Custom dataloader registry support #2932
Conversation
…try' into ori-2907-custom-dataloader-registry
…module / registry big change
for more information, see https://pre-commit.ci
…un, we will later adjust this file
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2932 +/- ##
==========================================
- Coverage 84.81% 83.90% -0.92%
==========================================
Files 173 173
Lines 14793 14920 +127
==========================================
- Hits 12547 12518 -29
- Misses 2246 2402 +156
|
and fix the test for custom dataloaders
src/scvi/model/_scvi.py
Outdated
| @setup_anndata_dsp.dedent | ||
| def setup_datamodule( | ||
| cls, | ||
| datamodule, # TODO: what to put here? |
There was a problem hiding this comment.
It should be pytorch.DataLoader right? Martin has done typing for it in the current code.
src/scvi/model/_scvi.py
Outdated
| "state_registry": { | ||
| "n_obs": datamodule.n_obs, | ||
| "n_vars": datamodule.n_vars, | ||
| "column_names": [str(i) for i in column_names], # TODO: from adata (czi)? |
| _validate_var_names(adata[modality], var_names[modality]) | ||
| logger.debug("Subsetting query vars to reference vars.") | ||
| adata._inplace_subset_var(var_names) | ||
| _validate_var_names(adata, var_names) |
There was a problem hiding this comment.
We need to verify for dataloaders that the gene names are matching.
| logger.debug("Subsetting query vars to reference vars.") | ||
| adata._inplace_subset_var(var_names) | ||
| _validate_var_names(adata, var_names) | ||
| registry = attr_dict.pop("registry_") |
There was a problem hiding this comment.
This check and the ones below are independent of datamodule or AnnData, right? Remove the indent.
There was a problem hiding this comment.
I am not sure why every code here is displayed as modified?
| @@ -202,7 +215,7 @@ def prepare_query_anndata( | |||
| Query adata ready to use in `load_query_data` unless `return_reference_var_names` | |||
| in which case a pd.Index of reference var names is returned. | |||
| """ | |||
| _, var_names, _ = _get_loaded_data(reference_model, device="cpu") | |||
| _, var_names, _ = _get_loaded_data(reference_model, device="cpu", adata=adata) | |||
There was a problem hiding this comment.
Does this work with a dataloader?
| @@ -350,15 +363,15 @@ def requires_grad(key): | |||
| par.requires_grad = False | |||
|
|
|||
|
|
|||
| def _get_loaded_data(reference_model, device=None): | |||
| def _get_loaded_data(reference_model, device=None, adata=None): | |||
| self.registry_ = registry | ||
| self.summary_stats = _get_summary_stats_from_registry(registry) | ||
| elif self.__class__.__name__ == "GIMVI": | ||
| # note some models do accept empty registry/adata (e.g: gimvi) |
There was a problem hiding this comment.
I'm not following this one. What is the exception with GIMVI?
src/scvi/model/base/_base_model.py
Outdated
| else: | ||
| return self._adata_manager.get_from_registry(registry_key) | ||
|
|
||
| # def get_from_registry(self, registry_key: str) -> np.ndarray | pd.DataFrame: |
| else: | ||
| # Case where correct AnnDataManager is found, replay registration as necessary. | ||
| adata_manager.validate() | ||
|
|
||
| return adata | ||
|
|
||
| def transfer_fields(self, adata: AnnOrMuData, **kwargs) -> AnnData: | ||
| """Transfer fields from a model to an AnnData object.""" | ||
| if self.adata: |
There was a problem hiding this comment.
where do we need transfer_fields? can we make it work with datamodule?
| @@ -627,8 +711,7 @@ def save( | |||
|
|
|||
| # save the model state dict and the trainer state dict only | |||
| model_state_dict = self.module.state_dict() | |||
|
|
|||
| var_names = _get_var_names(self.adata, legacy_mudata_format=legacy_mudata_format) | |||
| var_names = self.get_var_names(legacy_mudata_format=legacy_mudata_format) | |||
There was a problem hiding this comment.
do we need two get_var_names function?
| "Saved model does not contain original setup inputs. " | ||
| "Cannot load the original setup." | ||
| ) | ||
| _validate_var_names(adata, var_names) |
There was a problem hiding this comment.
should be validated also for a dataloader.
|
|
||
| def get_state_registry(self, registry_key: str) -> attrdict: | ||
| """Returns the state registry for the AnnDataField registered with this instance.""" | ||
| return attrdict(self.registry_[_FIELD_REGISTRIES_KEY][registry_key][_STATE_REGISTRY_KEY]) |
There was a problem hiding this comment.
