From 06a66462d1422cb3dae12f773fb903085d41bd86 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=BC=A0=E6=96=8C?= <259278618@qq.com>
Date: Tue, 1 Sep 2026 17:08:45 +0800
Subject: [PATCH] 1

---
 .../com/example/observer/CameraChannel.kt     |   5 +++--
 .../lib/detection/tflite_detector.dart        |   8 ++++----
 server/README.md                              |   2 +-
 server/biz/dao/model_training.go              |   2 +-
 server/biz/service/training.go                |   1 +
 server/common/training_runner.go              |   6 ++++--
 server/config.yml                             |   7 ++++---
 server/data/observer.db                       | Bin 425984 -> 425984 bytes
 server/training/train_server.py               |   9 +++++----
 server/技术设计.md                            |   8 ++++++--
 10 files changed, 29 insertions(+), 19 deletions(-)

diff --git a/flutter_app/android/app/src/main/kotlin/com/example/observer/CameraChannel.kt b/flutter_app/android/app/src/main/kotlin/com/example/observer/CameraChannel.kt
index 2f9fcec..d694df0 100644
--- a/flutter_app/android/app/src/main/kotlin/com/example/observer/CameraChannel.kt
+++ b/flutter_app/android/app/src/main/kotlin/com/example/observer/CameraChannel.kt
@@ -362,8 +362,9 @@ class CameraChannel(
 
     /**
      * 从设备流配置挑预览/分析共用分辨率（均为传感器方向尺寸，宽≥高）：
-     * 取两者交集，16:9 优先、长边 ≤1920 内取最大（推理输入 640x640，
-     * 更高只增帧传输与旋转开销，无精度收益）；无 16:9 时退回最大交集尺寸。
+     * 取两者交集，16:9 优先、长边 ≤1920 内取最大（推理输入长边 1280，
+     * 1920x1080 已留 1.5 倍降采样余量，更高只增帧传输与旋转开销，无精度收益）；
+     * 无 16:9 时退回最大交集尺寸。
      */
     private fun pickSize(previewSizes: Array<Size>, analysisSizes: Array<Size>): Size {
         val common = previewSizes.filter { analysisSizes.contains(it) }
diff --git a/flutter_app/lib/detection/tflite_detector.dart b/flutter_app/lib/detection/tflite_detector.dart
index b4ecaf8..7f20b5e 100644
--- a/flutter_app/lib/detection/tflite_detector.dart
+++ b/flutter_app/lib/detection/tflite_detector.dart
@@ -6,14 +6,14 @@ import 'package:tflite_flutter/tflite_flutter.dart';
 import 'detection_result.dart';
 import 'nms.dart';
 
-/// YOLOv8n 端侧推理实现（对应 Kotlin TFLiteDetector）。
+/// YOLOv8s 端侧推理实现（对应 Kotlin TFLiteDetector）。
 /// 模型输出布局（ultralytics litert 导出）：[1, 4 + nc, anchors]，
 /// cx/cy/w/h 已归一化，类别得分已过 sigmoid；按 out[dim][anchor] 索引。
-/// 输入为 NCHW [1, 3, 704, 704]（litert 导出保留 torch 布局）。
+/// 输入为 NCHW [1, 3, 1280, 1280]（litert 导出保留 torch 布局）。
