Trim the comments that were review notes

src/data_processors/process_dataset/script.R

@@ -90,8 +90,7 @@ if (ad2_mod == "ATAC") {
   # subset to make the task computationally feasible
   if (ncol(ad2) > 10000) {
     poss_ix <- which(Matrix::colSums(ad2$layers[["normalized"]]) > 0)
-    # sample by position -- sample(x, n) errors when n > length(x), and treats a
-    # length-1 x as seq_len(x)
+    # sample by position -- sample(x, n) treats a length-1 x as seq_len(x)
     sel_ix <- sort(poss_ix[sample.int(length(poss_ix), min(10000, length(poss_ix)))])
     ad2 <- ad2[, sel_ix]$copy()
     ad2_var <- ad2_var[sel_ix, , drop = FALSE]
@@ -109,9 +108,8 @@ if ("is_train" %in% colnames(ad1$obs)) {
   is_train <- which(ad1$obs[["is_train"]] == "train")
   is_test <- which(ad1$obs[["is_train"]] != "train")
 } else {
-  # No predefined split -- obs['is_train'] carries the NeurIPS 2021 competition
-  # split and other datasets have no reason to have it. Hold out a quarter of the
-  # batches instead, so the test cells come from donors the method has not seen.
+  # no predefined split -- hold out a quarter of the batches, so the test cells
+  # come from donors the method has not seen
   batches <- unique(as.character(ad1$obs[["batch"]]))
   if (length(batches) > 1) {
     test_batches <- sample(batches, max(1, floor(length(batches) / 4)))

src/methods/lm/script.R

@@ -31,7 +31,7 @@ ix <- seq_len(nrow(input_train_mod1))
 dr_train <- dr[ix, , drop = FALSE]
 dr_test <- dr[-ix, , drop = FALSE]
 
-# add an intercept column -- fastLm() uses the design matrix as is
+# add an intercept column
 dr_train <- cbind(intercept = 1, dr_train)
 dr_test <- cbind(intercept = 1, dr_test)
 

src/methods/novel/novel_predict/script.py

@@ -95,8 +95,6 @@
 dataloader_test = DataLoader(dataset_test, 32, shuffle = False, num_workers = 4)
 
 outputs = []
-# the weights are loaded with map_location='cpu', so move the model onto the device
-# we selected above -- otherwise this component requests a GPU and never uses it
 model = model.to(device)
 model.eval()
 with torch.no_grad():

src/metrics/correlation/script.R

@@ -39,9 +39,7 @@ pv_sd2 <- proxyC::colSds(pv)
 tv_sd1 <- proxyC::rowSds(tv)
 pv_sd1 <- proxyC::rowSds(pv)
 
-# correlate matching rows (margin 1) or columns (margin 2) of two matrices.
-# dynutils::calculate_similarity() would do this too, but it ends in as.matrix(),
-# densifying an n x n / p x p matrix just so we can read its diagonal.
+# correlate matching rows (margin 1) or columns (margin 2) of two matrices
 paired_similarity <- function(x, y, margin, method) {
   if (margin == 1) {
     x <- Matrix::t(x)
@@ -87,8 +85,7 @@ mean_spearman_per_gene <- mean(spearman_vec_2)
 tv_vec <- as.vector(tv)
 pv_vec <- as.vector(pv)
 
-# a constant prediction has no correlation to speak of -- score it as 0, the same
-# substitution the per-cell and per-gene metrics make above. cor() would return NA.
+# zero variance -- score as 0, like the per-cell and per-gene metrics above
 if (sd(tv_vec) == 0 || sd(pv_vec) == 0) {
   overall_pearson <- 0
   overall_spearman <- 0

src/metrics/mse/script.py

@@ -26,8 +26,7 @@
 
 logging.info("Computing MSE metrics")
 
-# coerce to sparse -- a method is free to return a dense layer, and
-# sparse - dense yields a np.matrix, which has no .power()
+# coerce to sparse -- sparse minus dense yields a np.matrix, which has no .power()
 tmp = csr_matrix(ad_sol.layers["normalized"]) - csr_matrix(ad_pred.layers["normalized"])
 rmse = np.sqrt(tmp.power(2).mean())
 mae = np.abs(tmp).mean()