---

name: pix2code description: "Operate the pix2code research repository for GUI screenshot-to-DSL experiments, legacy Keras/TensorFlow training, model sampling, and DSL compilation to web, Android, or iOS code." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

pix2code Repo Skill

Use this skill when a task involves the pix2code research codebase: turning GUI screenshot datasets into training examples, training or inspecting the legacy Keras/TensorFlow model, generating the pix2code DSL from screenshots, or compiling .gui DSL files into web, Android, or iOS scaffold code.

pix2code is an educational 2017 research prototype. Treat it as historical ML code, not a production UI generator. Prefer small, deterministic checks and explicit artifact validation before attempting model training or inference.

Start here

1. Read references/repo-provenance.md before deciding whether this skill matches a checkout or needs a refresh.
2. Read references/package-overview.md for repository layout, dependency constraints, and source facts shared across workflows.
3. Run or adapt scripts/check_pix2code_environment.py when the task starts with setup, dependency diagnosis, or legacy TensorFlow/Keras compatibility.
4. Route to the focused sub-skill below; do not reopen original repo scripts unless the user is explicitly maintaining the checkout.

Install and inspection baseline

The original README listed Python 2 or 3 and these historical pins:

Keras==2.1.2
numpy==1.13.3
opencv-python==3.3.0.10
h5py==2.7.1
tensorflow==1.4.0

For modern hosts, expect to use an isolated legacy Python environment. The old OpenCV pin may no longer be available from current package indexes; use the troubleshooting reference before changing versions. A minimal inspection check is:

python scripts/check_pix2code_environment.py --include-ml

The repository is not a packaged Python distribution with console entry points. Its scripts rely on working from either the model/ or compiler/ source directories. This generated skill therefore bundles safer helpers that are self-contained and runnable from arbitrary directories.

Route map

Task intent	Read this
Prepare pix2code datasets, validate .gui/image pairs, split train/eval sets, convert screenshots to .npz, or plan legacy training commands.	sub-skills/data-and-training/SKILL.md
Validate trained artifacts, choose greedy versus beam search, diagnose missing pix2code.json / .h5 / meta_dataset.npy / words.vocab, or explain batch versus single-image generation.	sub-skills/sampling-and-generation/SKILL.md
Compile a .gui DSL file to web Bootstrap HTML, Android XML, or iOS Storyboard, inspect DSL tokens, or debug malformed DSL nesting.	sub-skills/dsl-compilation/SKILL.md
Diagnose install/import failures, stale helper behavior, unavailable pretrained results, or research-prototype limitations shared by multiple workflows.	references/troubleshooting.md

Operating rules

• Keep user-facing instructions self-contained. Use bundled scripts and references in this skill rather than relying on original checkout scripts.
• Avoid full training unless the user explicitly accepts a long-running legacy TensorFlow/Keras job and has the data unpacked.
• Treat GPU acceleration as optional for this skill. CPU validates dataset handling and DSL compilers; paper-scale training speed or model quality is not verified by CPU checks.
• Validate generated or user-supplied .gui files before compiling them. Unknown tokens raise KeyError in the original compiler; use the bundled compiler's clearer diagnostics.
• Verify trained artifact directories before sampling. pix2code inference requires the JSON architecture, HDF5 weights, NumPy metadata, and vocabulary files to agree.
• Record any dependency substitutions separately from model-quality claims. A successful import of modern or substituted packages does not prove the 2017 paper result.

Common deliverables

• A cleaned or validated pix2code dataset layout with paired .gui and .png/.npz files.
• A command plan for legacy model training that distinguishes in-memory versus generator mode.
• A trained-artifact validation report before sampling screenshots.
• A compiled HTML/XML/Storyboard scaffold from a DSL fixture using bundled compiler logic.
• A troubleshooting note that names whether the blocker is dependency compatibility, missing data, missing trained weights, malformed DSL, or a research-code limitation.