rnnt_loss() gives different outputs (increasing) executed on same input

[stefan-falk]

I am using https://github.com/HawkAaron/warp-transducer indirectly via https://github.com/noahchalifour/rnnt-speech-recognition and I have noticed something odd when running the rnnt_loss() function in eager mode on the same input over and over again.

Basically: Running rnnt_loss(*args) repeatedly where args is always the same input results in an increasing loss.

from warprnnt_tensorflow import rnnt_loss
import numpy as np


def main():
    acts = np.asarray([
        [
            [[0.0, 0.0, 0.0],
             [0.0, 0.0, 0.0]],
            [[0.0, 0.0, 0.0],
             [0.0, 0.0, 0.0]],
            [[0.0, 0.0, 0.0],
             [0.0, 0.0, 0.0]],
        ]
    ])

    labels = np.asarray([[1, 2, 0]])
    label_lengths = [len(t) for t in labels]

    for i in range(10):
        loss = rnnt_loss(
            acts=acts,
            labels=labels,
            input_lengths=label_lengths,
            label_lengths=label_lengths
        )
        print(np.mean(loss))


if __name__ == '__main__':
    main()

Output:

1.0986123
2.1490226
5.274593
6.7222075
9.581686
11.274273
13.95323
15.808798
18.36151
20.329256

Is this expected behavior or am I doing something wrong here?

See also https://github.com/noahchalifour/rnnt-speech-recognition/issues/36

[cynecx]

I think the issue here is that the input is malformed. The label_lengths is [3] so U should be at least 3, hence the third dimension should be U+1=4 sized (So you could either adjust the label_length or adjust the 4-d acts tensor).

The validation of the input tensors is quite incomplete here which can cause Undefined Behavior and memory corruption as you could reproduce with your code. In my case it segfaults because of invalid writes (classic buffer overflow) which causes a corruption in the internal heap management structures.