def get_bayes_factor_scvi(cset_a,
cset_b,
sampling_n,
cells_sampled,
use_is,
force_batch=None):
subset_a = np.random.choice(cset_a, cells_sampled)
subset_b = np.random.choice(cset_b, cells_sampled)
posterior_a = trainer.create_posterior(trainer.model,
gene_dataset,
indices=subset_a)
posterior_b = trainer.create_posterior(trainer.model,
gene_dataset,
indices=subset_b)
px_scale_a, log_ratios_a, labels_a = posterior_a.differential_expression_stats(
M_sampling=sampling_n, force_batch=force_batch)
px_scale_b, log_ratios_b, labels_b = posterior_b.differential_expression_stats(
M_sampling=sampling_n, force_batch=force_batch)
px_scale = np.concatenate((px_scale_a, px_scale_b), axis=1)
log_ratios = np.concatenate((log_ratios_a, log_ratios_b), axis=1)
labels_de = np.concatenate((0 * labels_a, 0 * labels_b + 1), axis=0)
return get_IS_bayes_factors(px_scale,
log_ratios,
labels_de,
0,
other_cell_idx=1,
importance_sampling=use_is,
permutation=False)
cell_type[1], batch, nsamples).cpu())
px_scale_b = px_scale_b[np.newaxis, :]
return px_scale_a, px_scale_b
px_scale_a0, px_scale_b0 = scanvi_generate_scale(trainer_scanvi,
scanvi_labels,
use_agg_post,
couple_celltypes, 0,
n_cells, n_samples)
labels_de = np.concatenate((np.zeros(
(px_scale_a0.shape[1], )), np.ones((px_scale_b0.shape[1], ))))
px_scale_0 = np.concatenate((px_scale_a0, px_scale_b0), axis=1)
bayes_scanviAB1 = get_IS_bayes_factors(px_scale_0,
None,
labels_de,
0,
other_cell_idx=1,
importance_sampling=False,
permutation=False)
px_scale_a1, px_scale_b1 = scanvi_generate_scale(trainer_scanvi,
scanvi_labels,
use_agg_post,
couple_celltypes, 1,
n_cells, n_samples)
labels_de = np.concatenate((np.zeros(
(px_scale_a0.shape[1], )), np.ones((px_scale_b0.shape[1], ))))
px_scale_1 = np.concatenate((px_scale_a1, px_scale_b1), axis=1)
bayes_scanviAB2 = get_IS_bayes_factors(px_scale_1,
None,
labels_de,