Add code to run smash method

.gitignore

@@ -14,3 +14,5 @@ weights
 NSForest/*.csv
 benchmarks
 figures
+smash-fork/*
+

code/method-pars.R

@@ -142,7 +142,7 @@ method_pars_data <- list(
   random_pars,
   logfc_pars,
   glmgampoi_pars,
-  #smash_pars,
+  smash_pars,
   nsforest_pars,
   cepo_pars,
   cosg_pars

code/run-smash.R

@@ -13,9 +13,15 @@ run_smash <- function(sce, pars) {
   Sys.setenv(RETICULATE_PYTHON = "/home/jpullin/.conda/envs/mage/bin/python")
   library(reticulate)
 
+  old_wd <- getwd()
+
+  # Needed for the import statement below to work
+  smashpy_path <- here::here("smash-fork")
+  setwd(smashpy_path)
+
   # IPython and cuda errors seem spurious.
-  sm <- import("smashpy")
-  sm <- sm$smashpy()
+  smashpy <- py_run_string("from smashpy import *")
+  sm <- smashpy$smashpy()
 
   # Needed for AnnData conversion.
   sc <- import("scanpy")
@@ -50,8 +56,15 @@ run_smash <- function(sce, pars) {
                                             pct1 = 0.75, pct2 = 0.5)
 
     # Inverse PCA to remove unimportant genes
-    adata <- sm$scale_filter_features(adata, n_components = NULL,
+    # Note: For a small number of simulated datasets this function errors,
+    # likely dues to oddness of the simulated data. Therefore, for those
+    # datasets we skip this step.
+    tryCatch({
+      adata <- sm$scale_filter_features(adata, n_components = NULL,
                                       filter_expression = TRUE)
+      },
+      error = function(e) e
+    )
 
     # Run deep neural network to locate optimal markers for classification of
     # cells according to the original user annotations
@@ -92,5 +105,7 @@ run_smash <- function(sce, pars) {
 
   validate_mgs_result(result)
 
+  setwd(old_wd)
+
   list(time = time, result = result, raw_result = raw_cluster_mgs, pars = pars)
 }