CSV Stats

See what's in a CSV, clean it, and load it into pandas. Per-column statistics, one-click cross-filtering, cleaning steps that never touch the original, and a pd.read_csv call that matches what is really in the file — separator, decimal comma, dates.

Right-click a .csv file in the explorer → Open in CSV Stats.

A band above every column: type, missing, distinct and unique values, then a histogram or the most frequent values.

What's new

• Column details at a glance (0.32). The distribution and the column in row order come first, then the statistics in pairs, then what is worth checking with its fix beside it; the rest is one click away, under More details.
• Correlation heatmap (0.31). Dataset has a Heatmap button: every pair of numeric columns on one grid, red when they rise together, blue when one falls as the other rises, with the seaborn code to copy. Its section titles now line up on the left, in bold, with a small icon, like the column menu.
• Clean what is inside the cells (0.30). Replace text… finds a piece of text in each cell (€, kg) or a pattern written /like this/; Fill up, and Fill empty with… a fixed value, join Fill down. The field you type in unfolds right beside the menu.
• Remove the rows you are looking at (0.30). With a filter on — outliers, a value, a spelling — Remove these N rows takes them out of the cleaned table, as a step you can undo.
• More real files open (0.30). Tab and | separators, and files saved by Excel in Windows-1252 or UTF-16, are read correctly. Tooltips are instant, and clicking again closes a menu or field.
• Spelling, encoding, gaps (0.27). Merge similar spellings (Lyon, Lyons, lyon ) from the column's menu or Clean; fix garbled characters (Café → Café); see where an id or date series has holes; and read a measure's average by category.
• Column details tells what is wrong (0.26). Placeholder values (-999, 1900-01-01) with Treat as empty beside them, the column in row order with its outliers and empty cells, repeated values in an identifier column, advice for pandas (Int64, category…), and related columns in Dataset.
• Column details leads to the fix (0.25.1). Outliers, values that break the column's type and inconsistent formats (A-9999 against A-999) are one click from their rows, with the cleaning step that fixes them right beside.
• Clean, without touching the original (0.25.0). A sparkle icon on every column header (or right-click it) trims, fixes case, replaces values, changes the type, renames or removes the column; the Clean button lists the applied steps, holds the full catalogue — duplicate and blank rows included — and saves a cleaned copy next to the file, name.clean.csv, never over it. Changed cells are underlined; hover one to see what it was.
• Duplicated rows, side by side (0.24.4). Every occurrence, original included, grouped one under the other on alternating shades of the palette; hover a row number to see where the others are.
• Findings in colour (0.24.4). Each finding in Dataset wears its column's colours, so you see at once where to clean.
• Unique values (0.24.3). Next to Distinct, the count of values that occur only once — often typos hidden under Other. Click it to see those rows.
• Dates and times (0.24.3). 2015-01-31, 31/01/2015 and 12:52:01 are recognised: a timeline instead of a ranking, a chronological sort, a before / after / between filter, and parse_dates in the generated pandas code.
• Fixes (0.24.3). A column no longer changes type under a filter, and Read this column as now applies to the filtered rows and survives later filters.
• Headers in colour (0.30). With Dusk and Woodland, and in about half the random palettes, each column name wears its column's colour, adjusted to stay readable.
• Random palettes, much more varied (0.29). Random palette draws coloured text, on a plain background or on one veiled with each column's colour, with or without grid lines, and over the whole colour wheel instead of mostly blues and violets.
• Palettes and headers (0.28). Six palettes — four tone on tone, where one family of colour glides from column to column, and two with distinct colours — and a column header that wears its column's colour: a rule on top, a light veil behind.

Every version is listed in the changelog.

Features

• Statistics per column — type (number, date, time or text), missing, distinct and unique values, histogram, min / max, mean / median.
• Cross-filtering — click a bar or a value: the table keeps those rows and every other band is recomputed on them. Filters show as chips; one click removes each.
• Filter panel per column — a value list with search, like Excel; contains / starts with for text; greater than / between for numbers and dates.
• Cleaning steps — trim, invisible characters, case, replace values or text inside a cell (plain or regular expression), treat tokens as empty, fill down, fill up, fill empty cells with a value, change type, rename, remove a column, remove duplicate, blank or filtered rows (outliers included). Each step shows what it changed and can be removed; the original is never modified, and every save keeps a new copy.
• Column details — the chart button on each band: empty cells, cells of spaces and NULL counted apart, placeholder values, the distribution and the column in row order, outliers, formats, gaps in a series, average by category, what pandas would make of the column and what stands in the way.
• Dataset — the matching pd.read_csv, followed by every rename, drop, filter and sort you made, in order — plus duplicated rows and values that differ only by case or padding, each one click away.
• Correlation heatmap — one click on Heatmap, in Dataset, draws the Pearson correlation of every pair of numeric columns; hover a cell for the coefficient, copy the matching seaborn code.
• Raw view — the file exactly as written, each field in its column's colour.
• 20 palettes, and random ones to draw — hover to preview on your own data.
• Fast — 100,000 rows scroll smoothly; the file itself is never modified.

Right-click a header: the cleaning steps for that column, and the field they need unfolds right beside the menu.

Hover a palette and the sheet changes at once; only a click keeps it.

Fold the bands and every column narrows to its values: twelve columns on screen.

Raw: the unparsed file, each field still in its column's colour.

Good to know

• Comma, semicolon, tab and | files; in a semicolon file the comma is a decimal separator. UTF-8, Windows-1252 (what Excel writes) and UTF-16 are recognised, and the generated pd.read_csv says which.
• Up to 300 MB; beyond 500,000 rows only the first 500,000 are analysed, and the tab says so.
• How every part works, and why: the guide.

Install

Extensions view → … → Install from VSIX…, or:

code --install-extension csv-stats-0.25.8.vsix