 class TfliteDetector {
   // 输入尺寸取自模型本身（ultralytics litert 导出 NCHW [1,3,H,W]，各数据集
-  // 训练 imgsz 可不同），默认 704 兜底
-  static const int defaultInputSize = 704;
+  // 训练 imgsz 可不同），默认 1280 兜底
+  static const int defaultInputSize = 1280;
   // 环颈雉鸡数据置信度普遍偏低（0.1~0.2 量级），保留低分池供运动检测提升；
   // 可运行时调整（设置页滑块），默认 0.10
   double minScore = 0.10;
diff --git a/server/README.md b/server/README.md
index 2d5d46a..4b48c68 100644
--- a/server/README.md
+++ b/server/README.md
@@ -200,7 +200,7 @@ APK 下载引导页（静态页面，源码在 `h5/index.html`，由后端 `/dow
 | GET | `/admin/label-tasks` | 标注任务列表：`page/size` 分页（含 status/total/done） |
 | GET | `/admin/label-tasks/detail` | 标注任务详情：返回数据集全部图片 + 每张标注框（`boxes`，YOLO 归一化 xywh + 置信度 + 类别） |
 | POST | `/admin/label-tasks/save` | 保存单张标注 `{"datasetId":1,"filename":"a.jpg","boxes":[{"class":0,"cx":0.5,"cy":0.4,"w":0.1,"h":0.2}]}`：整体覆写该图 `labels_json`（空 boxes=清空标注），返回该数据集当前 `labeledCount` |
-| POST | `/admin/trainings` | 发起训练 `{"dataset":"yolo","imgsz":704,"epochs":150,"batch":16,"device":"0","name":"..."}`：先同步数据集到训练机 → 校验目录存在 → runner 启动训练；**并发度 1**，已有 running 任务时返回错误 |
+| POST | `/admin/trainings` | 发起训练 `{"dataset":"yolo","imgsz":1280,"epochs":150,"batch":16,"device":"0","name":"..."}`（imgsz/model 实际以 `config.yml` training 节点为准）：先同步数据集到训练机 → 校验目录存在 → runner 启动训练；**并发度 1**，已有 running 任务时返回错误 |
 | GET | `/admin/trainings` | 训练任务列表：`page/size` 分页，按下发时间倒序，含 status/进度/指标 |
 | GET | `/admin/trainings/detail` | 任务详情 `{"id":1}`：参数快照 + 进度 + 指标 + 日志尾部 |
 | POST | `/admin/trainings/cancel` | 取消训练 `{"id":1}`（仅 running）：杀训练进程，状态置 failed（记录 error） |
diff --git a/server/biz/dao/model_training.go b/server/biz/dao/model_training.go
index 6832fc7..e08a7f7 100644
--- a/server/biz/dao/model_training.go
+++ b/server/biz/dao/model_training.go
@@ -24,7 +24,7 @@ func init() {
   name          TEXT NOT NULL,
   status        TEXT NOT NULL DEFAULT 'running',
   dataset_id    INTEGER NOT NULL,
-  imgsz         INTEGER NOT NULL DEFAULT 704,
+  imgsz         INTEGER NOT NULL DEFAULT 1280,
   epochs        INTEGER NOT NULL DEFAULT 150,
   batch         INTEGER NOT NULL DEFAULT 16,
   device        TEXT NOT NULL DEFAULT '0',
diff --git a/server/biz/service/training.go b/server/biz/service/training.go
index dbd14f7..1720b36 100644
--- a/server/biz/service/training.go
+++ b/server/biz/service/training.go
@@ -425,6 +425,7 @@ func (s *trainingService) AdminStartTraining(ctx context.Context, req *dto.Admin
