[
https://issues.apache.org/jira/browse/BEAM-12955?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17436085#comment-17436085
]
Brian Hulette edited comment on BEAM-12955 at 10/29/21, 5:07 PM:
-----------------------------------------------------------------
Ah! ok thank you for the clarification about the TFRecord/tf.train.Example
duality. It looks like Example is a really flexible protobuf (just a list of
key-value pairs). Based on the
[docs|https://www.tensorflow.org/api_docs/python/tf/train/Example] it looks
like it does have structure, but it's defined in some separate configuration
files:
{quote}This flexible and compact format allows the storage of large amounts of
typed data, *but requires that the data shape and use be determined by the
configuration files and parsers that are used to read and write this
format.*{quote}
I don't think that would be compatible with what we're discussing here. The
idea with this task is to generate a schema from a proto message that fully
represents the structure of your data. Something like:
{code}
message Person {
string name = 1;
int64 age = 2;
bool knowsJavascript = 3;
}
{code}
When this is converted to dataframe you'd get one with a name, age, and
knowsJavascript column with the appropriate types. If we did that for the
[Example
proto|https://github.com/tensorflow/tensorflow/blob/919f693420e35d00c8d0a42100837ae3718f7927/tensorflow/core/example/example.proto#L91],
the result wouldn't really match the structure of your data. It would just
have a single column "features" with a list encapsulating all the key-value
pairs.
That being said, I think integrating TF(X) with beam schemas and the dataframe
API would be really powerful and we should look into it. I filed BEAM-13150 for
that, let's move any more discussion over there. I am aware of some work from
the TFX team to use pyarrow with Beam, I can definitely engage them to make
sure we don't duplicate effort.
was (Author: bhulette):
Ah! ok thank you for the clarification about the TFRecord/tf.train.Example
duality. It looks like Example is a really flexible protobuf (just a list of
key-value pairs). Based on the
[docs|https://www.tensorflow.org/api_docs/python/tf/train/Example] it looks
like it does have structure, but it's defined in some separate configuration
files:
{quote}This flexible and compact format allows the storage of large amounts of
typed data, *but requires that the data shape and use be determined by the
configuration files and parsers that are used to read and write this
format.*{quote}
I don't think that would be compatible with what we're discussing here. The
idea with this task is to generate a schema from a proto message that fully
represents the structure of your data. Something like:
{code}
message Person {
string name = 1;
int64 age = 2;
bool knowsJavascript = 3;
}
{code}
When this is converted to dataframe you'd get one with a name, age, and
knowsJavascript column with the appropriate types. If we did that for the
[Example
proto](https://github.com/tensorflow/tensorflow/blob/919f693420e35d00c8d0a42100837ae3718f7927/tensorflow/core/example/example.proto#L91),
the result wouldn't really match the structure of your data. It would just
have a single column "features" with a list encapsulating all the key-value
pairs.
That being said, I think integrating TF(X) with beam schemas and the dataframe
API would be really powerful and we should look into it. I filed BEAM-13150 for
that, let's move any more discussion over there. I am aware of some work from
the TFX team to use pyarrow with Beam, I can definitely engage them to make
sure we don't duplicate effort.
> Add support for inferring Beam Schemas from Python protobuf types
> -----------------------------------------------------------------
>
> Key: BEAM-12955
> URL: https://issues.apache.org/jira/browse/BEAM-12955
> Project: Beam
> Issue Type: Improvement
> Components: sdk-py-core
> Reporter: Brian Hulette
> Assignee: Svetak Vihaan Sundhar
> Priority: P2
> Labels: stale-assigned
>
> Just as we can infer a Beam Schema from a NamedTuple type
> ([code|https://github.com/apache/beam/blob/master/sdks/python/apache_beam/typehints/schemas.py]),
> we should have support for inferring a schema from a [protobuf-generated
> Python
> type|https://developers.google.com/protocol-buffers/docs/pythontutorial].
> This should integrate well with the rest of the schema infrastructure. For
> example it should be possible to use schema-aware transforms like
> [SqlTransform|https://beam.apache.org/releases/pydoc/2.32.0/apache_beam.transforms.sql.html#apache_beam.transforms.sql.SqlTransform],
>
> [Select|https://beam.apache.org/releases/pydoc/2.32.0/apache_beam.transforms.core.html#apache_beam.transforms.core.Select],
> or
> [beam.dataframe.convert.to_dataframe|https://beam.apache.org/releases/pydoc/2.32.0/apache_beam.dataframe.convert.html#apache_beam.dataframe.convert.to_dataframe]
> on a PCollection that is annotated with a protobuf type. For example (using
> the addressbook_pb2 example from the
> [tutorial|https://developers.google.com/protocol-buffers/docs/pythontutorial#reading-a-message]):
> {code:python}
> import adressbook_pb2
> import apache_beam as beam
> from apache_beam.dataframe.convert import to_dataframe
> pc = (input_pc |
> beam.Map(create_person).with_output_type(addressbook_pb2.Person))
> df = to_dataframe(pc) # deferred dataframe with fields id, name, email, ...
> # OR
> pc | beam.transforms.SqlTransform("SELECT name WHERE email = '[email protected]'
> FROM PCOLLECTION")
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
