This PR adds none_encoded_decoded hint usage mode, which preserves the same computation graphs to encoded_decoded mode but without computing hint losses.
On average this mode performs better than the ‘none’ hint usage mode, but the main goal of adding this mode is to provide a clearer comparison of Hints vs. No hints modes, keeping computation graphs the same.
This modification was inspired by the paper Neural Algorithmic Reasoning Without Intermediate Supervision (https://arxiv.org/abs/2306.13411), but note that the main results from this paper require additional modifications.
This PR adds none_encoded_decoded hint usage mode, which preserves the same computation graphs to encoded_decoded mode but without computing hint losses. On average this mode performs better than the ‘none’ hint usage mode, but the main goal of adding this mode is to provide a clearer comparison of Hints vs. No hints modes, keeping computation graphs the same. This modification was inspired by the paper Neural Algorithmic Reasoning Without Intermediate Supervision (https://arxiv.org/abs/2306.13411), but note that the main results from this paper require additional modifications.