Phd is a long and likely hard journey, if some tools can make the way easier we won’t complain about it. I want to gather tools that i find useful in this post.

It can be related to bibliography, writing, coding or anything that pass to mind, it should be updated at some point when new things come to my mind.

Research

Scholar inbox

This tool is so nice to be updated anytime an article in our field of research is published. You need to create an account an train a recommendation system based on literature that you are interested in. When it’s done, you will have new article pre-published in the past week that relates to your research.

S/O to MG that let me know this tool !

edit-0 : It makes the finding of new article very easy, but the rate at which new article are brought to you is higher than the reading speed unfortunately. So it just make your reading list increases haha. Still need to see how to manage it.

Bibliography

I can’t recommend Zotero enough. You can save article from the browser in one click. Read and annotate all in one place.

Folder VS Tags debate: For the document management, i would recommend you to think of all the folders and subfolders you would make and not create them. Usually, one document belongs to multiple categories, and you don’t want to duplicate all your sources et each folders/subfolders. That is why i would assign tags to each document at the time you add them to you library. That way, you can search your document by tag and it makes it very easy.

Note taking

I like to use obsidian with the citation extension. It connects the notes with the zotero library and allows you to link all the article and is very helpful when you want to write about any topic.

Literature Graph

Connected papers : not free. Lit map could be an alternative.

Redaction

my favorite software of this stack Quarto. It allow you to write anything in markdown and generate pdfs (with latex or typst), beamers or revealjs and this blog your are currently reading !

It is an amazing piece of software based on Pandoc the Haskell unicorn.

If you are in stats or working with any type of data, quarto is nice since you can code in the same file as the one you write in. It is like a notebook but plain text, that allows version control systems to follow it nicely.

Coding

When you code neural networks there are many things you don’t want to take care of. I understand we would want to control everything for pedagogy purposes but when you want to write reproducible research code that is going to be shared you want to write your pipeline with reproducible tools that make the training and the parametrization agnostic from the system features (os, devices, nodes …).

That is why for my Deep Learning stack i chose to use the following libraries/framework :

Lightning : library that handle distributed training, validation testing with a full control over training loop.
torchmetrics : Handle the computation of any metrics you would want in a distributed manner. You can write your own if you want and the device gathering will be taken care of.
SMP : Set of models usable in semantic segmentation, there are losses as well (be careful when using them on empty mask)
Torchcp : Library that contains the boilerplate to quantificate the uncertainty one gpu. I prefer to use this lib since all algorithms can be ran on GPU, and that is a requirement in computer vision.

--- title: Tools for Phd author: "Julien Combes" date: "2025-05-27" categories: [Phd, productivity, ENG] image: "image.jpeg" --- Phd is a long and likely hard journey, if some tools can make the way easier we won't complain about it. I want to gather tools that i find useful in this post. It can be related to bibliography, writing, coding or anything that pass to mind, it should be updated at some point when new things come to my mind. # Research ## Scholar inbox [This tool](https://www.scholar-inbox.com/) is so nice to be updated anytime an article in our field of research is published. You need to create an account an train a recommendation system based on literature that you are interested in. When it's done, you will have new article pre-published in the past week that relates to your research. S/O to MG that let me know this tool ! edit-0 : It makes the finding of new article very easy, but the rate at which new article are brought to you is higher than the reading speed unfortunately. So it just make your reading list increases haha. Still need to see how to manage it. ## Bibliography I can't recommend [Zotero](https://www.zotero.org/) enough. You can save article from the browser in one click. Read and annotate all in one place. **Folder VS Tags debate**: For the document management, i would recommend you to think of all the folders and subfolders you would make and not create them. Usually, one document belongs to multiple categories, and you don't want to duplicate all your sources et each folders/subfolders. That is why i would assign tags to each document at the time you add them to you library. That way, you can search your document by tag and it makes it very easy. ## Note taking I like to use [obsidian](https://obsidian.md/) with the citation extension. It connects the notes with the zotero library and allows you to link all the article and is very helpful when you want to write about any topic. ## Literature Graph Connected papers : not free. Lit map could be an alternative. ## Redaction my favorite software of this stack [Quarto](https://quarto.org/). It allow you to write anything in markdown and generate pdfs (with latex or typst), beamers or revealjs and this blog your are currently reading ! It is an amazing piece of software based on [Pandoc](https://pandoc.org/) the Haskell unicorn. If you are in stats or working with any type of data, quarto is nice since you can code in the same file as the one you write in. It is like a notebook but plain text, that allows version control systems to follow it nicely. # Coding When you code neural networks there are many things you don't want to take care of. I understand we would want to control everything for pedagogy purposes but when you want to write reproducible research code that is going to be shared you want to write your pipeline with reproducible tools that make the training and the parametrization agnostic from the system features (os, devices, nodes ...). That is why for my Deep Learning stack i chose to use the following libraries/framework : - [Lightning](https://lightning.ai/docs/pytorch/stable/) : library that handle distributed training, validation testing with a full control over training loop. - [torchmetrics](https://lightning.ai/docs/torchmetrics/stable/) : Handle the computation of any metrics you would want in a distributed manner. You can write your own if you want and the device gathering will be taken care of. - [SMP](https://segmentation-models-pytorch.readthedocs.io/en/latest/quickstart.html) : Set of models usable in semantic segmentation, there are losses as well (be careful when using them on empty mask) - [Torchcp](https://github.com/ml-stat-Sustech/TorchCP) : Library that contains the boilerplate to quantificate the uncertainty one gpu. I prefer to use this lib since all algorithms can be ran on GPU, and that is a requirement in computer vision.