non_vectorized_decode.py

from pathlib import Path
import json

with Path('resources/tokens.json').open() as f:
    tokens = json.load(f)

def decode(predictions):
    # 'predictions' is 2D array, shape (num_frames, 999)
    # For each frame, the index in range [0-998] with the largest value tells
    # you the predicted token.
    # Add the token to the text string unless it is the same as the previous token
    # or is equal to '_'.

    num_frames = predictions.shape[0]

    text = ''
    prev_token = '_'
    for i in range(num_frames):
        max_index = -1
        max_value = -1e20
        for j in range(999):
            if predictions[i, j] > max_value:
                max_index = j
                max_value = predictions[i, j]

        token = tokens[max_index]
        if token != '_' and token != prev_token:
            text += token
        prev_token = token

    return text