3. Single CRS Modules

Specially, as shown in Case 2, we can load CRS modules first. Also, we can use the CRS (recommender or generator) module independently, let us use the recommender module from ReDIAL as an example:

from recwizard import RedialRec, RedialRecTokenizer

# load redial recommender module
redial_rec_module = RedialRec.from_pretrained('RecWizard/redial-rec')

# load redial recommender module tokenizer (convert text data to tensor data)
redial_rec_tokenizer = RedialRecTokenizer.from_pretrained('RecWizard/redial-rec')

3.1. Deal with Text Data

We can play with the module directly using natural language inputs, where if you wrap all the mentioned item titles with <entity> labels, such as <entity>Titanic</entity>, the Titanic mention will be mapped to the corresponding item id. Otherwise, the Titanic mention will be mapped to normal corresponding word id:

redial_rec_module.response(
    raw_input="Recommend me some popular and classic movies, I like <entity>Titanic</entity>.",
    tokenizer=redial_rec_tokenizer,
    topk=5,
)

Then we obtain the top-5 results from the ReDIAL recommender module:

[
 'The Big Sick (2017)',
 'Bad Moms (2016)',
 'Dumb and Dumber (1994)',
 '21 Jump Street (2012)',
 'Star Wars: Episode VIII – The Last Jedi (2017)'
]

3.2. Deal with Tensor Data

All the above examples are using the .response method to deal with text data, but for the internal mechanism within (recommender or generator) modules, they are typically dealing with tensor data. So, a lower-level control is also provided if the users are curious about the tensor data. Bascially, response is equivalent to “tokenzier encode -> module forward -> tokenizer decode.

**3.2.1. Tokenzier encode**

raw_input='User: Hi I am looking for a movie like <entity>Super Troopers (2001)</entity>'
tensor_input = redial_rec_tokenizer([raw_input])
tensor_input

{'attention_mask': tensor([[[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]]), 'senders': tensor([[1]]), 'input_ids': tensor([[[    0, 30086,    38,   524,   546,    13,    10,  1569,   101,  1582,
           6354, 18158,     2]]]), 'movieIds': tensor([[1901]]), 'conversation_lengths': tensor([1]), 'movie_occurrences': [tensor([[[0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 0., 0., 0.]]])]}

**3.2.2. Module forward**

tensor_output = redial_rec_module(**tensor_input) # equivalent to redial_rec_module.forward(**tensor_input)
print(tensor_output)

tensor([[[-20.4876, -13.0947, -20.1804,  ..., -20.6572, -21.6375, -20.9953]]],
       grad_fn=<AsStridedBackward0>)

**3.2.2. Tokenizer decode**

movieIds = tensor_output.topk(k=3, dim=-1).indices[:, -1,
                   :].tolist()
tokenizer.decode(movieIds[0])

['The Big Sick (2017)', 'Bad Moms (2016)', 'Dumb and Dumber (1994)']

Additionally, if you are interestd in more existing modules, please check our resource section.