Supported and tested RNN classes and APIs: LSTMCell, BasicLSTMCell, GRUCell, GRUBlockCell, MultiRNNCell, and user defined RNN cells inheriting from rnn_cell_impl.RNNCell, used along with DropoutWrapper, BahdanauAttention, AttentionWrapper. Check here for well tested cases.
For other advanced RNN cells, it is supposed to good to convert as well, but there is no comprehensive testing against them.
tf.nn.dynamic_rnn and tf.nn.bidirectional_dynamic_rnn are common APIs to trigger RNN cell's run, both approaches are supported to convert.
Use following commands to have a quick trial on your model:
python -m tf2onnx.convert --input frozen_rnn_model.pb --inputs input1:0,input2:0 --outputs output1:0,output2:0 --opset 8 --output target.onnx --continue_on_error
Besides BasicLSTMCell/LSTMCell/GRUCell/GRUBlockCell conversion, all other conversion requires target onnx opset to be >= 8.
There are a few tests case of LSTMCell, BasicLSTMCell, GRUCell, GRUBlockCell for your reference.
For other advanced RNN cells, check here.
Use onnxruntime or caffe2 to test against converted models.
There is a simpler way to run your models and test its correctness (compared with TensorFlow run) using following command.
python tests\run_pretrained_models.py --backend onnxruntime --config rnn.yaml --tests model_name --onnx-file ".\tmp" --opset 8
The content of rnn.yaml looks as below. For inputs, an explicit numpy expression or a shape can be used. If a shape is specified, the value will be randomly generated.
model_name:
model: path/to/tf_frozen.pb
input_get: get_ramp
inputs:
"input1:0": np.array([60]) # numpy random function
"input2:0": [2, 1, 300] # shape for the input
outputs:
- output1:0
- output2:0