demo_t5_qa_pipe.py

import torch
import tensorflow as tf
import numpy as np
from transformers import Text2TextGenerationPipeline

class DemoT5QAPipeline(Text2TextGenerationPipeline):
    def _forward(self, model_inputs, **generate_kwargs):
        if self.framework == "pt":
            in_b, input_length = model_inputs["input_ids"].shape
        elif self.framework == "tf":
            in_b, input_length = tf.shape(model_inputs["input_ids"]).numpy()

        self.check_inputs(
            input_length,
            generate_kwargs.get("min_length", self.model.config.min_length),
            generate_kwargs.get("max_length", self.model.config.max_length),
        )
        outputs = self.model.generate(**model_inputs, **generate_kwargs, return_dict_in_generate=True, output_scores=True, max_new_tokens=75)
        
        # Code from the parent class
        output_ids = outputs.sequences
        out_b = output_ids.shape[0]
        if self.framework == "pt":
            output_ids = output_ids.reshape(in_b, out_b // in_b, *output_ids.shape[1:])
        elif self.framework == "tf":
            output_ids = tf.reshape(output_ids, (in_b, out_b // in_b, *output_ids.shape[1:]))

        output_sequences = outputs.sequences
        output_scores = outputs.scores
        return {"output_ids": output_ids, "output_sequences": output_sequences, "output_scores": output_scores}
    
    def postprocess(self, model_outputs):
        guess_text = super().postprocess(model_outputs)[0]['generated_text']

        transition_scores = self.model.compute_transition_scores(model_outputs['output_sequences'], model_outputs['output_scores'], normalize_logits=True)
        log_probs = np.round(np.exp(transition_scores.cpu().numpy()), 3)[0]
        guess_prob = np.product(log_probs)

        return {'guess': guess_text, 'confidence': guess_prob}