model-construction
Construct Keras segmentation models with Segmentation Models architectures, backbones, preprocessing, and input validation.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- model-construction
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Construct Keras segmentation models with Segmentation Models architectures, backbones, preprocessing, and input validation.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md7.8 KB
- 参考文档references/backbones-and-architectures.md6.2 KB
- 参考文档references/troubleshooting.md7.2 KB
- 脚本scripts/model_constructor_smoke.py7.6 KB
03 SKILL.md 正文
Model Construction
Use this sub-skill when a task asks how to instantiate or diagnose Segmentation Models Keras models: choosing an architecture, selecting a backbone, configuring input_shape, classes, activation, encoder_weights, encoder_freeze, preprocessing, and constructor-specific validation.
Do not use this sub-skill for loss/metric math or full training loops. Route losses and metrics to losses-metrics; route optimizer/fit utilities, set_trainable, and set_regularization to training-utilities.
Operating checklist
-
Select the framework before import. In modern TensorFlow environments prefer TensorFlow Keras:
import os os.environ["SM_FRAMEWORK"] = "tf.keras" # or "keras" import segmentation_models as smsm.set_framework("tf.keras")andsm.set_framework("keras")exist, but objects should be created only after the intended framework has been selected. -
Choose task outputs. For binary segmentation use
classes=1, activation="sigmoid". For mutually exclusive multiclass segmentation use a channel per class (often classes plus background) withactivation="softmax". For independent multi-label masks use multiple channels withactivation="sigmoid". -
Choose architecture/backbone. See
references/backbones-and-architectures.mdfor the supported names and selection guidance. -
Configure preprocessing by backbone. Use
preprocess_input = sm.get_preprocessing(backbone_name)and apply it to image tensors produced for the same backbone. -
Choose input shape safely.
Unet,Linknet, andFPNcan use dynamic spatial dimensions such as(None, None, 3), but actual image height/width should normally be divisible by 32.PSPNetrequires concrete height/width divisible by6 * downsample_factor. -
Avoid accidental network downloads.
encoder_weights="imagenet"may download pretrained encoder weights and expects RGB-compatible input. For smoke tests, offline builds, grayscale, multispectral, or arbitrary channel counts, useencoder_weights=Noneor map non-RGB inputs to 3 channels before the model. -
Smoke-check unusual configs. Use
scripts/model_constructor_smoke.pyfor a safe constructor check; it defaults toencoder_weights=Noneand does not run prediction unless--predictis passed.
Key references
references/api-reference.md— constructor signatures, defaults, framework and preprocessing examples.references/backbones-and-architectures.md— architecture/backbone trade-offs and supported names.references/troubleshooting.md— common constructor/import/shape failures and recoveries.scripts/model_constructor_smoke.py— CLI smoke script for offline model-construction checks.
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).