tensorflowtensorflow-litetensorflow-hubmovenet

How can I save a model with input shape (1, None, None, 3) with None fixed to 256?

I have questions about a model obtained from TensorFlow Hub.

I am trying to use "movenet" with the following source code.

https://www.kaggle.com/models/google/movenet/tensorFlow2

(1) The Input Shape of this model is (1, None, None, 3), but how can I fix the None part? To create a fixed TensorFlow Lite model, I want to fix the input shape to (1, 256, 256, 3).

(2) Also, I can't display the summary of this model. What should I do?

(3) I want to check the input shape and output shape after tf.saved_model.load, but when I run loaded.signatures['serving_default'], a KeyError occurs. How can I check the input shape and output shape after loading?

    input_size = 256

    model_url = "https://www.kaggle.com/models/google/movenet/TensorFlow2/multipose-lightning/1"
    model = tfhub.load(model_url)
    movenet = model.signatures['serving_default']
    print(movenet.structured_input_signature)
    print(movenet.structured_outputs)
    tf.saved_model.save(model, './saved_model')

    model.summary() # -->> ERROR

    loaded = tf.saved_model.load('./saved_model')
    print(list(loaded.signatures.keys())) # -->> [] # empty!
    movenet = loaded.signatures['serving_default'] # -->> ERROR
    print(movenet.structured_input_signature)
    print(movenet.structured_outputs)

((), {'input': TensorSpec(shape=(1, None, None, 3), dtype=tf.int32, name='input')})
{'output_0': TensorSpec(shape=(1, 6, 56), dtype=tf.float32, name='output_0')}

### model.summary() # -->> ERROR
'_UserObject' object has no attribute 'summary'
  File "/home/debian/sandbox/movenet/python/movenet_multi.py", line 126, in main
    model.summary() # -->> ERROR
  File "/home/debian/sandbox/movenet/python/movenet_multi.py", line 356, in <module>
    main()
AttributeError: '_UserObject' object has no attribute 'summary'

### movenet = loaded.signatures['serving_default'] # -->> ERROR
'serving_default'
  File "/home/debian/sandbox/movenet/python/movenet_multi.py", line 127, in main
    movenet = loaded.signatures['serving_default'] # -->> ERROR
  File "/home/debian/sandbox/movenet/python/movenet_multi.py", line 355, in <module>
    main()
KeyError: 'serving_default'
Solution

  • I solved it myself.

    (1) I was able to solve the issue by converting the model using the flow below.

    (2) It is not a Keras model, so it seems that the internals cannot be seen with summary(). I was able to solve it by exporting it to TensorFlow Lite format and visualizing it using Netron.

    (3) Just like issue (2), you can check it using Netron.

    Download a model from TensorFlow Hub and save it in Keras and TensorFlow Lite formats.

    ### Save as tflite format
# -->> for movenet
import tensorflow as tf
import tensorflow_hub as hub

model = tf.keras.Sequential(
    [
        hub.KerasLayer(
            "https://www.kaggle.com/models/google/movenet/TensorFlow2/multipose-lightning/1",
            trainable=False,
            signature="serving_default", 
            signature_outputs_as_dict=True,
        ),
    ]
)
model.build([1, 256, 256, 3])
model.summary()
model.save("movenet-multipose-lightning")

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

    Convert TensorFlow Lite to ONNX, fix input shape and convert back to TensorFlow Lite

    ### Convert to ONNX format
$ python -m tf2onnx.convert --tflite converted_model.tflite --output converted_model_nchw_output.onnx --inputs-as-nchw serving_default_keras_layer_input:0

### Fixing input shape
$ python -m onnxruntime.tools.make_dynamic_shape_fixed --input_name serving_default_keras_layer_input:0 --input_shape 1,3,256,256 converted_model_nchw_o
utput.onnx converted_model_nchw_output_fixed.onnx

