Frequently asked questions
How to use the GPU?
Using a GPU may significantly speed up Novae's training or inference.
If you have a valid GPU for PyTorch, you can set the accelerator argument (e.g., one of ["cpu", "gpu", "tpu", "hpu", "mps", "auto"]) in the following methods: model.fit(), model.fine_tune(), model.compute_representations().
When using a GPU, we also highly recommend setting multiple workers to speed up the dataset __getitem__. For that, you'll need to set the num_workers argument in the previous methods, according to the number of CPUs available (num_workers=8 is usually a good value).
For more details, refer to the API of the PyTorch Lightning Trainer and to the API of the PyTorch DataLoader.
How to load a pretrained model?
We highly recommend loading a pre-trained Novae model instead of re-training from scratch. For that, choose an available Novae model name on our HuggingFace collection, and provide this name to the model.save_pretrained() method:
from novae import Novae
model = Novae.from_pretrained("MICS-Lab/novae-human-0") # or any valid model name
How to avoid overcorrecting?
By default, Novae corrects the batch-effect to get shared spatial domains across slides.
The batch information is used only during training (fit or fine_tune), which should prevent Novae from overcorrecting in zero_shot mode.
If not using the zero_shot mode, you can provide the min_prototypes_ratio parameter to control batch effect correction: either (i) in the fine_tune method itself, or (ii) during the model initialization (if retraining a model from scratch).
For instance, if min_prototypes_ratio=0.5, Novae expects each slide to contain at least 50% of the prototypes (each prototype can be interpreted as an "elementary spatial domain"). Therefore, the lower min_prototypes_ratio, the lower the batch-effect correction. Conversely, if min_prototypes_ratio=1, all prototypes are expected to be found in all slides (this doesn't mean the proportions will be the same overall slides, though).
How do I save my own model?
If you have trained or fine-tuned your own Novae model, you can save it for later use. For that, use the model.save_pretrained() method as below:
Then, you can load this model back via the model.from_pretrained() method:
How to turn lazy loading on or off?
By default, lazy loading is used only on large datasets. To enforce a specific behavior, you can do the following:
# never use lazy loading
novae.settings.disable_lazy_loading()
# always use lazy loading
novae.settings.enable_lazy_loading()
# use lazy loading only for AnnData objects with 1M+ cells
novae.settings.enable_lazy_loading(n_obs_threshold=1_000_000)
How to update the logging level?
The logging level can be updated as below:
import logging
from novae import log
log.setLevel(logging.ERROR) # or any other level, e.g. logging.DEBUG
How to disable auto-preprocessing?
By default, Novae will preprocess your adata object if you provide raw counts. More specifically, it will consider you have raw counts if adata.X.max() > 10, and then it will run sc.pp.normalize_total and sc.pp.log1p. To avoid that, you can run:
How to disable the multimodal mode?
By default, Novae will use multimodal mode if and only if you ran novae.compute_histo_embeddings and novae.compute_histo_pca.
To force not using the multimodal mode although you already computed H&E embeddings per cell, you can run:
How long does it take to use Novae?
The pip installation of Novae usually takes less than a minute on a standard laptop. The inference time depends on the number of cells, but typically takes 5-20 minutes on a CPUs, or 30sec to 2 minutes on a GPU (expect it to be roughly 10x times faster on a GPU).
How to train a new Novae model?
You can decide to train your own Novae model. To do that, simply create a new model as below and use the model.fit method.
model = novae.Novae(adata) # one or a list of adata objects
model.fit(adata, accelerator="cuda", num_workers=4)
How to monitor the model training?
Since we use Pytorch Lightning to train Novae, you can provide an existing Lightning Logger to the fit or fine_tune methods. For instance, there is an existing logger for Weight and Biases, which will allow you to aggregate plenty of training metrics, such as loss curves, hardware usage, time, and many others.
If you don't want to use a monitoring library but just want to see the model loss curve, you can also use a CSVLogger and then use our novae.plot.loss_curve function. For instance:
from lightning.pytorch.loggers import CSVLogger
# save the logs in a directory called "logs"
model.fine_tune(logger=CSVLogger("logs"), log_every_n_steps=10)
novae.plot.loss_curve("logs")
How to contribute?
If you want to contribute, check our contributing guide.
How to resolve any other issue?
If you have any bugs/questions/suggestions, don't hesitate to open a new issue. We'll be happy to help!