VR-7870 Support passing a dataframe in the client predict method

[nhatsmrt]

I'm thinking of sth like this, in order to allow the user to define their model in terms of data frame (modulo some wrapper). Need to actually verify that it works 😬 But let me know if sth like this would be problematic, design-wise 🤔

[convoliution]

I think it would be important to maintain backwards compatibility (users who are passing a DataFrame to predict(), but are handling it as a plain array of values). @conradoverta What do you think?

[conradoverta]

I think it would be important to maintain backwards compatibility (users who are passing a DataFrame to predict(), but are handling it as a plain array of values). @conradoverta What do you think?

We need to ship this asap. I'm ok breaking compatibility on this and adding a note.

[nhatsmrt]

Make sure that we provide the conversion back too.

@conradoverta what do you mean by conversion back 🤔

[conradoverta]

Make sure that we provide the conversion back too.

@conradoverta what do you mean by conversion back 🤔

How do I get my dataframe back inside my prediction model?

[nhatsmrt]

Make sure that we provide the conversion back too.

@conradoverta what do you mean by conversion back 🤔

How do I get my dataframe back inside my prediction model?

@conradoverta I think what you meant is:

class MyModel(object):
  @prediction_input_df
  def predict(self, input):
    # input here should be a dataframe

right? It's handled by the prediction_input_df decorator, which, receiving the dictionary, will convert it back to a data frame and feed it to the predict method. This is analogous to other helper decorators that I found in DeployedModel.

[convoliution]

@conradoverta

Make sure that we provide the conversion back too.

@conradoverta what do you mean by conversion back 🤔

How do I get my dataframe back inside my prediction model?

Did Jim's reply address this?

[liuverta]

We're revisiting model I/O very, very soon. And that may come with changes in how the client handles predictions.

And I'm under the impression that users already find these sorts of decorators rather confusing 😬