Model hyperparameter tuning with scVI

Warning

scvi.autotune development is still in progress. The API is subject to change.

Finding an effective set of model hyperparameters (e.g. learning rate, number of hidden layers, etc.) is an important component in training a model as its performance can be highly dependent on these non-trainable parameters. Manually tuning a model often involves picking a set of hyperparameters to search over and then evaluating different configurations over a validation set for a desired metric. This process can be time consuming and can require some prior intuition about a model and dataset pair, which is not always feasible.

In this tutorial, we show how to use scvi’s autotune module, which allows us to automatically find a good set of model hyperparameters using Ray Tune. We will use SCVI and a subsample of the heart cell atlas for the task of batch correction, but the principles outlined here can be applied to any model and dataset. In particular, we will go through the following steps:

1. Installing required packages

2. Loading and preprocessing the dataset

3. Defining the tuner and discovering hyperparameters

4. Running the tuner

5. Comparing latent spaces

6. Optional: Monitoring progress with Tensorboard

7. Optional: Tuning over integration metrics with scib-metrics

Installing required packages

Note

Running the following cell will install tutorial dependencies on Google Colab only. It will have no effect on environments other than Google Colab.

!pip install --quiet scvi-colab
from scvi_colab import install

install()

import os
import tempfile

import ray
import scanpy as sc
import scvi
import seaborn as sns
import torch
from ray import tune
from scvi import autotune

os.environ["JAX_PLATFORMS"] = "cpu"
os.environ["TUNE_DISABLE_STRICT_METRIC_CHECKING"] = "1"

scvi.settings.seed = 0
print("Last run with scvi-tools version:", scvi.__version__)

Last run with scvi-tools version: 1.2.2

Note

You can modify save_dir below to change where the data files for this tutorial are saved.

sc.set_figure_params(figsize=(6, 6), frameon=False)
sns.set_theme()
torch.set_float32_matmul_precision("high")
save_dir = tempfile.TemporaryDirectory()
scvi.settings.logging_dir = save_dir.name

%config InlineBackend.print_figure_kwargs={"facecolor": "w"}
%config InlineBackend.figure_format="retina"

Loading and preprocessing the dataset

adata = scvi.data.heart_cell_atlas_subsampled(save_path=save_dir.name)
adata.layers["counts"] = adata.X.copy()  # preserve counts
adata

INFO     Downloading file at /tmp/tmpv690v5hw/hca_subsampled_20k.h5ad

AnnData object with n_obs × n_vars = 18641 × 26662
    obs: 'NRP', 'age_group', 'cell_source', 'cell_type', 'donor', 'gender', 'n_counts', 'n_genes', 'percent_mito', 'percent_ribo', 'region', 'sample', 'scrublet_score', 'source', 'type', 'version', 'cell_states', 'Used'
    var: 'gene_ids-Harvard-Nuclei', 'feature_types-Harvard-Nuclei', 'gene_ids-Sanger-Nuclei', 'feature_types-Sanger-Nuclei', 'gene_ids-Sanger-Cells', 'feature_types-Sanger-Cells', 'gene_ids-Sanger-CD45', 'feature_types-Sanger-CD45', 'n_counts'
    uns: 'cell_type_colors'
    layers: 'counts'

The only preprocessing step we will perform in this case will be to subsample the dataset for 2000 highly variable genes using scanpy for faster model training.

sc.pp.highly_variable_genes(adata, n_top_genes=2000, flavor="seurat_v3", subset=True)
adata

AnnData object with n_obs × n_vars = 18641 × 2000
    obs: 'NRP', 'age_group', 'cell_source', 'cell_type', 'donor', 'gender', 'n_counts', 'n_genes', 'percent_mito', 'percent_ribo', 'region', 'sample', 'scrublet_score', 'source', 'type', 'version', 'cell_states', 'Used'
    var: 'gene_ids-Harvard-Nuclei', 'feature_types-Harvard-Nuclei', 'gene_ids-Sanger-Nuclei', 'feature_types-Sanger-Nuclei', 'gene_ids-Sanger-Cells', 'feature_types-Sanger-Cells', 'gene_ids-Sanger-CD45', 'feature_types-Sanger-CD45', 'n_counts', 'highly_variable', 'highly_variable_rank', 'means', 'variances', 'variances_norm'
    uns: 'cell_type_colors', 'hvg'
    layers: 'counts'

Defining the tuner and discovering hyperparameters

The first part of our workflow is the same as the standard scvi-tools workflow: we start with our desired model class, and we register our dataset with it using setup_anndata. All datasets must be registered prior to hyperparameter tuning.

model_cls = scvi.model.SCVI
model_cls.setup_anndata(adata, layer="counts", labels_key="cell_type", batch_key="cell_source")

Our main entry point to the autotune module is the ModelTuner class, a wrapper around ray.tune.Tuner with additional functionality specific to scvi-tools. We can define a new ModelTuner by providing it with our model class.

