Description

Just a simple script to inspect TF Lite metadata

Usage

simple use:

python metadata_viewer.py \
    model_file=./some_tf_lite_model.tflite

to visualize metadata

python metadata_viewer.py \
    model_file=./some_tf_lite_model.tflite \
    appended_resource_id=0

where appended_resource_id should be the index (0 based) of the appended resource file

Helper function

If you just want to get the label names that are embedded within the .tflite file (i.e like for edge models trained in Google Cloud AutoML) then you can use this helper function:

import zipfile
from tflite_support import metadata

def get_labels_from_tflite(model_file_name, format=0):
  '''
  Unpack the metadata from a tflite model and return and parse the first 
  associated file which is assumed to be the label file.
  input : str (model file name), int (format 0=list, 1=dict)
  output : str (text from labels file)
  '''

  displayer = metadata.MetadataDisplayer.with_model_file(model_file_name)
  associate_files= displayer.get_packed_associated_file_list()
  resource_file = associate_files[0]
  archive = zipfile.ZipFile(model_file_name, 'r')
  labels = archive.read(resource_file).decode().split('\n')

  if format == 1:
    labels = {i:label  for (i, label) in enumerate(labels)}

  return labels
  
  
# Example usage
# get labels as  list >> ["cat", "dog", "mouse"]
labels_list = get_labels_from_tflite('my_model.tflite', format=0)

# get labels as  dict >> {0:"cat", 1:"dog", 2:"mouse"}
labels_dict = get_labels_from_tflite('my_model.tflite', format=1)
 

Resources

Oficial documentation for TF Lite metadata