 		taskJSON, _ := json.Marshal(map[string]any{
 			"workdir":     cfg.Workdir,
 			"yolo":        filepath.ToSlash(filepath.Join(cfg.DatasetDir, "yolo", dataset.Name)),
+			"model":       cfg.Model,
 			"imgsz":       imgsz,
 			"epochs":      epochs,
 			"batch":       batch,
diff --git a/server/common/training_runner.go b/server/common/training_runner.go
index 1c0c24e..7a311eb 100644
--- a/server/common/training_runner.go
+++ b/server/common/training_runner.go
@@ -64,7 +64,7 @@ type TrainingJob struct {
 	TaskId      int64
 	DatasetName string // yolo 数据集目录名（训练机 datasetDir/yolo/<name>）
 	Python      string // 训练机 venv python 路径
-	Workdir     string // 训练机工作目录（train_server.py / yolov8n.pt 所在）
+	Workdir     string // 训练机工作目录（train_server.py / yolov8s.pt 所在）
 	DatasetDir  string // 训练机数据集根目录（相对 workdir）
 }
 
@@ -87,6 +87,7 @@ type TrainingConfig struct {
 	DatasetDir  string
 	Python      string
 	TimeoutMins int
+	Model       string // 训练基座权重（训练机 workdir 下，如 yolov8s.pt）
 	Imgsz       int
 	Epochs      int
 	Batch       int
@@ -100,7 +101,8 @@ func TrainingConfigOf(ctx context.Context) (TrainingConfig, bool) {
 		DatasetDir:  g.Cfg().MustGet(ctx, "training.datasetDir", "datasets").String(),
 		Python:      g.Cfg().MustGet(ctx, "training.venvPython").String(),
 		TimeoutMins: g.Cfg().MustGet(ctx, "training.timeoutMinutes", 600).Int(),
-		Imgsz:       g.Cfg().MustGet(ctx, "training.imgsz", 704).Int(),
+		Model:       g.Cfg().MustGet(ctx, "training.model", "yolov8s.pt").String(),
+		Imgsz:       g.Cfg().MustGet(ctx, "training.imgsz", 1280).Int(),
 		Epochs:      g.Cfg().MustGet(ctx, "training.epochs", 150).Int(),
 		Batch:       g.Cfg().MustGet(ctx, "training.batch", 16).Int(),
 		Device:      g.Cfg().MustGet(ctx, "training.device", "0").String(),
diff --git a/server/config.yml b/server/config.yml
index d9ca12d..e04ee80 100644
--- a/server/config.yml
+++ b/server/config.yml
@@ -70,11 +70,12 @@ training:
     port: 22
     privateKeyPath: ""  # 私钥路径与 password 二选一
     password: "123"
-  workdir: /opt/pheasant_data   # 训练机工作目录（train_server.py / yolov8n.pt 所在）
+  workdir: /opt/pheasant_data   # 训练机工作目录（train_server.py / yolov8s.pt 所在）
   venvPython: /opt/pheasant_data/venv/bin/python
   datasetDir: datasets  # 训练机数据集根目录（相对 workdir，yolo/<name> 为子目录）
   timeoutMinutes: 240   # 训练超时判死（started_at 起算；实测 100 张约 3 分钟，千图级约 1 小时）
-  imgsz: 704            # 训练/导出分辨率（须与 App 端推理输入对齐）
