Design and acceptance notes
This is an interactive educational page, not an installation landing page.
The design uses the existing tensorViz white background, dark green mark and
system fonts. The canonical supplied logo is copied unchanged and never redrawn.
The Hugging Face configuration reference
supports a thumbnail URL for social sharing. The README points that field at
the immutable public copy of the supplied logo; the lesson loads its local copy.
Surface and controls
- Header: supplied logo,
tensorViz / shape lab, runnable-example link. - Introduction: one problem statement and an explicit arithmetic-only boundary.
- Main lesson: batch/height/width inputs, three spatial presets and three classifier choices alongside an ordered shape trace. Mismatch/success uses both text and restrained color. Linear weight/transposed dimensions are explicit.
- Reshape counterexample: compare original feature-map rows with the inferred first dimension. Distinguish changed batch, impossible reshape and redundant reshape when the width already matches.
- Conclusion: preserve batch, declare the intended input contract, and separate shape compatibility from numerical correctness. Optional public source and recorded-model links are available without an embedded remote page.
Main content is at most 1,040 pixels wide inside a 1,136-pixel outer container. Below 760 pixels, controls precede results. Below 390 pixels, shape labels and values stack. Interactive targets have native labels and keyboard focus styles; validation/results use status announcements. No animation or external fonts.
Rendered acceptance
The design follows the existing brand directly; no generated concept is used. Automated arithmetic checks do not establish rendered acceptance.
Check the actual page on desktop and at 320-pixel width before publication:
logo proportions, legible controls/shapes, no horizontal
overflow, all presets/modes/reset, empty and invalid fields, keyboard focus,
long error wrapping and the optional links. The exact example must begin with
(2x512 and 128x5), and the unsafe reshape must show two images becoming eight
rows. Record observed results separately from arithmetic tests.