Hello!
System information (version)
- OpenVINO => 2022.2
- Operating System / Platform => Ubuntu 20.04.5 LTS
- Compiler =>
- Problem classification: Model Conversion
- Framework: TensorFlow
There is an issue, when converting a simple TF network with GRU layer which was trained on TensorFlow 2.8 and saved into SavedModel format, there is no problem. I can convert my model easily without any error, with this code:
mo --saved_model_dir /work/user/saved_model/my_model/ --input_shape [?,28,28] --output_dir /work/user/saved_model/vino
But when I change Tensorflow version to 2.9.1, and I train a TF network with GRU layer and save into SavedModel format, I get an error like this:
[ ERROR ] Cannot infer shapes or values for node "StatefulPartitionedCall/sequential/gru/PartitionedCall/TensorArrayV2_1".
[ ERROR ] Graph execution error:
Tensorflow type 21 not convertible to numpy dtype.
[ ERROR ]
[ ERROR ] It can happen due to bug in custom shape infer function <function tf_native_tf_node_infer at 0x7f6bf1b1c700>.
[ ERROR ] Or because the node inputs have incorrect values/shapes.
[ ERROR ] Or because input shapes are incorrect (embedded to the model or passed via --input_shape).
[ ERROR ] Run Model Optimizer with --log_level=DEBUG for more information.
[ ERROR ] Exception occurred during running replacer "REPLACEMENT_ID" (<class 'openvino.tools.mo.middle.PartialInfer.PartialInfer'>): Stopped shape/value propagation at "StatefulPartitionedCall/sequential/gru/PartitionedCall/TensorArrayV2_1" node.
For more information please refer to Model Optimizer FAQ, question #38. (https://docs.openvino.ai/latest/openvino_docs_MO_DG_prepare_model_Model_Optimizer_FAQ.html?question=38#question-38)
A short code, to reproduce this problem:
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
model = keras.Sequential()
model.add(layers.GRU(64, input_shape=(28, 28)))
model.add(layers.BatchNormalization())
model.add(layers.Dense(10))
print(model.summary())
mnist = keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train/255.0, x_test/255.0
x_validate, y_validate = x_test[:-10], y_test[:-10]
x_test, y_test = x_test[-10:], y_test[-10:]
model.compile(
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer="sgd",
metrics=["accuracy"],
)
model.fit(
x_train, y_train, validation_data=(x_validate, y_validate), batch_size=64, epochs=5
)
model.save('saved_model/my_model')
Then converting with this code:
mo --saved_model_dir /work/user/saved_model/my_model/ --input_shape [?,28,28] --output_dir /work/user/saved_model/vino
It works well with TF 2.8, but after the change TF to 2.9.1, I get an error, like above.