+  model: yolov8s.pt     # 训练基座权重（须已同步到训练机 workdir；s 容量适配小目标+遮挡，详见技术设计）
+  imgsz: 1280           # 训练/导出分辨率（须与 App 端推理输入对齐；1280 提升远距离小目标保留像素）
   epochs: 150           # 训练轮数（patience 30 早停，设大可自动停）
   batch: 16             # 批大小（按训练机显存调整）
   device: "0"           # GPU 编号（cpu 用 cpu）
@@ -87,7 +88,7 @@ localAi:
   threshold: 0.08       # 候选置信度阈值（宁多勿漏）
   confConfirmed: 0.2    # 高于此视为确认(class 0)，否则疑似(class 1)
   overlapThreshold: 0.3  # 重叠去重阈值（交叠/较小框面积，NMS 风格；RF-DETR 同目标常输出一大一小两框，minIoU 比 IoU 更能命中）
-  inputSize: 700        # 提交前整图等比缩放最长边
+  inputSize: 700        # 提交前整图等比缩放最长边；RF-DETR 模型包固定 700x700 输入（stretch 拉伸、不支持动态尺寸，见模型包 inference_config.json），提交更大最终也拉伸到 700 无增益，2026-09-01 实测回退
   # 标注类别名（写入 data.yaml，随 result.json 存模型 labels）从数据集表取：
   # 第一类 = gen_species（空回退数据集名），第二类 = gen_classes（空回退 class0/class1）
 
diff --git a/server/data/observer.db b/server/data/observer.db
index cd4b5a6d4899c61900772c5fc475c046ca8e8eab..ef0418cdfcc6faa7685ba081e1a0623321c4cca9 100644
GIT binary patch
delta 11599
zcmaKyeUM#sdB@NBoqPB0-t4`*31suSdE+gG5YGEc$U*{<XmpWa*xf`R8-Yd=5<>tn
zfj|<tCxKBxF&uSp8b^v#Er_LZ3Z2>@3awS7o!Zi=(+Nzq!cY^d<7h|PexCEY=j`%N
zjlkyaz2|v;&+~o0&-ZzL_u#?#2M^AFc)nV6^=B4d#=j3uerwWv{C_8}?Kx!(ygl&R
zz;_3p9(Ziv=)m5AYX`0vum|S1f6@L<`%imLwI6LCY;S5`(q7)4)%tzwXRTLT&$k|L
zeWLY|)(2W)Ye}ow|LgvL>;IenKkI+2e^39G{&oE;`{(wn8E?<{rx_=X?NMi4)7$r7
zeLv}Ysqg8&FZSKvx4Unm@3KCl_pRQ4?7h494|}ibUDbPbZ@cl^#v6?vG@fld*tov&
z-iF&)*y#EH{yejE$r(M_sUFo)jpNzJ8|wU8?`Msn?D2;BMXhhhxGa0LR~?++Tx6_J
zMrY5>+i&{V?yXU<a>t(SmTwrm)9&M)GrndFJ@djFN6lJ)%4<)}P_NZmpEic_2m4jc
zthIL-LytH=d#%=fjo<EWs~xq08bAG{t(Mn{r}Pe}-?i#LHY|0$(YR7w-@GJi52}Z1
zy|1We^DhsozcSnDZD}~eQ0L8jU|{sVeR*S8-L0+(+}N{S-*#-<O`LGWvK#g-TV<~d
z{LuDdCw3h-@PcH;vODEf+qXkIa(Qd)`JT(mA1z+?LMIOVh&S3kZwWK!$@+rqx}TbT
zn`1u;{2&hG{rroMXaimpMqcDc4wjw3bt?;@<0U*N3?d#HdVXaCmprUHc!m8i+xEP<
zpw|C8b6@^1v(+8uoX#z2P#o{*ouh}XzB1U^H=4DL+2mgH{n;O#QSZq=J*2*8jt6lZ
zxrrY{iHC#Lu@m2wlQ^yu1hy^VD9-J0rT`~x&v)cFysi%(297V03PV41y~Goz7Pg`^
z(vHj+JzUSf^h@<|Gh5tWdm=mWgtcgspG1-Cg}xp8Zo+3dVK*e0P5jsuAEKZ#%_|aC
zCyHZCRWZjdS&MvEUe8{CsJ>KLjr^U3s%8z;E;8!I2aMLG&8T75qHJ(dt!^LLnx3cK
z+Bw>}BY)pgwNq_&k}x7;ap=b(|MN;;2v6wZCdPxvb$GoC06%hLCyax{wZkMza8pOZ
z4LDXnDg{hl9Aq=Uq9M@Sl<+F(@Dm4T-O2=LlMjUf|EV5t`yh+Y+m1XgntHlp#|e=s
zg5U;4_UCJaMZTW|t{o(f8`)7fl}pzp)O^fN+`x~W$_B`a$)@8ZcHj}g$_OD1<Po0B
zFR=lN7Y;R+8@&>?&aU*_3SRYm<ZN}T8uPp)CVu3HqlCDtGa2)Dm$dO4!3VEv7`b-h
z^C#i`p+ojc%Tb*8e&8ie5P`i!Vz+Rpxw30`e|k<~n6f?G$iwp}+i^dR?I84QA|CN?
