Question Strange input shape after export from keras to ONNX and import in unity

Hello,

this is my neural network model:

The input is an example of 10000 features. Each feature is a number (0 or 1).
The output is a number between 0 and 1.

Code (CSharp):

1. from tensorflow.keras.datasets import imdb
2. (train_data, train_labels), (test_data, test_labels) = imdb.load_data(
3. num_words=10000)
4.  
5. import numpy as np
6. def vectorize_sequences(sequences, dimension=10000):
7.   results = np.zeros((len(sequences), dimension))
8.   for i, sequence in enumerate(sequences):
9.     for j in sequence:
10.       results[i, j] = 1.
11.   return results
12.  
13. x_train = vectorize_sequences(train_data)
14. x_test = vectorize_sequences(test_data)
15.  
16. y_train = np.asarray(train_labels).astype("float32")
17. y_test = np.asarray(test_labels).astype("float32")
18.  
19. from tensorflow import keras
20. from tensorflow.keras import layers
21.  
22. model = keras.Sequential([
23. layers.Dense(16, activation="relu"),
24. layers.Dense(16, activation="relu"),
25. layers.Dense(1, activation="sigmoid")
26. ])
27.  
28. model.compile(optimizer="rmsprop",
29. loss="binary_crossentropy",
30. metrics=["accuracy"])
31.  
32. x_val = x_train[:10000]
33. partial_x_train = x_train[10000:]
34. y_val = y_train[:10000]
35. partial_y_train = y_train[10000:]
36.  
37. model.fit(partial_x_train,
38. partial_y_train,
39. epochs=20,
40. batch_size=512,
41. validation_data=(x_val, y_val))

This is how I export the model:

Code (CSharp):

1. import tensorflow as tf
2.  
3. spec = (tf.TensorSpec(model.inputs[0].shape, tf.float32, name="my input"),)
4.  
5. nchw_inputs_list = [model.inputs[0].name]
6.  
7. import tf2onnx
8.  
9. model_proto, _ = tf2onnx.convert.from_keras(model, input_signature=spec, custom_ops=None, opset=9, inputs_as_nchw=nchw_inputs_list, output_path="example.onnx")

When I import it in Unity:



What I am doing wrong? How can I export it to ONNX and import it in unity in the right way?

Thank you

 

As far as I can tell from my experiments, Barracuda has not fully implemented all functions so it will only work with some onnx if you're lucky.