目录文档-技术白皮书44-EFT.WP.Data.ModelCards v1.0

第6章 任务与输入输出模式


I. 章节目的与范围

与落盘口径,覆盖输入/输出字段、shape/dtype/range/semantics、批处理与流式、约束与校核规则;确保与评测协议、部署接口及计量章保持一致。 规范性定义固化 task 与 io_schema 的

II. 术语与依赖


III. 任务类型与 task 字段(规范性)

task: "<classification|retrieval|generation|asr|segmentation|detection|timeseries|forecasting|ranking|regression>"

;若为多任务模型,使用数组并在 io_schema 为每个任务提供命名子模式。一致task 必须与 evaluation.protocol、io_schema 的 I/O 契约

IV. io_schema 顶层结构(规范性)

io_schema:

version: "v1.0"

inputs:

- {name:"<string>", shape:"<(…)>", dtype:"<fp32|int8|uint8|…>", range:"<[lo,hi] 或 N/A>", semantics:"<tokenized|waveform|rgb|…>"}

outputs:

- {name:"<string>", shape:"<(…)>", dtype:"<fp32|int8|…>", range:"<[lo,hi]|[0,1]|N/A>", semantics:"<logits|probs|classes|spans|boxes|…>"}

batching: {mode:"<static|dynamic>", max_batch:<int>}

streaming: {enabled:<bool>, chunk_ms:<int?>, lookahead_ms:<int?>}

constraints:

- {type:"shape_compatible", of:["inputs[0]","outputs[0]"]}

- {type:"range", target:"outputs[probs]", rule:"[0,1] & sum==1±1e-6"}

see:

- "EFT.WP.Core.Metrology v1.0:check_dim"

shape/dtype/range 的合法值在 Schema 中固化;分布式/流式场景需在 streaming 显式窗口与延迟字段。

V. 任务特定 I/O 模式(规范示例)

task: "classification"

io_schema:

inputs:

- {name:"image", shape:"(H,W,3)", dtype:"uint8", range:"[0,255]", semantics:"rgb"}

outputs:

- {name:"probs", shape:"(K,)", dtype:"float32", range:"[0,1]", semantics:"softmax"}

batching: {mode:"dynamic", max_batch: 128}

constraints:

- {type:"range", target:"outputs[probs]", rule:"[0,1] & sum==1±1e-6"}

task: "generation"

io_schema:

inputs:

- {name:"tokens_in", shape:"(T_in,)", dtype:"int32", range:"[0,V)", semantics:"tokenized"}

outputs:

- {name:"tokens_out", shape:"(T_out,)", dtype:"int32", range:"[0,V)", semantics:"tokenized"}

- {name:"logprobs", shape:"(T_out,V)", dtype:"float32", range:"(-∞,0]", semantics:"log-softmax"}

streaming: {enabled:true, chunk_ms: 50, lookahead_ms: 0}

task: "asr"

io_schema:

inputs:

- {name:"waveform", shape:"(T,)", dtype:"float32", range:"[-1,1]", semantics:"pcm"}

outputs:

- {name:"text", shape:"()", dtype:"string", range:"N/A", semantics:"utf-8"}

constraints:

- {type:"sampling_rate", target:"inputs[waveform]", rule:"f_samp==16000 Hz"}

task: "detection"

io_schema:

inputs:

- {name:"image", shape:"(H,W,3)", dtype:"uint8", range:"[0,255]", semantics:"rgb"}

outputs:

- {name:"boxes", shape:"(N,4)", dtype:"float32", range:"[0,1]", semantics:"xywh_norm"}

- {name:"scores", shape:"(N,)", dtype:"float32", range:"[0,1]", semantics:"objectness"}

- {name:"labels", shape:"(N,)", dtype:"int32", range:"[0,K)", semantics:"class_id"}

constraints:

- {type:"range", target:"outputs[boxes]", rule:"[0,1]"}

task: "forecasting"

io_schema:

inputs:

- {name:"series", shape:"(T, C)", dtype:"float32", semantics:"zscore"}

- {name:"time_index", shape:"(T,)", dtype:"int64", semantics:"unix_ms"}

outputs:

- {name:"y_hat", shape:"(H, C)", dtype:"float32", semantics:"forecast"}

- {name:"q_hat", shape:"(H, C, Q)", dtype:"float32", semantics:"quantiles"}

constraints:

- {type:"unit", target:"inputs[series]", rule:"SI; check_dim==true"}

(涉及物理量或时间/频率时,单位与量纲由《Core.Metrology v1.0》统一校核,check_dim=true。)


VI. 多任务与多模态 I/O

task: ["classification","retrieval"]

io_schema:

version: "v1.0"

modes:

classification:

inputs: [{name:"image", shape:"(H,W,3)", dtype:"uint8", range:"[0,255]"}]

outputs: [{name:"probs", shape:"(K,)", dtype:"float32", range:"[0,1]", semantics:"softmax"}]

retrieval:

inputs: [{name:"query_emb", shape:"(D,)", dtype:"float32"}]

outputs: [{name:"doc_ids", shape:"(M,)", dtype:"int64"}]

多任务时,evaluation.metrics 需分任务列出;deployment.forms 若路由不同须在第11/第16章接口中体现。

VII. 校核规则与 Lint 约束(规范性)

lint_rules:

- id: IO.RANGE_PROBS

when: "$.io_schema.outputs[?(@.semantics=='softmax')]"

assert: "range == '[0,1]'"

level: error

- id: IO.SHAPE_NONEMPTY

when: "$.io_schema.inputs[*].shape"

assert: "matches('^\\(') and contains(',')"

level: error

- id: IO.DTYPE_ALLOWED

when: "$.io_schema.inputs[*].dtype"

assert: "in_(['float16','float32','int8','int16','int32','uint8','string'])"

level: error

- id: IO.METROLOGY_CHECKDIM

when: "$.io_schema"

assert: "metrology.units=='SI' and metrology.check_dim==true"

level: error

(上述 Lint 与第15章 Schema 配合执行;失败为阻断项。)


VIII. 与评测协议、部署接口的一致性


IX. 计量与路径依赖(如适用)

若模型 I/O 涉及路径依赖量(如 T_arr),需在模型卡登记:delta_form、path="gamma(ell)"、measure="d ell";两种等价表达:

X. 机器可读片段(可直接嵌入)

task: "classification"

io_schema:

version: "v1.0"

inputs: [{name:"image", shape:"(H,W,3)", dtype:"uint8", range:"[0,255]", semantics:"rgb"}]

outputs: [{name:"probs", shape:"(K,)", dtype:"float32", range:"[0,1]", semantics:"softmax"}]

batching: {mode:"dynamic", max_batch: 128}

streaming: {enabled:false}

see:

- "EFT.WP.Core.Metrology v1.0:check_dim"


XI. 本章合规自检


版权与许可(CC BY 4.0)

版权声明:除另有说明外,《能量丝理论》(含文本、图表、插图、符号与公式)的著作权由作者(“屠广林”先生)享有。
许可方式:本作品采用 Creative Commons 署名 4.0 国际许可协议(CC BY 4.0)进行许可;在注明作者与来源的前提下,允许为商业或非商业目的进行复制、转载、节选、改编与再分发。
署名格式(建议):作者:“屠广林”;作品:《能量丝理论》;来源:energyfilament.org;许可证:CC BY 4.0。

首次发布: 2025-11-11|当前版本:v5.1
协议链接:https://creativecommons.org/licenses/by/4.0/