Could tft support embedding transform function? about transform HOT 11 CLOSED

some function like tf.keras.layers.Embedding

from transform.

You could call either tf.nn.embedding_lookup directly or tf.keras.layers.Embedding in the preprocessing_fn.
Have you ran into issues doing that?

from transform.

@zhangjunqiang

Apologies for the delay and as of now we don't have such function for embedding transform and you can check all available functions with Tensorflow Transform here and I have attached one good resource for you with tensorflow transform examples, please have a look once here and if you're looking to explore how Transform component works with TFX pipeline then please check this article I hope it will help you to resolve your issue, If your issue still persists or you need any further assistance please let us know ?

Thank you!

from transform.

You could call either tf.nn.embedding_lookup directly or tf.keras.layers.Embedding in the preprocessing_fn. Have you ran into issues doing that?

because tf.keras.layers.Embedding is return a layer object not a tensor. All of the tfx function in the preprocessing_fn must return a tensor.
if i use tf.keras.layers.Embedding, i must build a tf model, and then model.compile() 、 model.fit() and the predict the input data(It's a whole train flow). but i don't want a whole train flow in the preprocessing_fn code. i want use the embedding function like tft.scale_to_0_1(). is this possible？

from transform.

thanks for the reply. if i can't use the embedding function in the preprocessing_fn function. i will put the embedding layer in the train stage. not the data preprocess stage

from transform.

Could you not call it and apply it directly to your data?
tf.keras.layers.Embedding(dim_a, dim_b)(inputs['x'])

from transform.

Could you not call it and apply it directly to your data? tf.keras.layers.Embedding(dim_a, dim_b)(inputs['x'])

I had try. but return an error.

return self._stateless_fn._get_concrete_function_garbage_collected(  # pylint: disable=protected-access
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2576, in _get_concrete_function_garbage_collected
    graph_function, _ = self._maybe_define_function(args, kwargs)
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2760, in _maybe_define_function
    graph_function = self._create_graph_function(args, kwargs)
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2670, in _create_graph_function
    func_graph_module.func_graph_from_py_func(
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py", line 1247, in func_graph_from_py_func
    func_outputs = python_func(*func_args, **func_kwargs)
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 677, in wrapped_fn
    out = weak_wrapped_fn().__wrapped__(*args, **kwds)
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow_transform/impl_helper.py", line 683, in transform_fn
    transformed_features = preprocessing_fn(inputs_copy)
  File "tfx_transform.py", line 133, in preprocessing_fn
    outputs['embedding_age'] = tf.keras.layers.Embedding(input_dim=100, output_dim=10)(inputs["age"])
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/keras/utils/traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/opt/conda/envs/feast-python-sdk/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 790, in invalid_creator_scope
    raise ValueError(
ValueError: tf.function only supports singleton tf.Variables created on the first call. Make sure the tf.Variable is only created once or created outside tf.function. See https://www.tensorflow.org/guide/function#creating_tfvariables for more information.

from transform.

What about using tf.nn.embedding_lookup directly? (since keras is creating trainable variables on that layer's build implementation)

from transform.

What about using tf.nn.embedding_lookup directly? (since keras is creating trainable variables on that layer's build implementation)
tf.nn.embedding_lookup The first param of this function is params(A single tensor representing the complete embedding tensor,or a list of tensors all of same shape except for the first dimension, representing sharded embedding tensors following "div" partition strategy.)
I did't have this embedding tensor. This embedding tensor should train by embedding function.

from transform.

In that case, the embedding training shouldn't happen in transform.
If there are pre-trained embeddings those can be applied during transform.

from transform.

In that case, the embedding training shouldn't happen in transform. If there are pre-trained embeddings those can be applied during transform.

Get it, i should split embedding transform from pre-trained stage into training stage.
But in other words, embedding transform is one type of feature data transform.

from transform.