Error when batch_size[1] <= number of genes

[le-ander]

when I run batchglm through diffxpy (diffxpy.test.wald()) and my dataset has less or equal the number of features as I have set as the second dimension of the batch size, I get an error.

So for example using data with 255 (or 256) genes and a batch_size of (1e9, 256), I get the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-29-2450422e51e3> in <module>
     11 
     12 with dask.config.set(**{'array.slicing.split_large_chunks': True}):
---> 13     dxp_hsp_d70 = de.test.wald(
     14                     data = adata_de,
     15                     dmat_loc = patsy.dmatrix(form, adata_de.obs),

~/.local/lib/python3.8/site-packages/diffxpy/testing/tests.py in wald(data, factor_loc_totest, coef_to_test, formula_loc, formula_scale, as_numeric, init_a, init_b, gene_names, sample_description, dmat_loc, dmat_scale, constraints_loc, constraints_scale, noise_model, size_factors, batch_size, backend, train_args, training_strategy, quick_scale, dtype, **kwargs)
    722 
    723     # Fit model.
--> 724     model = _fit(
    725         noise_model=noise_model,
    726         data=data,

~/.local/lib/python3.8/site-packages/diffxpy/testing/tests.py in _fit(noise_model, data, design_loc, design_scale, design_loc_names, design_scale_names, constraints_loc, constraints_scale, init_model, init_a, init_b, gene_names, size_factors, batch_size, backend, training_strategy, quick_scale, train_args, close_session, dtype)
    242         pass
    243 
--> 244     estim.train_sequence(
    245         training_strategy=training_strategy,
    246         **train_args

~/.local/lib/python3.8/site-packages/batchglm/models/base/estimator.py in train_sequence(self, training_strategy, **kwargs)
    122                         (x, str(d[x]), str(kwargs[x]))
    123                     )
--> 124             self.train(**d, **kwargs)
    125             logger.debug("Training sequence #%d complete", idx + 1)
    126 

~/.local/lib/python3.8/site-packages/batchglm/train/numpy/base_glm/estimator.py in train(self, max_steps, method_b, update_b_freq, ftol_b, lr_b, max_iter_b, nproc, **kwargs)
    137                 idx_update = self.model.idx_not_converged
    138                 if self._train_loc:
--> 139                     a_step = self.iwls_step(idx_update=idx_update)
    140                     # Perform trial update.
    141                     self.model.a_var = self.model.a_var + a_step

~/.local/lib/python3.8/site-packages/batchglm/train/numpy/base_glm/estimator.py in iwls_step(self, idx_update)
    299             else:
    300                 if np.linalg.cond(a.compute(), p=None) < 1 / sys.float_info.epsilon:
--> 301                     delta_theta[:, idx_update] = np.expand_dims(
    302                         np.linalg.solve(a[0], b[0]).compute(),
    303                         axis=-1

ValueError: assignment destination is read-only

[ilan-gold]

@le-ander Sorry to just get to this, I was taking some time off. I understand if it's out of your scope to make a smaller reproducible example, but could you highlight what makes you think that this difference in size is the issue? I saw this type of error while fixing the unit tests a few months ago so hopefully this goes away with the upcoming release.

[le-ander]

Hey there! :) I tested a couple of different scenarios back then:

• 2 genes, batch_size=(1e9, 256): error
• 255 genes, batch_size=(1e9, 256): error
• 256 genes, batch_size=(1e9, 256): error
• 257 genes, batch_size=(1e9, 256): no error
• 300 genes, batch_size=(1e9, 256): no error

So I figured that the batch_size <= number of features must be the culprit here.