blog: Converting a Jupyter Notebook to a DVC Project #3624
Conversation
flippedcoder
force-pushed
the
blog/vscode-dvc-tutorial
branch
from
b37a154 to
9bc45ea
Compare
Link Check Report
All 6 links passed! |
content/blog/2022-07-28-switching-to-dvc-from-jupyter-vscode.md
Outdated
Show resolved
Hide resolved
content/blog/2022-07-28-switching-to-dvc-from-jupyter-vscode.md
Outdated
Show resolved
Hide resolved
| metrics, and then save the model. That's what we're doing in the | ||
| `bicycle_experiments.ipynb` file. | ||
|
|
||
|  |
There was a problem hiding this comment.
What do you think about starting with the parameters inside the notebook instead of reading them from params.yaml? I know it's harder to then convert to DVC, but it feels a bit unrealistic to be starting in a notebook but already have a parameters file configured.
There was a problem hiding this comment.
It looks like this was not addressed. And I'm finding that for some reason when the notebook opens, all the cells are in Markdown. I thought this was a problem in VS Code, but when I open in Jupyter in a browser it's the same. @RCdeWit we are going to have to figure out how to fix this notebook to make it more representative of real life.
| Next, we'll make a `evaluate.py` file that will take a saved model and get the | ||
| metrics for how well it performs. This file will have the `Set test variables`, | ||
| `Load model and test data`, `Get model predictions`, | ||
| `Calculate model performance metrics`, and `Save model performance metrics` |
There was a problem hiding this comment.
Similar to params.yaml above, it seems unlikely that someone is already saving the model and especially the performance metrics to a file if they are working in a notebook.
There was a problem hiding this comment.
This one also needs to be addressed in the notebook.
content/blog/2022-07-28-switching-to-dvc-from-jupyter-vscode.md
Outdated
Show resolved
Hide resolved
|
Great subject! I think it's worth spending a little time iterating on this to make it as helpful as possible since it's a common topic. |
Co-authored-by: Rob de Wit <[email protected]>
There was a problem hiding this comment.
might be worth mentioning (as an alternative) https://papermill.readthedocs.io/en/latest/usage-execute.html#execute-via-cli, i.e. instead of:
python src/train.py ./data/ ./models/model.pklyou do:
papermill src/train.ipynb -r data ./data/ -r model ./models/model.pklwith a param cell in the src/train.ipynb notebook:
# parameters
data = "./data/"
model = "./models/model.pkl"
| extension, we can show you how to make those experiments reproducible with the | ||
| addition of the DVC VS Code extension. | ||
| picture: 2022-07-28/jupyter-to-dvc.png | ||
| pictureComment: Using the DVC VS Code Extension with a Jupyter Notebook |
There was a problem hiding this comment.
| pictureComment: Using the DVC VS Code Extension with a Jupyter Notebook | |
| pictureComment: Using the DVC Extension for VS Code with a Jupyter Notebook |
|
|
||
| Each of these stages has a `cmd` that executes the Python scripts we wrote with | ||
| the required arguments. They both have defined dependencies in `deps` that let | ||
| DVC know what needs to be available for a stage to execute before it starts |
There was a problem hiding this comment.
To make it even more explicit, maybe add something along the lines here of
"As you can see, for example, the training stage is listed as a requirement for the evaluation stage. This ensures that the latter will only start once the first has been completed."
There was a problem hiding this comment.
And maybe as an admonition somewhere saying that when using dvc exp run only the stages downstream from your changes are triggered?
|
Closing in favor of https://github.com/iterative/iterative.ai/pull/550. |
|
@yathomasi I noticed as I was going over comments and checking Milecia's repo that her link to her blog post in this repo is broken. I thought they would redirect to the new site. |
|
The link was missing. I have added the link. Readme |
Tentative publish date:
07/19/2207/28/2208/24/22Things left to update: