I want to use tvm as a compiler for a CNN built with Keras 2.1.6 (and tensorflow 1.9.0). But I have this error when importing the keras model to relay with the latest version of TVM:
Traceback (most recent call last):
File "/home/julien/recherche/pose_estimation/ursonet/keras_tvm_example.py", line 39, in <module>
mod, params = relay.frontend.from_keras(keras_resnet50, shape_dict)
File "/home/julien/outils/tvm/python/tvm/relay/frontend/keras.py", line 1130, in from_keras
and not model._node_key(keras_layer, node_idx) in model._network_nodes
AttributeError: 'Model' object has no attribute '_network_nodes'
Which version of TVM should I use with Keras 2.1.6? Knowing that I will not be able to update the model to the latest version of Keras.
thank you for your answer @mbrookhart. But I may not have been clear enough.
For the purpose of my research, I got a Keras (2.1.6) / TensorFlow (1.9.0) code on GitHub. I want to bring this network to TVM. But I don’t want to spend time on bringing the code to Keras 2.3.1 (or more recent version).
So do someone have a table or anything that says: for this particular version of TensorFlow/Keras, you should use that version of TVM.
Looking at the commit history, I’m not sure we had robust tracking for TF/Keras versions back in 2018 I’m not sure I can recommend a version that will definitely support this combination.