Wrapper class for end-to-end t5 model

[pmabbo13]

Description

Add a wrapper class for the T5 model that takes text as input and outputs the generated text from the decoder

Process

The wrapper class performs the following end-to-end process:

1. transforms input text using T5Transform to get truncated and padded token IDs
2. passes resulting model input through t5 model, using a simple beam search generator to produce the output sequence
3. uses transform pipeline to translate the generator output (sequence of token IDs) to text.

Wrapper class accepts either: configuration parameter which is a string indicating which bundler object to use (currently only supports 'base', but will update once other configurations are added), or parameters checkpoint, t5_conf, and transform if custom model architecture and weights are to be used.

Testing

Feed wrapper class input text and ensure that it returns the expected output. Tests are performed for wrapper instantiation using both configuration and custom settings. Scripted and non-scripted versions are also tested.

pytest test/prototype/integration_tests/test_models.py

[Nayef211]

Overall LGTM. How difficult would it be to enable the wrapper class to work with other model configurations?

[pmabbo13]

Overall LGTM. How difficult would it be to enable the wrapper class to work with other model configurations?

It would be straight forward. We could update line 43 in wrapper.py to assert that the value passed in for configuration is in ('base', 'small', 'large', '3b', '11b'), and then use if statements to load the bundler object that corresponds to the input configuration.