Reshape

Coral.Reshape changes a frame between wide and long layouts. Every function returns a new Frame[m].

Operation	Signature	Result
melt	(df: Frame[n], id_cols: List[string], value_cols: List[string]) -> Frame[m]	Repeats string id columns and places float value columns into variable and value columns
pivot	(df: Frame[n], index_col: string, columns_col: string, values_col: string) -> Frame[m]	Makes one float output column per distinct string value in columns_col
stack	(df: Frame[n]) -> Frame[m]	melt with no id columns and every column as a value column
unstack	(df: Frame[n], index_col: string) -> Frame[m]	pivot with variable as the category column and value as the values column

melt and stack

melt needs string id columns and float value columns. It emits all rows for the first value column, then all rows for the next. The output column order is variable, value, then the id columns. Columns named in neither list are dropped. Do not name an id column variable or value: the id column silently replaces that output column. Melting a string id column named value returns only variable and value, where value now holds the id strings.

module Coral.BookReshape
import Coral.Frame (FloatCol, StringCol, from_pairs, nrows)
import Coral.Reshape (melt)
export (main)
def main() -> i64 = {
  frame = from_pairs([("city", StringCol(["london", "paris", "oslo"])), ("qty", FloatCol(to_tensor([5.0f32, 6.0f32, 7.0f32]))), ("price", FloatCol(to_tensor([10.0f32, 20.0f32, 30.0f32])))])
  melted = melt(frame, ["city"], ["qty", "price"])
  nrows(melted)
}

main returns 6: three source rows for each of two value columns.

stack passes every column as a value column, so every column must be float. A frame with a string or integer column fails with melt: value_cols must be float; use melt and name the id columns instead. Stacking float columns a = [1.0, 2.0] and b = [3.0, 4.0] gives columns variable and value, with values [1.0, 2.0, 3.0, 4.0].

pivot and unstack

pivot needs string index and category columns and a float values column; another type fails with pivot: index_col must be string, pivot: columns_col must be string, or pivot: values_col must be float.

• Layout. The index column comes first, then one float column per distinct category. Rows follow the order in which index values first appear; generated columns follow the order in which categories first appear.
• Gaps and duplicates. A missing index/category combination gives NaN. When several input rows share a combination, the first one wins.
• Other columns are dropped.
• Name clash. A category equal to the index column's name silently replaces the index column with that category's float values. Rename the index column first if a category could share its name.

For an input with day = ["tue", "mon", "tue", "mon", "tue"], field = ["qty", "qty", "price", "price", "qty"], and v = [1.0, 2.0, 3.0, 4.0, 9.0], pivot(long, "day", "field", "v") returns columns day, qty, price with day = ["tue", "mon"], qty = [1.0, 2.0], and price = [3.0, 4.0]. The second tue/qty row (9.0) is ignored.

unstack(frame, index_col) reverses a melt that had one id column: melting city, qty, price on city and then unstacking on city returns columns city, qty, price again.