He Jia comments

Results 87 comments of


                                            He Jia

ShadowVariable Post applying gradients gives NaN value for embedding

Sorry, PS mode is only supported for TF1-like estimator building now.

Support for ragged arrays, like torch.nested

Any progress?

Performance comparison with the Pytorch ecosystem

> Hi @monatis, and apologies for the late reply! > > Thanks for the kind words and for checking out the repo. I haven’t put together a formal benchmark report...

fail to run demo:movielens-1m-keras-with-horovod

Try xla jit level 1 for now. TFRA with XLA support will be soon available.

How to save and restore a model with dynamic embeddings?

Sorry, TFRA is hard to support tf.keras.models.load_model API. Because load_model will create trainable variable object from TensorFlow, but TFRA trainable wrapper is not in TF code.

How to save and restore a model with dynamic embeddings?

@PhyCoe load_weight API https://github.com/tensorflow/recommenders-addons/blob/b3bc3d4691a9421dbe837ee9708c504217fdc8f9/demo/dynamic_embedding/movielens-1m-keras/movielens-1m-keras.py#L133

item embedding and user sequence item embedding can not use the same embedding variable

TFRA Keras Embedding layers could not be called twice and more for now. Please concat all inputs in one tensor and then lookup. If you use GPU kernel, this would...

`jnp.unique` along an axis is very slow for large arrays

@jakevdp Any progress? Is migrate the unique operator from TensorFlow as an independent primitive a good option? like TopK operator.

`jnp.unique` along an axis is very slow for large arrays

@jakevdp So would it be better to abandon the XLA variadic sort and change it to a XLA custom call kernel?

`jnp.unique` along an axis is very slow for large arrays

What about directly migrate hard codes from TensorFlow op kernel? Since the most common solution of Unique OP which is Radix sort and it’s hard to be expressed by XLA....