Python T5 base model

Pre-trained model on CodeSearchNet Python dataset using a span-masking objective. The training objective and model were introduced in this paper and first released in this repository. PyT5 model used git-t5 framework built on top of JAX/Flax to pre-train the model on a TPU v3-8 node.

How to use

You can use this model to denoise span-masked sequences.

First, install the git-t5 pip package:

> pip install git-t5

Next, download the model and tokenizer:

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, 

model = AutoModelForSeq2SeqLM.from_pretrained("formermagic/pyt5-base")

tokenizer = AutoTokenizer.from_pretrained("formermagic/pyt5-base")

Finally, encode your input and generate the output sequence:

from git_t5.utils import encode_input

text = """
def alias(self, annotationtype, set, fallback=False):
    if inspect.isclass(annotationtype): annotationtype = annotationtype.ANNOTATIONTYPE
    if annotationtype in self.set_alias and set in self.set_alias[annotationtype]:
        return self.set_alias[annotationtype][set]
    elif fallback:
        return set
        raise KeyError("No alias for set " + set)

batch, max_length = encode_input(tokenizer, text, seed=22)
outputs = model.generate(batch["input_ids"], max_length=max_length, num_beams=1)
print(tokenizer.batch_decode(outputs[..., 1:]))

You should see the following output:

['<extra_id_0>, fallback=<extra_id_1> inspect<extra_id_2>.set_alias<extra_id_3> return self.set<extra_id_4>) def fallback']
['<extra_id_0>, fallback=<extra_id_1> inspect<extra_id_2>.set_alias<extra_id_3> return self.set<extra_id_4>) </s></s>']

As you can see, the predicted result is very close to the target sequence.

Downloads last month
Hosted inference API
This model can be loaded on the Inference API on-demand.