def postprocess()

in training/flax/distil_whisper/pipeline.py [0:0]

• 18 lines of code
• 5 McCabe index (conditional complexity)

    def postprocess(self, model_outputs, return_timestamps=None, return_language=None):
        # unpack the outputs from list(dict(list)) to list(dict)
        model_outputs = [dict(zip(output, t)) for output in model_outputs for t in zip(*output.values())]

        time_precision = self.feature_extractor.chunk_length / self.model.config.max_source_positions
        # Send the chunking back to seconds, it's easier to handle in whisper
        sampling_rate = self.feature_extractor.sampling_rate
        for output in model_outputs:
            if "stride" in output:
                chunk_len, stride_left, stride_right = output["stride"]
                # Go back in seconds
                chunk_len /= sampling_rate
                stride_left /= sampling_rate
                stride_right /= sampling_rate
                output["stride"] = chunk_len, stride_left, stride_right

        text, optional = self.tokenizer._decode_asr(
            model_outputs,
            return_timestamps=return_timestamps,
            return_language=return_language,
            time_precision=time_precision,
        )
        return {"text": text, **optional}