44-EFT.WP.Data.ModelCards v1.0 | Chapter 7 Architecture & Parameters

Home ／ Docs-Technical WhitePaper ／ 44-EFT.WP.Data.ModelCards v1.0

Chapter 7 Architecture & Parameters

I. Chapter Purpose & Scope

II. Terminology & Dependencies

• Terminology source: Comprehensive Template v0.1; this chapter only adds fields directly tied to architecture/parameters.
• Dependent volumes: Core.DataSpec v1.0 (contracts/exports); Core.Metrology v1.0 (units & dimensional checks); Core.Equations v1.1 (if path/arrival-time appears). Inline symbols use backticks; any division/integral/composite operator must use parentheses; no Chinese in formulas/symbols/definitions.

III. Fields & Structure (Normative)

architecture:

version: "v1.0"

backbone: "<string>" # e.g., resnet50 | vit-b | conformer-xs | transformer-base

topology:

stages: # ordered module/stage list (order = topology)

- {name:"stem", type:"conv", params:{out:64, k:7, s:2, norm:"bn", act:"relu"}}

- {name:"stage1", type:"resblk", repeat:3, params:{out:256, bottleneck:true}}

- {name:"stage2", type:"resblk", repeat:4, params:{out:512}}

- {name:"head", type:"linear", params:{out_dim:1000}}

positional_encoding: {type:"sinusoidal|learned|none", dim: 768?}

norm: {type:"bn|ln|rmsnorm", eps:1e-5, affine:true}

act: {type:"relu|gelu|silu|tanh"}

dropout: {p: 0.1}

attention: {type:"msa|lsa|flash", heads:12?, window:16?}

mixed_precision: {train:"fp16|bf16|fp32", infer:"fp16|bf16|fp32", loss_scale:"dynamic|static|none"}

init:

scheme: "kaiming_uniform|xavier_normal|trunc_normal"

seed: 1701

params_report:

M_param: 25.6 # million (M)

FLOPs: 4.1e9 # per-sample inference

T_inf: 3.8 # ms/sample (batch=1; record device/driver elsewhere)

constraints:

grad_ckpt: true

amp_safe_ops: ["conv","gemm"]

see:

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

(M_param/FLOPs/T_inf units and posture are validated by the Metrology chapter; any I/O-related shapes must match Chapter 6.)

IV. Parameter Counting & Metrology Posture

• M_param: Count trainable parameters; optimizer state excluded by default. If including frozen parameters or sparsity masks, note it under params_report.notes.
• FLOPs: Use single-sample forward as default; if dependent on sequence length/resolution, provide a function form or nominal points (e.g., 224×224, T=16000).
• T_inf: Record device/batch/framework & kernel versions; unit ms; report a reproducible statistic (e.g., median of 5 runs under fixed env).
• Dimensional consistency: Declare units in Schema and pass check_dim.

V. Module Catalog & Constraints (Common Types)

• Conv/Residual blocks: conv/resblk; params: out,k,s,p,bottleneck; standardize residual connections and downsampling triggers.
• Transformer encoder/decoder: self_attn/cross_attn/ffn; enforce heads * head_dim = model_dim; expose masking/causality under constraints.
• Normalization/Activation: Declare at norm/act top level; layer overrides must be explicit in module params.
• Positional encoding: sinusoidal|learned|rope etc.; align with input length/padding policy.
• Attention impl: msa (standard) | lsa (local) | flash (kernel-optimized); state windowing and memory/complexity improvements.

VI. Initialization, Precision & Device Policy

• init.scheme: Specify weight/bias strategy; for distributed/sharded init, provide pseudorandom graph and seed matrix.
• mixed_precision: Training/inference precision and loss scaling; list unsafe ops allow/deny lists and fallback rules.
• Gradient checkpointing/recompute: Flag via constraints.grad_ckpt; document impact on FLOPs/T_inf.

VII. Regularization & Structured Techniques (linked to Chapter 5 Compression)

• Weight decay/dropout: Declare in optimization / architecture.dropout.
• Structured sparsity/channel pruning: If enabled, put details in compression and reflect topology/shape impacts in this chapter’s constraints.
• Distillation alignment: Ensure compression.distillation.teacher is shape-compatible with architecture.backbone.

VIII. Consistency with Task I/O & Evaluation Protocol

• Shape coherence: Spatial/temporal changes in architecture.topology must align with io_schema (Chapter 6).
• Evaluation reproducibility: List checkpoint and weight hashes used for evaluation in export_manifest.artifacts[]; record seed/repeats in evaluation.

IX. Machine-Readable Fragment (Drop-in)

architecture:

version: "v1.0"

backbone: "vit-b"

topology:

- {name:"patchify", type:"conv", params:{k:16, s:16, out:768}}

- {name:"enc1", type:"transformer_block", repeat:12,

params:{dim:768, heads:12, mlp_ratio:4.0, act:"gelu", norm:"ln"}}

- {name:"head", type:"linear", params:{out_dim:1000}}

positional_encoding: {type:"sinusoidal", dim:768}

mixed_precision: {train:"bf16", infer:"bf16", loss_scale:"dynamic"}

init: {scheme:"trunc_normal", seed:1701}

params_report: {M_param: 86.6, FLOPs: 1.8e10, T_inf: 6.2}

X. Linkage to Path-Dependent Quantities (if applicable)

1. path_dependence.delta_form, path="gamma(ell)", measure="d ell".
2. Two equivalent T_arr expressions:
   • T_arr = ( 1 / c_ref ) * ( ∫ n_eff d ell )
   • T_arr = ( ∫ ( n_eff / c_ref ) d ell )
     and pass check_dim.

XI. Machine-Readable Schema (Excerpt, Normative)

# I15-7 Architecture & Params (excerpt)

properties:

architecture:

type: object

required: [version, backbone, topology]

properties:

version: {type: string}

backbone:{type: string}

topology: {type: array, items:{type: object, properties:{

name:{type:string}, type:{type:string}, repeat:{type:integer},

params:{type:object}}}}

positional_encoding: {type: object}

norm: {type: object}

act: {type: object}

dropout: {type: object}

attention: {type: object}

mixed_precision: {type: object}

init: {type: object, properties:{scheme:{type:string}, seed:{type:integer}}}

params_report: {type: object, properties:{M_param:{type:number}, FLOPs:{type:number}, T_inf:{type:number}}}

constraints: {type: object}

(Declare units at the Schema layer and validate via Core.Metrology v1.0; citations use “Volume vX.Y:Anchor”.)

XII. Chapter Compliance Checklist

• architecture declares backbone/topology/init/precision and parameter report, consistent with Chapter 6 I/O and evaluation scripts.
• M_param/FLOPs/T_inf units & measurement posture are explicit and pass check_dim; device/batch/environment locks ensure reproducibility.
• If path dependence exists, delta_form/path/measure registered and metrology checks passed; all formulas use backticks and parentheses with no Chinese.
• Interfaces to compression/explainability are clear; export artifacts and hashes are listed in export_manifest.