first try

Author: xadupre

Troubleshooting.md

@@ -36,3 +36,7 @@ An example of this is the [ONNX Slice operator before opset-10](https://github.c
 You can pass the options ```--fold_const```(removed after tf2onnx-1.9.3) in the tf2onnx command line that allows tf2onnx to apply more aggressive constant folding which will increase chances to find a constant.
 
 If this doesn't work the model is most likely not to be able to convert to ONNX. We used to see this a lot of issue with the ONNX Slice op and in opset-10 was updated for exactly this reason.
+
+## cudaSetDevice() on GPU:0 failed. Status: CUDA-capable device(s) is/are busy or unavailable
+
+See [Regression: TF 2.18 crashes with cudaSetDevice failing due to GPU being busy](https://github.com/tensorflow/tensorflow/issues/78784).

examples/end2end_tfkeras.py

@@ -8,11 +8,9 @@
 *onnxruntime*, *tensorflow* and *tensorflow.lite*.
 """
 from onnxruntime import InferenceSession
-import os
 import subprocess
 import timeit
 import numpy as np
-import tensorflow as tf
 from tensorflow import keras
 from tensorflow.keras import layers, Input
 

examples/tf_custom_op/double_and_add_one_custom_op.py

@@ -7,7 +7,6 @@
 import os
 from tf2onnx import utils
 from tf2onnx.handler import tf_op
-from tf2onnx.tf_loader import tf_placeholder
 
 
 DIR_PATH = os.path.realpath(os.path.dirname(__file__))

tests/keras2onnx_unit_tests/test_subclassing.py

@@ -49,6 +49,43 @@ def call(self, inputs, **kwargs):
         return output
 
 
+def get_save_spec(model, dynamic_batch=False):
+  """Returns the save spec of the subclassing keras model."""
+  from tensorflow.python.framework import tensor_spec
+  shapes_dict = getattr(model, '_build_shapes_dict', None)
+  # TODO: restore dynamic_batch
+  # assert not dynamic_batch, f"get_save_spec: dynamic_batch={dynamic_batch}, shapes_dict={shapes_dict}"
+  if not shapes_dict:
+    return None
+
+  if 'input_shape' not in shapes_dict:
+    raise ValueError(
+        'Model {} cannot be saved because the input shapes have not been set.'
+    )
+
+  input_shape = shapes_dict['input_shape']
+  if isinstance(input_shape, tuple):
+    shape = input_shape
+    shape = (None,) + shape[1:]
+    return tensor_spec.TensorSpec(
+        shape=shape, dtype=model.input_dtype
+    )
+  elif isinstance(input_shape, dict):
+    specs = {}
+    for key, shape in input_shape.items():
+      shape = (None,) + shape[1:]
+      specs[key] = tensor_spec.TensorSpec(
+          shape=shape, dtype=model.input_dtype, name=key
+      )
+    return specs
+  elif isinstance(input_shape, list):
+    specs = []
+    for shape in input_shape:
+      shape = (None,) + shape[1:]
+      specs.append(tensor_spec.TensorSpec(shape=shape, dtype=model.input_dtype))
+    return specs
+
+
 class SimpleWrapperModel(tf.keras.Model):
     def __init__(self, func):
         super(SimpleWrapperModel, self).__init__()
@@ -57,6 +94,9 @@ def __init__(self, func):
     def call(self, inputs, **kwargs):
         return self.func(inputs)
 
+    def _get_save_spec(self, dynamic_batch=False):
+        return get_save_spec(self, dynamic_batch=dynamic_batch)
+
 
 def test_lenet(runner):
     tf.keras.backend.clear_session()
@@ -198,7 +238,10 @@ def _tf_where(input_0):
     swm = SimpleWrapperModel(_tf_where)
     const_in = [np.array([2, 4, 6, 8, 10]).astype(np.int32)]
     expected = swm(const_in)
-    swm._set_inputs(const_in)
+    if hasattr(swm, "_set_input"):
+        swm._set_inputs(const_in)
+    else:
+        swm.inputs_spec = const_in
     oxml = convert_keras(swm)
     assert runner('where_test', oxml, const_in, expected)
 

tf2onnx/convert.py

@@ -447,7 +447,7 @@ def from_keras(model, input_signature=None, opset=None, custom_ops=None, custom_
     function = _saving_utils.trace_model_call(model, input_signature)
     try:
         concrete_func = function.get_concrete_function()
-    except TypeError as e:
+    except (TypeError, AttributeError) as e:
         # Legacy keras models don't accept the training arg tf provides so we hack around it
         if "got an unexpected keyword argument 'training'" not in str(e):
             raise e