data-and-tracking
Use DALLE2-pytorch dataloaders, WebDataset and embedding layouts, tracker/log/save/load configuration, and safe data-layout diagnostics for training workflows.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- data-and-tracking
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Use DALLE2-pytorch dataloaders, WebDataset and embedding layouts, tracker/log/save/load configuration, and safe data-layout diagnostics for training workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/data-formats.md4.9 KB
- 参考文档references/tracking-and-checkpoints.md4.0 KB
- 参考文档references/troubleshooting.md3.5 KB
- 脚本scripts/validate_webdataset_layout.py5.6 KB
03 SKILL.md 正文
data-and-tracking
Use this sub-skill when the task is to prepare DALLE2-pytorch training data, validate WebDataset/embedding shard layout, configure ImageEmbeddingDataset or PriorEmbeddingDataset, or set up experiment trackers, checkpoint loading, and checkpoint saving.
Load This When
- Checking decoder WebDataset
.tarshards,.jpg/.npykeys, sidecar image/text embedding folders, shard widths, orindex_width. - Using
create_image_embedding_dataloader,ImageEmbeddingDataset,get_reader,make_splits, orPriorEmbeddingDataset. - Debugging S3/fsspec data URLs,
s3cmd,s3fs, missing embeddings, zero sidecar embeddings, or resampling/shuffling behavior. - Configuring console/W&B logging, local/URL/W&B checkpoint loading, or local/W&B/HuggingFace saving.
- Separating public package workflow guidance from credentials, private data, and tracker artifacts.
Route Elsewhere
- Model architecture, prior/decoder construction,
DALLE2,dream, inpainting, or CLIP adapters:../generation-and-api/SKILL.md. - JSON training configs, trainer APIs, launcher commands, and distributed training:
../training-and-configs/SKILL.md.
Runtime Contract
- Public package:
pip install dalle2-pytorch. - Public data imports:
from dalle2_pytorch.dataloaders import ImageEmbeddingDataset, create_image_embedding_dataloader, get_reader, make_splits, PriorEmbeddingDataset. - Public tracker/config imports:
from dalle2_pytorch.train_configs import TrackerConfig; tracker provider behavior lives behind config classes. - This skill does not provide datasets, credentials, model weights, S3 buckets, W&B runs, or HuggingFace tokens. It helps validate user-provided paths/configs safely.
References And Bundled Script
- Decoder and prior data formats: references/data-formats.md.
- Tracker, loader, saver, and checkpoint destinations: references/tracking-and-checkpoints.md.
- Data/tracker troubleshooting: references/troubleshooting.md.
- Safe layout checker: scripts/validate_webdataset_layout.py.
Typical Flow
-
Read references/data-formats.md to decide whether the task uses decoder WebDataset shards or prior EmbeddingReader folders.
-
For decoder configs, inspect shard and sidecar naming without loading large data:
python scripts/validate_webdataset_layout.py --tar-pattern 'data/shards/{}.tar' --start-shard 0 --end-shard 9 --shard-width 1 --index-width 4 -
Add sidecar folders when embeddings are outside the tar files:
python scripts/validate_webdataset_layout.py --tar-pattern 'data/shards/{}.tar' --start-shard 0 --end-shard 9 --shard-width 1 --index-width 4 --image-embeddings data/img_emb --text-embeddings data/text_emb -
Read references/tracking-and-checkpoints.md before enabling W&B, HuggingFace, S3, URL loaders, or auto-resume.
Safety Defaults
- Prefer local filesystem paths and
log_type: consolefor first checks. - Do not inspect or upload private tracker credentials; require the user to supply and approve them at execution time.
- Treat network, S3, W&B, HuggingFace, and URL loaders/savers as credentialed or networked workflows, not default smoke tests.
- Use
--inspect-taron at most a small shard; otherwise the bundled checker only validates names and file existence.
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).