ModelTuner will register all tunable hyperparameters in SCVI and all of them are accessible to tune starting version 1.2.0.

Running the tuner

Now that we know what hyperparameters are available to us, we can define a search space using the search space API in ray.tune. For this tutorial, we choose a simple search space with two model hyperparameters and one training plan hyperparameter. These can all be combined into a single dictionary that we pass into the fit method.

search_space = {
    "model_params": {"n_hidden": tune.choice([64, 128, 256]), "n_layers": tune.choice([1, 2, 3])},
    "train_params": {"max_epochs": 100, "plan_kwargs": {"lr": tune.loguniform(1e-4, 1e-2)}},
}

There are a couple more arguments we should be aware of before fitting the tuner:

• num_samples: The total number of hyperparameter sets to sample from ``. This is the total number of models that will be trained.

  For example, if we set num_samples=2, based on the above search space, we might sample two models with the following hyperparameter configurations:

  model1 = {
      "n_hidden": 64,
      "n_layers": 1,
      "lr": 0.001,
  }
  model2 = {
      "n_hidden": 64,
      "n_layers": 3,
      "lr": 0.0001,
  }
  

• max_epochs: The maximum number of epochs to train each model for.

  Note: This does not mean that each model will be trained for max_epochs. Depending on the scheduler used, some trials are likely to be stopped early.

• resources: A dictionary of maximum resources to allocate for the whole experiment. This allows us to run concurrent trials on limited hardware.

Now, we can call fit on the tuner to start the hyperparameter sweep. This will return a TuneAnalysis dataclass, which will contain the best set of hyperparameters, as well as other information.

ray.init(log_to_driver=False)

results = autotune.run_autotune(
    model_cls,
    data=adata,
    mode="min",
    metrics="validation_loss",
    search_space=search_space,
    num_samples=5,
    resources={"cpu": 10, "gpu": 1},
)

Tune Status

Current time:	2025-02-16 17:24:37
Running for:	00:02:52.19
Memory:	22.2/377.2 GiB

System Info

Using AsyncHyperBand: num_stopped=5
Bracket: Iter 64.000: -465.25177001953125 | Iter 32.000: -465.0001220703125 | Iter 16.000: -469.39208984375 | Iter 8.000: -475.1295166015625 | Iter 4.000: -489.63555908203125 | Iter 2.000: -503.7332458496094 | Iter 1.000: -572.58056640625
Logical resource usage: 10.0/64 CPUs, 1.0/2 GPUs (0.0/1.0 accelerator_type:G)

Trial Status

Trial name	status	loc	model_params/n_hidde n	model_params/n_layer s	train_params/max_epo chs	train_params/plan_kw args/lr	iter	total time (s)	validation_loss
_trainable_92f2cbb7	TERMINATED	132.77.199.125:2917995	128	1	100	0.000582248	100	62.4929	468.383
_trainable_c6809c70	TERMINATED	132.77.199.125:2918336	256	3	100	0.00253497	100	73.6775	468.378
_trainable_925d2317	TERMINATED	132.77.199.125:2919407	128	2	100	0.00146047	1	1.18282	572.581
_trainable_ed6c9bcb	TERMINATED	132.77.199.125:2919527	64	2	100	0.000402269	1	1.18551	729.555
_trainable_16e11c59	TERMINATED	132.77.199.125:2919649	256	3	100	0.00314606	100	74.1104	468.666

print(results.result_grid)

ResultGrid<[
  Result(
    metrics={'validation_loss': 468.38323974609375},
    path='/tmp/tmpv690v5hw/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/_trainable_92f2cbb7_1_n_hidden=128,n_layers=1,max_epochs=100,lr=0.0006_2025-02-16_17-21-45',
    filesystem='local',
    checkpoint=None
  ),
  Result(
    metrics={'validation_loss': 468.3775634765625},
    path='/tmp/tmpv690v5hw/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/_trainable_c6809c70_2_n_hidden=256,n_layers=3,max_epochs=100,lr=0.0025_2025-02-16_17-21-53',
    filesystem='local',
    checkpoint=None
  ),
  Result(
    metrics={'validation_loss': 572.58056640625},
    path='/tmp/tmpv690v5hw/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/_trainable_925d2317_3_n_hidden=128,n_layers=2,max_epochs=100,lr=0.0015_2025-02-16_17-22-01',
    filesystem='local',
    checkpoint=None
  ),
  Result(
    metrics={'validation_loss': 729.5545043945312},
    path='/tmp/tmpv690v5hw/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/_trainable_ed6c9bcb_4_n_hidden=64,n_layers=2,max_epochs=100,lr=0.0004_2025-02-16_17-23-04',
    filesystem='local',
    checkpoint=None
  ),
  Result(
    metrics={'validation_loss': 468.6656188964844},
    path='/tmp/tmpv690v5hw/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/scvi_18d70fdd-bf19-4348-92c7-251479b6153d/_trainable_16e11c59_5_n_hidden=256,n_layers=3,max_epochs=100,lr=0.0031_2025-02-16_17-23-13',
    filesystem='local',
    checkpoint=None
  )
]>

Comparing latent spaces

Work in progress: please check back in the next release!

