TFLite New Converter with 1.X frozen graph

[mhs4670go]

I tried to covert 1.X frozen graph with TFLite New Converter following this guidelines(https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt).

Here is my full python script.

graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(open(flags.input_path, 'rb').read())

wrap_func = wrap_frozen_graph(
    graph_def,
    inputs=[_str + ":0" for _str in _parse_array(flags.input_arrays)],
    outputs=[_str + ":0" for _str in _parse_array(flags.output_arrays)])
converter = tf.lite.TFLiteConverter.from_concrete_functions([wrap_func])

It works well. But, is there any method to set input shapes from users? It's very hard to find this:(

like tf.compat.v1.lite.TFLiteConverter.from_frozen_graph method in 1.X version.

def from_frozen_graph(cls,
                        graph_def_file,
                        input_arrays,
                        output_arrays,
                        input_shapes=None)

I found something similar like

# Set the correct data type and shape; shape can be (None, 224, 224, 3) also
new_placeholder = tf.placeholder(tf.float32, shape=(1, 224, 224, 3), name='inputs_new_name') 
# here you need to state the name of the placeholder you used in your original input placeholder  

saver = tf.import_graph_def(path/to/.meta, input_map={"original_inputs_placeholder_name:0": new_placeholder})

But, this is very inconvenient to use.

I'm looking for similar features in 2.X converter, but I haven't found any:(

[jvishnuvardhan]

@mhs4670go Yes. there is an option to change size during inference. Please check this comment. Please post a standalone code to reproduce your issue. Thanks!

[mhs4670go]

@jvishnuvardhan Thank you for your comment. Changing size during inference might be enough, But, actually, I want to know if I can convert .pb file with None shape to .tflite file with user specified shape not None shape like v1 converter as I said above.

# This is v1 converter having `input_shape` option
input_shapes = None
  if flags.input_shapes: # By this line, `pb` with None shape can be converted to `.tflite` with known shape. 
                                     # Does v2 converter have similar method or sth? 
    input_shapes_list = [
        _parse_array(shape, type_fn=int)
        for shape in six.ensure_str(flags.input_shapes).split(":")
    ]
    input_shapes = dict(list(zip(input_arrays, input_shapes_list)))

[MeghnaNatraj]

The link you are referring to discusses how a frozen graph (in TensorFlow 1 model format) can be converted to a TensorFlow 2 model: https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt. This is not required.

To convert the frozen graph model to a TensorFlow Lite model:

1. If you are using TensorFlow 1, the API is tf.lite.TFLiteConverter.from_frozen_graph
2. If you are using TensorFlow 2, the API is tf.compat.v1.lite.TFLiteConverter.from_frozen_graph

# TensorFlow Version 2.2
import tensorflow as tf
converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(
    graph_def_file='....path/to/frozen_graph.pb', 
    input_arrays=...,
    output_arrays=....,
    input_shapes={'...' : [_, _,....]}
)
tflite_model = converter.convert()

tflite_model_size = open('model.tflite', 'wb').write(tflite_model)
print('TFLite Model is %d bytes' % tflite_model_size)

The .from_frozen_graph API can be defined this way and the attributes which can be added are here. To find the names of these arrays, visualize the .pb file in Netron