorflow from source
2. Use google collab if you don't want to install tensorflow from source Here i made video on how to fix this Error

Answer 4

I just experienced this problem when quantizing a model into integer-precision.

RuntimeError: TOCO failed see console for info.
b'/bin/sh: 1: toco_from_protos: not found\n'

The key is to find where does toco_from_protos actually reside.

I use locate command to search across ALL files in my Linux system. (newly installed Linux may need to call updatedb on the first time, to create an initial database).

bash   > locate toco_from

result > /usr/local/intelpython3/bin/toco_from_protos

Bingo, let see if there are more things in this folder

ls /usr/local/intelpython3/bin/toco*

/usr/local/intelpython3/bin/toco
/usr/local/intelpython3/bin/toco_from_protos

My custom python installation requires manual configuration in order to be properly exposed to PATH

So I added the following lines in my ~/.profile

PYTHONBIN=/usr/local/intelpython3/bin
export PATH=$PATH:$PYTHONBIN

and I made it in effect immediately by:

source ~/.profile

Then everything goes well again.

Answer 5

I would recommend to try it on latest tensorflow (not tensorflow-gpu). Worked for me