Optional: Monitoring progress with Tensorboard

Ray by default plots the metrics and results into tensorboard. It can be accessed usually under http://localhost:6006/ once the optimization is done

Optional: Tuning over integration metrics with scib-metrics

Starting SCVI-Tools v1.3 we are now able to hyperparameter tune for optimal scib-metrics of any choice. This is currently works only for SCVI and SCANVI models. Note that youll need to update for the recent scib-metrics jax based package.

The idea in performing this is very much the same as before, we just need to select a metric that can be optimize from the pool of metrics that scib-metrics holds as seen in the following list: “Silhouette label”, “Silhouette batch”, “Isolated labels”, “Leiden NMI”, “Leiden ARI”, “KMeans NMI”, “KMeans ARI”, “cLISI”, “iLISI”, “KBET”, “Graph connectivity”, “PCR comparison”, “Batch correction”, “Bio conservation”, “Total” The last 3 are the aggegative metrics that are average of the relevant group.

Those metrics are the columns names of the output result table of scib metrics. For all metrics, higher means better and range [0..1]. For more information about those metrics see the scib-metric docs

Note that usually in order to compare bio conservation metrics we need to apply a semi supervised model. Also note that in order to save run time we select a random number of cells to calculate the scib metrics. The default is 100 but it can be changed with the parameter scib_subsample_rows, like in the following example: Here we apply the hyper parameter tuning for the same model in the previous example, while optimizing for iLISI (a batch correction metric, see here)

search_space_scib = {
    "model_params": {"n_hidden": tune.choice([64, 128, 256]), "n_layers": tune.choice([1, 2, 3])},
    "train_params": {"max_epochs": 10, "plan_kwargs": {"lr": tune.loguniform(1e-4, 1e-2)}},
}

results_scib = autotune.run_autotune(
    model_cls,
    data=adata,
    metrics="iLISI",
    mode="max",
    search_space=search_space_scib,
    num_samples=2,
    seed=0,
    scheduler="asha",
    searcher="hyperopt",
    resources={"cpu": 20, "gpu": 0},
    scib_subsample_rows=1000,
    local_mode=True,
    ignore_reinit_error=True,
)

Tune Status

Current time:	2025-02-16 17:27:16
Running for:	00:02:38.67
Memory:	23.5/377.2 GiB

System Info

Using AsyncHyperBand: num_stopped=1
Bracket: Iter 64.000: None | Iter 32.000: None | Iter 16.000: None | Iter 8.000: None | Iter 4.000: None | Iter 2.000: None | Iter 1.000: 0.3259779065847397
Logical resource usage: 20.0/64 CPUs, 0/2 GPUs (0.0/1.0 accelerator_type:G)

Trial Status

Trial name	status	loc	model_params/n_hidde n	model_params/n_layer s	train_params/max_epo chs	train_params/plan_kw args/lr	iter	total time (s)	iLISI
_trainable_a334893d	TERMINATED	132.77.199.125:2919986	128	1	10	0.000582248	1	133.307	0.331632
_trainable_bfb1b38d	TERMINATED	132.77.199.125:2920123	256	3	10	0.00253497	1	141.401	0.320323

The first run ended with a higher iLISI score. We can visualize the different UMAPs to see that indeed the 1st choice gives a better integration for cell source than the other:

model1 = scvi.model.SCVI(adata, n_hidden=128, n_layers=1)
model1.train(max_epochs=10, plan_kwargs={"lr": 0.0005})
adata.obsm["X_scVI1"] = model1.get_latent_representation()

sc.pp.neighbors(adata, use_rep="X_scVI1")
sc.tl.umap(adata, min_dist=0.3)
sc.pl.umap(
    adata,
    color=["cell_type"],
    frameon=False,
)
sc.pl.umap(
    adata,
    color=["cell_source"],
    ncols=2,
    frameon=False,
)

model2 = scvi.model.SCVI(adata, n_hidden=256, n_layers=3)
model2.train(max_epochs=10, plan_kwargs={"lr": 0.0025})
adata.obsm["X_scVI2"] = model2.get_latent_representation()

sc.pp.neighbors(adata, use_rep="X_scVI2")
sc.tl.umap(adata, min_dist=0.3)
sc.pl.umap(
    adata,
    color=["cell_type"],
    frameon=False,
)
sc.pl.umap(
    adata,
    color=["cell_source"],
    ncols=2,
    frameon=False,
)

As can be seen, the 1st option of hyper paramters, did gave a bit better cell_source integration, as the ray tune suggested. This can be understood from seeying that the 1st option clusters have a better random distribution of colors (batches) in each of them, not only partialy integrated and that there are more clusters which are more distinct with each other.