add first version of SCOT demonstration notebook (Gromov-Wasserstein for multi-omics) #64
Conversation
antoinebelloir
changed the title
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:52Z the title here will define the length of that item in the left-menu, so maybe something shorter. antoinebelloir commented on 2022-05-08T23:24:54Z Changed to "Gromov-Wasserstein for multi-omics"
|
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:53Z Maybe follow the rest of the templates and do not provide a tl;dr but rather an introducing sentence |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:54Z typo in the Wasserstein |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:54Z I see \ characters there.
I think to keep in line with the other notebooks, the motivation should be shorter. antoinebelloir commented on 2022-05-08T23:27:47Z Agreed, the introduction paragraphs were too long. I changed it to a way shorter presentation, that sticks better with other notebooks
|
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:55Z I think it's fine to say that you substitute POT for OTT, but if you mention "found to be faster" I think it might be important to put a comparison. Otherwise maybe something a bit more informative (POT can be slower if, e.g., it does not have the same convergence settings).
I think you can remove the two bottom sentences starting with "At the end..." and "Note that..." antoinebelloir commented on 2022-05-08T23:34:10Z I agree that a comparison of OTT vs POT is useful. I added it to the notebook. You are right to point out the convergence settings : for the contest to be fair, I therefore used POT's default parameters for OTT (maximum number of iterations=1000 and threshold=1e-9) and added a comment to make this explicit. You will see that OTT runs way faster on GPU. I did not run comparison experiences on CPU.
|
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:56Z try to use code lines that have 80 characters max to follow python guidelines. antoinebelloir commented on 2022-05-08T23:36:48Z Following your comment I used Black to format most of my cells. It's cleaner now
|
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:57Z Line #3. %pip install seaborn Here you could remove the # antoinebelloir commented on 2022-05-08T23:36:58Z Done |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:57Z Line #6. def find_correspondences(self, e, verbose=True): call e epsilon would make it more readable antoinebelloir commented on 2022-05-08T23:40:23Z Agreed, I made a modification within "fin_correspondances" 's body but I could to change "e" to "epsilon" in the arguments without making the code heavier, since "e" is actually a keyword argument used by another method of the SCOT class |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:58Z Line #11. epsilon=e, max_iterations=200, are you using a different setting than that used by default for max_iterations? if yes might be worth commenting on this. antoinebelloir commented on 2022-05-08T23:41:39Z See my reply about OTT vs POT comparison. I changed max_iterations to 1000 and I set threshold to 1e-9 to match the POT implementation. |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:00:59Z for epsilon and k, you could use quotes or latex antoinebelloir commented on 2022-05-08T23:41:47Z Corrected |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:01:00Z same here for X and Y, use latex (markdown) antoinebelloir commented on 2022-05-08T23:41:56Z Also corrected |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:01:00Z check this sentence ("we actually") antoinebelloir commented on 2022-05-08T23:42:39Z I rewrote this cell to make it clearer and to mention the OTT vs POT comparison |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:01:01Z Line #25. anim = animation.FuncAnimation(fig, animate, init_func = init, if this is an animation with 2 plots, maybe plot them directly? antoinebelloir commented on 2022-05-08T23:47:36Z I see how using an animation for only two plots looks like an overkill, but I find it useful though to see the crosses (representing chromatin accessibility points mapped to gene expression domain) appear on top of the gene expression domain points (circles). I deleted the precedent visualisation that was a single plot of the two set of points (ie the second frame of the animation) as it was a duplicate |
|
View / edit / reply to this conversation on ReviewNB marcocuturi commented on 2022-05-05T21:01:02Z Line #26. frames= [0,1], # Ic, c'est la valeur de k french comment antoinebelloir commented on 2022-05-08T23:47:49Z Deleted |
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
Changed to "Gromov-Wasserstein for multi-omics"
View entire conversation on ReviewNB |
|
Agreed, the introduction paragraphs were too long. I changed it to a way shorter presentation, that sticks better with other notebooks
View entire conversation on ReviewNB |
|
I agree that a comparison of OTT vs POT is useful. I added it to the notebook. You are right to point out the convergence settings : for the contest to be fair, I therefore used POT's default parameters for OTT (maximum number of iterations=1000 and threshold=1e-9) and added a comment to make this explicit. You will see that OTT runs way faster on GPU. I did not run comparison experiences on CPU.
View entire conversation on ReviewNB |
|
Following your comment I used Black to format most of my cells. It's cleaner now
View entire conversation on ReviewNB |
|
Done View entire conversation on ReviewNB |
|
Agreed, I made a modification within "fin_correspondances" 's body but I could to change "e" to "epsilon" in the arguments without making the code heavier, since "e" is actually a keyword argument used by another method of the SCOT class View entire conversation on ReviewNB |
|
See my reply about OTT vs POT comparison. I changed max_iterations to 1000 and I set threshold to 1e-9 to match the POT implementation. View entire conversation on ReviewNB |
|
Corrected View entire conversation on ReviewNB |
|
Also corrected View entire conversation on ReviewNB |
|
I rewrote this cell to make it clearer and to mention the OTT vs POT comparison View entire conversation on ReviewNB |
|
I see how using an animation for only two plots looks like an overkill, but I find it useful though to see the crosses (representing chromatin accessibility points mapped to gene expression domain) appear on top of the gene expression domain points (circles). I deleted the precedent visualisation that was a single plot of the two set of points (ie the second frame of the animation) as it was a duplicate View entire conversation on ReviewNB |
|
Deleted View entire conversation on ReviewNB |
|
This is starting to look good! May I suggest that you also add the link to that notebook in https://github.com/ott-jax/ott/blob/main/docs/index.rst so that it is directly referenced in the doc? I would probably add it in the "advanced applications" section |
|
Thanks ! I added it to the "advanced applications" list just after the "Single-cell genomics" notebook in order to group biology-related examples |
|
Thanks! |
|
It was very interesting to contribute to this notebook ! Just a detail, the merged notebook has a nice display in the repo pre-visualisation but has two little display issues on the readthedocs.io website : |
|
Thanks Antoine! I think I fixed that. it seems you cannot put an url on top of a |
add first version of SCOT demonstration notebook (Gromov-Wasserstein for multi-omics)
Adds a second biology example to the documentation notebooks.
The notebook presents an application of OTT's Gromov Wasserstein optimal transport to match single-cell points clouds from two different measurement spaces (e.g. mapping gene expressions measurements to chromatine accessibility measurements).
It is adapted from Demetci et al, Gromov-Wasserstein optimal transport to align single-cell multi-omics data, ICML 2020 Workshop on Computational Biology, 2020 (pdf here).