Pixel Image Preview

Use me when your data is secretly a picture. If your dataset has hundreds of columns named pixel_0, pixel_1, ... pixel_783, those are images hiding in tabular form. I'll take any row you pick, reshape its flat pixel array into a 2D grid, and render it as a heatmap so you can actually see what it looks like. Perfect for verifying MNIST digits, spot-checking CV datasets, or just confirming your image data loaded correctly.

Overview

The Pixel Image Preview plot takes a single row from your dataset, treats a selected set of numerical columns as pixel intensities, reshapes them into a 2D grid, and renders the result as an image heatmap. No separate image file is needed — the image is reconstructed entirely from the column values stored in your dataset.

Best used for:

• Visualizing individual training examples from image datasets stored as flat arrays
• Inspecting MNIST, Fashion-MNIST, CIFAR, or similar benchmark datasets
• Verifying that image data loaded and preprocessed correctly
• Spot-checking labels by seeing what a specific row actually looks like
• Debugging pixel normalization or scaling issues (e.g., values stuck at 0 or 255)
• Exploring any computer vision dataset where images are stored column-per-pixel

Common Use Cases

MNIST and Handwritten Digits

The most common use case. MNIST stores each 28×28 image as 784 columns (pixel_0 through pixel_783). Set Pixel Columns to all 784, Image Width to 28, and use the Greys colorscale to see individual digit images exactly as your model sees them.

Fashion-MNIST and Grayscale Datasets

Identical structure to MNIST — 28×28 = 784 pixel columns. Use Greys colorscale. Useful for confirming that labels match the image (e.g., verifying that row 42 labeled "Sneaker" actually shows a sneaker).

CIFAR Patches and Color Channel Inspection

If your CIFAR data has been flattened and stored per-channel (e.g., 1024 columns for a single 32×32 channel), set Image Width to 32. Use Viridis or Plasma to visualize channel intensity distributions.

Verifying Preprocessing

After normalization, clipping, or augmentation, use Row Index to step through multiple rows and confirm pixel values are in the expected range. If the image looks black or washed out, your pixel values may be outside the expected 0–255 or 0–1 range.

Custom CV Datasets

Any dataset where images were exported as flat CSVs. As long as you know the image dimensions, you can reconstruct any row.

Options

Data Selection

Pixel Columns

Required — Select all columns that contain pixel values.

Choose every column that represents a pixel in the flattened image array. For MNIST, this is pixel_0 through pixel_783 (784 columns). Use the Select All option in the multi-select to quickly add all numerical columns at once, then remove any non-pixel columns (e.g., label, id) if needed.

Accepts: NUMERICAL columns only

Image Row Index

Required — Which row of the dataset to visualize as an image.

Use the slider to select which row (data sample) to display. Row 0 is the first row of your dataset, row 1 is the second, and so on. Slide through different values to browse individual training examples.

Range: 0–10

Default: 0

Reshape Settings

Image Width

Required — The number of pixels across (width) for reshaping the flat array into a 2D grid.

The total number of selected pixel columns divided by this width gives the image height. You must know your image dimensions in advance:

• MNIST / Fashion-MNIST: 784 pixels ÷ 28 width = 28 height → set to 28
• CIFAR-10 (single channel): 1024 pixels ÷ 32 width = 32 height → set to 32
• Custom 64×64: 4096 pixels ÷ 64 width = 64 height → set to 64

If the total number of selected pixel columns is not evenly divisible by Image Width, the reshape will fail. Double-check that you have selected the correct number of pixel columns.

Default: 28

Color Scale

Optional — Color palette used to render pixel intensities as colors.

Options:

• Greys (default) — Black-to-white grayscale. Best for MNIST and any naturally grayscale dataset.
• Viridis — Perceptually uniform blue-to-yellow scale. Useful for general intensity visualization or when you want to highlight mid-range values.
• Plasma — High-contrast purple-to-yellow. Good for false-color visualization or when making posters/reports.
• Hot — Black-red-yellow-white thermal scale. Useful for emphasizing high-intensity regions (e.g., detecting bright spots or saturated pixels).

Default: Greys

Tips

1. Use Select All for large pixel sets: For datasets with 784+ pixel columns, use the multi-select's "Select All" option, then manually deselect the label column and any ID columns.

2. Verify your column count before setting Image Width: If you select 784 columns and enter width=28, the height will be 784/28=28 (valid). If you select 785 columns by accident, the reshape fails — the error will tell you the shapes don't align.

3. Scroll through multiple rows with Row Index: The slider goes from 0 to 10, so you can quickly scan the first 11 examples in your dataset to get a feel for the data distribution and check label alignment.

4. Greys with reversescale for dark-on-light: By default, low pixel values render dark and high values render light (like a typical black-digit-on-white-background). If your dataset is inverted (white digits on black), try the Hot colorscale for a different perspective.

5. If the image looks scrambled: Your Image Width is likely wrong. Try common square roots of your pixel count — for 1024 columns, try width=32; for 4096, try width=64.

6. Check pixel value range when using colorscales: Greys and other colorscales normalize to the min/max of the displayed row. If a single row has all pixels at the same value, the image will appear uniformly grey — this is a sign of a preprocessing issue (e.g., all zeros after incorrect normalization).

7. Not a substitute for full dataset inspection: This plot shows one row at a time. For understanding the overall distribution of your image data, combine it with a Histogram on a specific pixel column to check for class imbalance or data corruption.