alignment.py

# SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: MIT
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.
import sys
import numpy as np
from matplotlib import pylab as plt
from numba import jit


def save_plot(fname, attn_map):
    plt.imshow(attn_map)
    plt.savefig(fname)

@jit(nopython=True)
def mas_width1(attn_map):
    """mas with hardcoded width=1"""
    # assumes mel x text
    opt = np.zeros_like(attn_map)
    attn_map = np.log(attn_map)
    attn_map[0, 1:] = -np.inf
    log_p = np.zeros_like(attn_map)
    log_p[0, :] = attn_map[0, :]
    prev_ind = np.zeros_like(attn_map, dtype=np.int64)
    for i in range(1, attn_map.shape[0]):
        for j in range(attn_map.shape[1]):  # for each text dim
            prev_log = log_p[i-1, j]
            prev_j = j

            if j-1 >= 0 and log_p[i-1, j-1] >= log_p[i-1, j]:
                prev_log = log_p[i-1, j-1]
                prev_j = j-1

            log_p[i, j] = attn_map[i, j] + prev_log
            prev_ind[i, j] = prev_j

    # now backtrack
    curr_text_idx = attn_map.shape[1]-1
    for i in range(attn_map.shape[0]-1, -1, -1):
        opt[i, curr_text_idx] = 1
        curr_text_idx = prev_ind[i, curr_text_idx]
    opt[0, curr_text_idx] = 1
    return opt

if __name__ == '__main__':
    attn_ = np.load(sys.argv[1])
    attn = attn_.squeeze()
    save_plot('orig.png', attn)
    binarized = mas(attn)
    save_plot('binarized.png', binarized)