"Cannot compute the outputs" for dynamic ops (TensorArrayStack)

TensorFlow.js version

1.2.11

Browser version

Google Chrome Version 77.0.3865.120 (Official Build) (64-bit)

Describe the problem or feature request

I have a graph model (built using tf.contrib.seq2seq) which uses control flow ops for dynamic RNN decoding. I exported it with TF 1.15 (using simple_save) and converted with TF.js 1.2.11 (command: tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model --saved_model_tags=serve).

When trying to run the model with executeAsync, I get the following error:

Uncaught (in promise) Error: Cannot compute the outputs [decoder/decode_sample/decoder_1/transpose_1] from the provided inputs [inputs,softmax_temperature]. Consider providing the following inputs: []. Alternatively, to avoid the dynamic ops, use model.execute() and specify the inputs [decoder/decode_sample/decoder_1/TensorArrayStack_1/TensorArrayGatherV3]
    at t.<anonymous> (graph_executor.ts:318)
    at callbacks.ts:253
    at Object.next (callbacks.ts:253)
    at o (callbacks.ts:253)

Does this mean TF.js currently cannot handle these ops? Is it the TensorArrayStack/TensorArrayGatherV3 operation or something else?

[pyu10055]

yes, that is the exact result I got [37, 37, 162, 73, 162, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55] length: 1 proto: Array(0)

[pyu10055]

@cifkao in that case you can rely on our converter to do that, and the fix I had will fix your issue.