Layernaut Studio
Year: 2026
Type: software
Abstract: Added Exported image scripts now verify the class count against the dataset. The tool cannot see your data, so a head built for N classes against a directory containing M produced an opaque shape error inside the loss function minutes into training. The script now stops immediately with the counts, the class names, and what to change. Fixed "Freeze all except last layer" froze every layer with weights. Most backbones end in a pooling layer that has no parameters, so taking the last layer literally made the option identical to freezing the whole backbone. It now leaves the last layer that actually has weights trainable, and reports which one, giving four distinct behaviours: head only, last convolutional block, half the backbone, or full fine-tuning.
Author: Neili, Zakaria
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204434864