I am trying the SavedModel format export method

[famy040326]

I am trying the SavedModel format export method described in the link below.
However, I get an error and cannot execute it.
I would be grateful if you could help me with a solution.

https://mindee.github.io/doctr/using_doctr/using_model_export.html#using-savedmodel

import tensorflow as from doctr.models import db_resnet50 model = db_resnet50(pretrained=True) input_t = tf.random.uniform(shape=[1, 1024, 1024, 3], maxval=1, dtype=tf.float32) _ = model(input_t, training=False) tf.saved_model.save(model, './doctr_SavedModel/')

The following is the content of the error.
`Traceback (most recent call last):
File "", line 1, in
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\saved_model\save.py", line 1334, in save
save_and_return_nodes(obj, export_dir, signatures, options)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\saved_model\save.py", line 1369, in save_and_return_nodes
_build_meta_graph(obj, signatures, options, meta_graph_def))
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\saved_model\save.py", line 1536, in _build_meta_graph
return _build_meta_graph_impl(obj, signatures, options, meta_graph_def)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\saved_model\save.py", line 1477, in _build_meta_graph_impl
signatures = signature_serialization.find_function_to_export(
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\saved_model\signature_serialization.py", line 103, in find_function_to_export
for name, child in children:
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\saved_model\save.py", line 133, in list_children
super(_AugmentedGraphView, self).list_children(
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\training\tracking\graph_view.py", line 256, in list_children
in obj._trackable_children(save_type, **kwargs).items()]
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\training\tracking\base.py", line 1479, in _trackable_children
return self._get_legacy_saved_model_children(cache)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\training\tracking\base.py", line 1490, in _get_legacy_saved_model_children
functions = self._list_functions_for_serialization(serialization_cache)
File "C:\Python_dev\env1\lib\site-packages\keras\engine\training.py", line 3081, in _list_functions_for_serialization
functions = super(
File "C:\Python_dev\env1\lib\site-packages\keras\engine\base_layer.py", line 3168, in _list_functions_for_serialization
return (self._trackable_saved_model_saver
File "C:\Python_dev\env1\lib\site-packages\keras\saving\saved_model\base_serialization.py", line 93, in list_functions_for_serialization
fns = self.functions_to_serialize(serialization_cache)
File "C:\Python_dev\env1\lib\site-packages\keras\saving\saved_model\layer_serialization.py", line 72, in functions_to_serialize
return (self._get_serialized_attributes(
File "C:\Python_dev\env1\lib\site-packages\keras\saving\saved_model\layer_serialization.py", line 88, in _get_serialized_attributes
object_dict, function_dict = self._get_serialized_attributes_internal(
File "C:\Python_dev\env1\lib\site-packages\keras\saving\saved_model\model_serialization.py", line 51, in _get_serialized_attributes_internal
default_signature = save_impl.default_save_signature(self.obj)
File "C:\Python_dev\env1\lib\site-packages\keras\saving\saved_model\save_impl.py", line 208, in default_save_signature
fn = saving_utils.trace_model_call(layer)
File "C:\Python_dev\env1\lib\site-packages\keras\saving\saving_utils.py", line 148, in trace_model_call
return _wrapped_model.get_concrete_function(*model_args, **model_kwargs)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\def_function.py", line 1264, in get_concrete_function
concrete = self._get_concrete_function_garbage_collected(*args, **kwargs)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\def_function.py", line 1244, in _get_concrete_function_garbage_collected
self._initialize(args, kwargs, add_initializers_to=initializers)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\def_function.py", line 785, in _initialize
self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\function.py", line 2983, in _get_concrete_function_internal_garbage_collected
graph_function, _ = self._maybe_define_function(args, kwargs)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\function.py", line 3292, in _maybe_define_function
graph_function = self._create_graph_function(args, kwargs)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\function.py", line 3130, in _create_graph_function
func_graph_module.func_graph_from_py_func(
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\framework\func_graph.py", line 1161, in func_graph_from_py_func
func_outputs = python_func(*func_args, **func_kwargs)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\eager\def_function.py", line 677, in wrapped_fn
out = weak_wrapped_fn().wrapped(*args, **kwds)
File "C:\Python_dev\env1\lib\site-packages\tensorflow\python\framework\func_graph.py", line 1147, in autograph_handler
raise e.ag_error_metadata.to_exception(e)
AttributeError: in user code:

File "C:\Python_dev\env1\lib\site-packages\keras\saving\saving_utils.py", line 138, in _wrapped_model  *
    outputs = model(*args, **kwargs)
File "C:\Python_dev\env1\lib\site-packages\keras\utils\traceback_utils.py", line 67, in error_handler  **
    raise e.with_traceback(filtered_tb) from None

AttributeError: Exception encountered when calling layer "db_net" (type DBNet).

in user code:

    File "C:\Python_dev\env1\doctr\doctr\models\detection\differentiable_binarization\tensorflow.py", line 242, in call  *
        out["preds"] = [preds[0] for preds in self.postprocessor(prob_map.numpy())]

    AttributeError: 'Tensor' object has no attribute 'numpy'


Call arguments received:
  • x=tf.Tensor(shape=(None, 1024, 1024, 3), dtype=float32)
  • target=None
  • return_model_output=False
  • return_preds=False
  • kwargs={'training': 'False'}`

[felixdittrich92]

Hi @famy040326 👋,
you are right we have on track to update this part of the docs in 0.7.0 . (will be changed to onnx)
So in short the saved model export cannot work because the model includes the postprocessing (numpy / opencv /shapely) which cannot be stored in the saved format.

A workaround would be to follow the way we used in tests for onnx export:

doctr/tests/tensorflow/test_models_detection_tf.py

Line 175 in ef4773e

def test_models_onnx_export(arch_name, input_shape, output_size):

So the key is to set the arg exportable to True which skips the postprocessing and returns the raw logits (without applying sigmoid activation), then the model should be saved model exportable.
NOTE: after loading the model you have to perform the sigmoid activation + postprocessing on your own :)

https://github.com/mindee/doctr/blob/main/doctr/models/detection/differentiable_binarization/base.py

prob_map = tf.math.sigmoid(logits)
out["preds"] = [preds[0] for preds in self.postprocessor(prob_map.numpy())]

[famy040326]

Thank you for your response.
I understood what you are saying.

I am not sure how to proceed and would like your advice.
We are looking to train docTR models and leverage them in TensorFlow.js.
At this point, we have utilized the sample scripts to train on our own data. And we have checked the output of checkpoint,weights.data-00000-of-00001,weights.index.
python references/recognition/train_tensorflow.py crnn_vgg16_bn --train_path path/to/your/train_set --val_path path/to/your/val_set --epochs 5

I don't know what more steps I need to take to be able to utilize this in TensorFlow.js.
We would like to utilize only the recognition model.
I would be grateful if you could tell me.

[felixdittrich92]

Hi @famy040326 👋,
have you checked out the TF.js demo : https://github.com/mindee/doctr-tfjs-demo ? :)

same case as described for the detection model use exportable=True the model can be exported to Onnx or Saved Model format and will return the logits vector .. post processing is done here:
https://github.com/mindee/doctr-tfjs-demo/blob/1a15e35bfe033612d2d9de8b247472175f512b9d/src/utils/index.ts#L190

I have not worked on this app so maybe @charlesmindee @frgfm can provide a bit more details / tips 😅

[famy040326]

Thank you for your kind attention.
I will check the TF.js demo and try some things.