[FIXED] Convert TF 2 saved model to frozen graph - no attribute model.inputs[0]

Issue

I created/trained a model via TF 2.4 (w/ CUDA 11.0, Python 3.7) /models/research/object_detection tutorial. No errors, appears to have run fine for 25000 steps. Everything looked normal, Tensorboard showed total loss < 0.5. It produced a saved_model.pb per the tutorial. I now want to convert to a frozen graph for inferences.

It appears to load fine (this code was run in Jupyter notebook):

!ls {model_path} -l
model = tf.compat.v2.saved_model.load(export_dir=model_path)
print (type(model))

output:

total 13232
drwxr-xr-x 2 jay jay     4096 Dec 21 10:41 assets
-rw-r--r-- 1 jay jay 13538598 Dec 21 10:41 saved_model.pb
drwxr-xr-x 2 jay jay     4096 Dec 21 10:41 variables
<class 'tensorflow.python.saved_model.load.Loader._recreate_base_user_object.<locals>._UserObject'>

however, when I begin to convert it, I get an error

full_model = tf.function(lambda x: model(x))
full_model = full_model.get_concrete_function(
    tf.TensorSpec(model.inputs[0].shape, model.inputs[0].dtype))

output:

AttributeError                            Traceback (most recent call last)
<ipython-input-73-50e1947f8357> in <module>
      2 full_model = tf.function(lambda x: model(x))
      3 full_model = full_model.get_concrete_function(
----> 4     tf.TensorSpec(model.inputs[0].shape, model.inputs[0].dtype))

AttributeError: '_UserObject' object has no attribute 'inputs'

in addition, the model cli seems to work:

!saved_model_cli show --dir {model_path} --all

abbreviated output:

2020-12-22 11:38:23.453843: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['__saved_model_init_op']:
  The given SavedModel SignatureDef contains the following input(s):
  The given SavedModel SignatureDef contains the following output(s):
    outputs['__saved_model_init_op'] tensor_info:
        dtype: DT_INVALID
        shape: unknown_rank
        name: NoOp
  Method name is: 

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['input_tensor'] tensor_info:
        dtype: DT_UINT8
        shape: (1, -1, -1, 3)
        name: serving_default_input_tensor:0

<content removed for brevity>

Defined Functions:
  Function Name: '__call__'
    Option #1
      Callable with:
        Argument #1
          input_tensor: TensorSpec(shape=(1, None, None, 3), dtype=tf.uint8, name='input_tensor')

Is my model bad or am I doing something wrong here? Should I be using tf.keras to load the model?

tf.keras.models.load_model(model_path, custom_objects=None, compile=True, options=None)

when I used tf.keras, I received an error on loading:

~/anaconda3/envs/tf24/lib/python3.7/site-packages/tensorflow/python/keras/saving/saved_model/load.py in infer_inputs_from_restored_call_function(fn)
    980     return tensor_spec.TensorSpec(defun.common_shape(x.shape, y.shape),
    981                                   x.dtype, x.name)
--> 982   spec = fn.concrete_functions[0].structured_input_signature[0][0]
    983   for concrete in fn.concrete_functions[1:]:
    984     spec2 = concrete.structured_input_signature[0][0]

IndexError: list index out of range

Solution

You could use Keras to help you get to a frozen graph but (as of 2021.01.04), that doesn't work and you'll encounter issue 43527 as noted.

There is a workaround - not using Keras. Go through the colab tutorials: tensorflow/models/research/object_detection/colab_tutorials/

Specifically, go through: inference_from_saved_model_tf2_colab.ipynb With minor edits, this will run locally - you don't have to run on colab. This works well and it will show you the pattern for utilizing your model without the Keras problems.

Answered By - jduff1075