### Remove suffix
$ python script.py converted_model_nchw_output_fixed.onnx converted_model_nchw_output_mod.onnx

### Converting back to TensorFlow format
$ onnx2tf -i converted_model_nchw_output_mod.onnx -osd

### Convert back to TensorFlow Lite format
$ python tf2tflite_no_optim.py -m saved_model -o movenet_multi_nchw_output_fixed.tflite

    script.py

    import onnx
import sys

def remove_suffix_from_names(model_path, output_model_path, suffix=':0'):
    onnx_model = onnx.load(model_path)
    graph_input_names = [input.name for input in onnx_model.graph.input]
    graph_output_names = [output.name for output in onnx_model.graph.output]
    print('graph_input_names =', graph_input_names)
    print('graph_output_names =', graph_output_names)

    for input in onnx_model.graph.input:
        input.name = input.name.removesuffix(suffix)

    for output in onnx_model.graph.output:
        output.name = output.name.removesuffix(suffix)

    for node in onnx_model.graph.node:
        for i in range(len(node.input)):
            if node.input[i] in graph_input_names:
                node.input[i] = node.input[i].removesuffix(suffix)

        for i in range(len(node.output)):
            if node.output[i] in graph_output_names:
                node.output[i] = node.output[i].removesuffix(suffix)

    onnx.save(onnx_model, output_model_path)

if __name__ == "__main__":
    if len(sys.argv) != 3:
        print("Usage: python3 script.py <input_model.onnx> <output_model.onnx>")
        sys.exit(1)

    input_model_path = sys.argv[1]
    output_model_path = sys.argv[2]

    remove_suffix_from_names(input_model_path, output_model_path)

    tf2tflite_no_optim.py

    import tensorflow as tf
import numpy as np
import argparse
import os

def representative_dataset():
    for _ in range(100): 
        data = tf.random.uniform((1, 544, 960, 3), minval=-1, maxval=1, dtype=tf.float32)  
        yield [data]

def convert_model(model_path, output_model_path):
    converter = tf.lite.TFLiteConverter.from_saved_model(model_path)
    tflite_model = converter.convert()

    with open(output_model_path, "wb") as f:
        f.write(tflite_model)
    print(f"Successfully converted and saved to {output_model_path}")

def main():
    parser = argparse.ArgumentParser(description="Convert a TensorFlow SavedModel to a quantized TFLite model.")
    parser.add_argument("-m", "--model_path", type=str, required=True, help="Path to the input TensorFlow SavedModel directory.")
    parser.add_argument("-o", "--output_path", type=str, help="Path to save the converted TFLite model.  If not specified, defaults to the same directory as the input model.")

    args = parser.parse_args()

    if args.output_path is None:
        base_name = os.path.basename(os.path.normpath(args.model_path))
        output_model_path = os.path.join(os.path.dirname(args.model_path), f"{base_name}_int8.tflite")
    else:
        output_model_path = args.output_path


    convert_model(args.model_path, output_model_path)

if __name__ == "__main__":
    main()

    requirements.txt

    absl-py==2.2.1
astunparse==1.6.3
cachetools==5.5.2
certifi==2025.1.31
charset-normalizer==3.4.1
coloredlogs==15.0.1
flatbuffers==1.12
gast==0.4.0
google-auth==2.38.0
google-auth-oauthlib==0.4.6
google-pasta==0.2.0
grpcio==1.71.0
h5py==3.13.0
humanfriendly==10.0
idna==3.10
keras==2.9.0
Keras-Preprocessing==1.1.2
libclang==18.1.1
Markdown==3.7
MarkupSafe==3.0.2
mpmath==1.3.0
numpy==1.26.4
oauthlib==3.2.2
onnx==1.14.1
onnx-graphsurgeon==0.5.7
onnx2tf==1.26.9
onnxruntime==1.21.0
opt_einsum==3.4.0
packaging==24.2
protobuf==3.20.3
psutil==7.0.0
pyasn1==0.6.1
pyasn1_modules==0.4.2
requests==2.32.3
requests-oauthlib==2.0.0
rsa==4.9
six==1.17.0
sng4onnx==1.0.4
sympy==1.13.3
tensorboard==2.9.0
tensorboard-data-server==0.6.1
tensorboard-plugin-wit==1.8.1
tensorflow==2.9.0
tensorflow-estimator==2.9.0
tensorflow-hub==0.16.1
tensorflow-io-gcs-filesystem==0.37.1
tensorflow-neuron==1.0
termcolor==3.0.0
tf-keras==2.14.1
tf2onnx==1.13.0
typing_extensions==4.13.0
urllib3==2.3.0
Werkzeug==3.1.3
wrapt==1.17.2