Description
I have a simple Keras model (see below) with a GRU which keras2onnx was mapping just fine to onnx GRU.
Now with tf2onnx.convert.from_keras I get a huge very complicated onnx model with loops and initializers other stuff and does not map to the onnx GRU. This breaks several inference backends.
Urgency
Urgency is high - we have many customer models with GRU/LSTM/RNN that need to move to latest release now that keras2onnx is deprecated.
System information
- OS Platform and Distribution: Linux Ubuntu 18.04
- Tensorflow Version: 2.3.1
- Python version: 3.6
To Reproduce
In python, build the model below and run onnx_model,_ = tf2onnx.convert.from_keras(model)
from tensorflow.keras.models import Model, load_model
from tensorflow.keras.layers import Input, Dense, GRU, Dropout, Activation # dense-gru
model_in = Input(tuple(in_shape),batch_size=batch_size)
x = Dense(192, activation='relu')(model_in)
x = Dropout(0.5)(x)
x = GRU(32, return_sequences=True)(x)
x = Dropout(0.5)(x)
model_out = Dense(outlen, activation='softmax')(x)
model = Model(inputs=model_in, outputs=model_out)
Note: this model trains and works fine with keras2onnx v1.7 resulting in an .onnx model with a similar structure as the Keras definition (dropouts removed):
model.summary()
Model: "functional_1"
Layer (type) Output Shape Param #
input_1 (InputLayer) [(32, 16, 192)] 0
dense (Dense) (32, 16, 192) 37056
dropout (Dropout) (32, 16, 192) 0
gru (GRU) (32, 16, 32) 21696
dropout_1 (Dropout) (32, 16, 32) 0
dense_1 (Dense) (32, 16, 4) 132
Total params: 58,884
Trainable params: 58,884
Non-trainable params: 0