adding the time warping functionality

[noahweber1]

Difference of this notebook to the previous baseline https://github.com/pinellolab/DNA-Diffusion/blob/dna-diffusion/notebooks/experiments/conditional_diffusion/dna_diff_baseline_conditional_UNET.ipynb

is time warping.

Time warping (from https://arxiv.org/pdf/2211.15089.pdf)

The basic idea here is that we can smartly select the noise and use it to learn more efficiently. Selection of the noise happens through selecting optimal time values.

It has been tested and we have similiar computation time while having lower KL values under same seed setting as previously.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[ttunja]

Fixed Noahs version:

1. Erased initial time_warping (no use of time_warping as in CDCD paper, since our t space discrete and taking CDF of discrete t-s doesn't make much sense). Instead implemented simple version of "time_warping" where we learn for t-s where loss is big.
2. Fixed kl_divergence metrics (before was relative entropy, similar)
3. Added Jensen-Shannon divergence as a symetric metrics.
4. Trained and seen the convergence of the loss (converges faster than without time_warping with small additional time complexitiy)