Does this work with dataloader. Documentation should be updated then.
| @@ -133,7 +133,10 @@ def _initialize_model(cls, adata, attr_dict): | |||
| if "pretrained_model" in non_kwargs.keys(): | |||
| non_kwargs.pop("pretrained_model") | |||
|
|
|||
| model = cls(adata, **non_kwargs, **kwargs) | |||
| if not adata: | |||
| adata = None | |||
There was a problem hiding this comment.
Is adata false here? Do we need a default value for registry?
| if max_epochs is None: | ||
| if datamodule is None: | ||
| if self.adata is not None: |
There was a problem hiding this comment.
We should take here n_obs from summary stats to make it compatible with a dataloader.
There was a problem hiding this comment.
See below we don't need the if statement.
| experiment_name = "mus_musculus" | ||
| obs_value_filter = 'is_primary_data == True and tissue_general in ["kidney"] and nnz >= 3000' | ||
|
|
||
| # This is under comments just to save time (selecting highly varkable genes): |
There was a problem hiding this comment.
Remove this block, we don't need it.
| dataloader_kwargs={"num_workers": 0, "persistent_workers": False}, | ||
| ) | ||
|
|
||
| # table of genes should be filtered by soma_joinid - but we should keep the encoded indexes |
| model = scvi.model.SCVI(adata_orig, n_latent=10) | ||
| model.train(max_epochs=1) | ||
|
|
||
| # TODO: do we need to apply those functions to any census model as is? |
| _ = model.get_reconstruction_error(dataloader=dataloader) | ||
| _ = model.get_latent_representation(dataloader=dataloader) | ||
|
|
||
| scvi.model.SCVI.prepare_query_anndata(adata_orig, reference_model=model) |
There was a problem hiding this comment.
test this with a second model trained using dataloader
| n_layers = 1 | ||
| n_latent = 50 | ||
|
|
||
| scvi.model._scvi.SCVI.setup_datamodule(datamodule) # takes time |
|
|
||
| pprint(datamodule.registry) | ||
|
|
||
| batch_size = 1024 |
There was a problem hiding this comment.
does batch size have an effect. I thought it's defined by the datamodule?
There was a problem hiding this comment.
yes it is redundant
| # _ = model_census2.get_latent_representation() | ||
|
|
||
| # takes time | ||
| adata = cellxgene_census.get_anndata( |
There was a problem hiding this comment.
just download 10 cells - see below with obs_value_filter.
| var_coords=hv_idx, | ||
| ) | ||
|
|
||
| # TODO: do we need to put inside (or is it alrady pre-made) - perhaps need to tell CZI |
| adata.obs["batch"] = adata.obs[batch_keys].agg("".join, axis=1).astype("category") | ||
|
|
||
| scvi.model.SCVI.prepare_query_anndata(adata, save_path) | ||
| scvi.model.SCVI.load_query_data(registry=datamodule.registry, reference_model=save_path) |
There was a problem hiding this comment.
We should have more tests that actually fail - using different genes without prepare_query_anndata and different batch categories. Assert that it fails.
|
|
||
| scvi.model.SCVI.prepare_query_anndata(adata, model_census2) | ||
|
|
||
| scvi.model.SCVI.setup_anndata(adata, batch_key="batch") # needed? |
There was a problem hiding this comment.
checking that an AnnData model can be trained using datamodule. Do we really want it?
|
|
||
| user_attributes_model_census3 = model_census3._get_user_attributes() | ||
| pprint(user_attributes_model_census3) | ||
| _ = model_census3.get_elbo() |
| scvi.model.SCVI.prepare_query_anndata(adata, model_census3) | ||
| scvi.model.SCVI.load_query_data(adata, model_census3) | ||
|
|
||
| datamodule_inference = CensusSCVIDataModule( |
There was a problem hiding this comment.
check here that using different genes and different batches fails. You can take much fewer cells here, like 1000.
| # Create a dataloder of a CZI module | ||
| datapipe = datamodule_inference.datapipe | ||
| dataloader = experiment_dataloader(datapipe, num_workers=0, persistent_workers=False) | ||
| mapped_dataloader = ( |
|
|
||
| model = SCVI(adata, n_latent=n_latent) | ||
| model.train(max_epochs=1) | ||
| dataloader = model._make_data_loader(adata) |
There was a problem hiding this comment.
Does model._make_data_loader exist for all models? We should then add the test to the other models as well?
There was a problem hiding this comment.
Is the dataloader sufficient to also setup the model and does setup_datamodule work for it?
No description provided.