data-and-text-conditioning
Prepare Imagen-Pytorch image folders, Hugging Face dataset items, collators, and T5 or precomputed text conditioning inputs.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- data-and-text-conditioning
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Prepare Imagen-Pytorch image folders, Hugging Face dataset items, collators, and T5 or precomputed text conditioning inputs.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/data-formats.md2.3 KB
- 参考文档references/text-conditioning.md1.8 KB
- 参考文档references/troubleshooting.md2.1 KB
- 脚本scripts/check_image_folder.py5.7 KB
- 脚本scripts/text_embedding_shape_check.py5.7 KB
03 SKILL.md 正文
Data and Text Conditioning
Use this sub-skill for the preprocessing path that turns local image collections or Hugging Face rows into Imagen-ready batches and text-conditioning inputs.
Owns
- Local image-folder validation for unconditional training.
Dataset(folder, image_size, exts, convert_image_to_type)andget_images_dataloader(...).Collator(image_size, url_label, text_label, image_label, name, channels)and its URL/image/text row handling.- Raw texts,
text_embeds,text_masks, and image-channel conversion for Imagen / ImagenTrainer. - Shape and mode checks for precomputed text conditioning.
Route elsewhere
- Training loop, optimizers, EMA, checkpoints, and
train_step:../training-and-checkpointing/SKILL.md - Config files and CLI flags:
../configuration-and-cli/SKILL.md - Image generation and sampling APIs:
../image-generation/SKILL.md - Video and inpainting sampling APIs:
../video-and-inpainting/SKILL.md
Short workflow
- Identify the input path: local folder, HF row schema, or precomputed text tensors.
- Check the folder or tensor metadata with the bundled scripts:
- For local folders, use
Dataset(...)orget_images_dataloader(...)to produce image-only batches. - For HF rows, use
Collator(...)to produce image tensors and encoded text tensors. - For text conditioning, choose raw texts or precomputed
text_embeds/text_masksand verify the shapes before training. - Hand the validated batches to the training or sampling sub-skill.
Key contracts
- The local-folder path is image-only.
- The collator path can download URLs, skip failing rows, and return
Noneif every row in a batch fails. - Trainer dataloaders are consumed by keyword order, so tuple layout must match the configured output names.
- T5 text handling truncates at 256 tokens and uses the configured T5 model name.
- Channel conversion must match the requested
channelsvalue and the input image modes.
References
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).