zIYe2wcHsIRxFl`9L_?dVn34$8g-()4GD_0~!wo@KjNKr~*8bF7IQP0%V)PXzJNKk#
zYumF&5BKG(&sPW4W^4kM*e4|t$kMW(zze=lCgMmkUYfRjCyZ=6@;t{*;vlL{O^`h=
zN*vpTIeg-tEj?f@n%OhM=!~Ri=~&N^3R-`mTC2t>dwh&s0a++d=^i)`P(f2MhvsqB
zf$<X46uReluuys65dH|rb4<R1C`t+*nboQtGHT|FM)RWDXUrEHS2wPP*#E>FYHUz1
zi6@;^=>^)Onc~skI%-6X2cgdaq%7eu->cFeIE2{t?8I{;Qc(uN1ru=siUXVKqSQ;5
zVi#VMa<db-$j;2~TEoNV7=u*^X2WJ9>G{>)>-j^zTBF8cB~l3~If0+VaRvR*p_Bu1
zYD@5?<O&eLYGD|~RA(R&^C4YbywHs=gcl0Md9gUKu-*A!x<XsLSdKgqsjIPA^aGnS
zgQYmEk`EmHIIuZz;y7?+nT5!Q<)CzaBVC}f7)AKSTd5l?$3>}~4eXkQ#Rt-!^bhn;
z=Yjc(%eRJpw(Tl)X?E`s_4(}dsCp(lbxeIKJ2*>i$d0d83&*5I_z`_2aKk970uoUz
zj!nOT44{ZIBS4ii$Fu>Wn1JXqAfD@xkpw1y*ZnxVb-X@IpH(ksBO~f?>;92+c~d&+
z(V3&IeEo>>6$rWn1Z^`~a~r+2_nF(wZ6N4<=9%iJa*)pEbe<08s1)(`wdz_mVdG)s
z;1llJ(gHQ230VlO&`rYtpDIK{xDNCHXW8Uc4xzILxA`!&5FkB>a2JrC*BmrvXp?g~
z7pHT(KaL(Y^AD^8JlDWO9)%oHg^<G(frM}(^__%L!%@jyP$lVeJs2;P>W)k6F3!ie
zkdR~BDX1&&U9WmA;$`(_XKqkw>;Bu)Asw#+#B0$8wL{JB>`Z6lQ|HFwKZJ7Y=;0ap
z;!$;z+5*1hq?8JJ!>a;=UlFHID3@4hukwxBgT;6~N>DNIUg;d6B7+hX-pw1LZnZZ)
zZq&^SjMhMNUE^N!g4(yp)jYF(q$eFN;`G3UqxYNny_?i0)n<>%AP)FKD*<0+ngqQ(
zodyDcp#Lu9sW^HX*`NRu$hPuDnh@PSroa-Uf?fDh^swdx^~<Xc+zAib(^@2iKtGBk
zx-fca0s^54qmx8S5R~$`ZV))4Rlqe>1-P^x8A%sa*{|o%ZY28?f-oKtg+kagz4BBj
zJ}-=;7~+&VuBZbB4`}W{lhUB4bxjfi^6tPX($3++BEZX<BV_|wl`hn_g*T@PImng{
zTWj*;SE*hzTk@QFIJ+vTZ^$MevlfkEmG%O4hIEF4GrCM4Lj68Kgi@g~5U5)lh_~^Y
ze$L1Ev5W>p4-ZNvptT^gpR5nhZq<#3pl&2xFfAnS-%_wCL3SxDDuP5ORfiEaai|o;
zDu)y}YP)n0=+~Bl6ir$sIu>?>Ia~s4hfcQSfVDX5AL;3xyDOcqt)ft;1;fw&eyz18
zTXRKyGXM5<$kVYXLY6voCyz|h9LnqjHcE<M$I?N<xXg|%;tG9)lcIzqG+k;C<RU{v
ztD!3n{oEXesE$HZvy8@r&G*vmW|_0TvUCa4!QOPBa89O!JM%4*G9A#80HBA=mDo#j
z)v1VGn*=c+Akz4!8xhg1^dtHk9l!8edqIqN6{2;$I7$k6qo;!j`VRa@$wJUlrsea=
zF*t@3u|u6zA*Vt@FwekXqnD;gh_@q>p?(AgeMqBlPGf(G!kf~z4*X5RzLPhokE(Hx
zV*!8E3X&qJLQGF5j4@ILyPmvWWE))22!xHGJ33dDZCp&GKw%)%PwW+~?EF@L*}lQO
zwC`1EmyF+C2cDmOc^$bw{kXL!TN2aIjyGygXD4q{i^mW)5z_#1=djaSEy$2`i^@vq
z<7Mz&f{gHRz`O0y;L8ey5q;)0kC=v#c(yUG54YB&Egg_y>FrnUP?xK5<}nejh%#sa
ziYk+VG!)wRTo3(1mMZ*1FhD&<zJP+T1RW}m6jWWrEdnCV^Vj*{F7-<_v$hZN{kGA#
zt9H1xvKgA+hJ2qeht-SXQ71~z)Sk)kR1rpYD5@XH|L1;ng_*5+Os&r*KWPn5AaSCI
z5+e{X8l!>_*g%uJlqTmc7@^A)kcR*YgrZ#i(qsODgml3-XQ8Nd1D(^@Rl;X`I#c_!
zQ=)a`b|$B5C;-3%BT)LFx(pVhD@q93mySfdO8A7Rerga>gPJLkvO<{y;97(k%{#WU
znb*~a7q-qax^N{1bBIV+2^QnzU3aQgY65mA{UjVaK_tyb7bZqU(UWuBqaw;&dGt~o
zC>wqW4#sI_Ef00hVR%}%VceXaQJuFhzjQz6od8i3AHkM+efk)O=LtfAu}x;e5{|;7
zkgr2n`C#Hz7F}XP*Mg`RBI$PpJhGM7)qBnmL3MHG%5;`Cj_kB@=YYsg+5p`#_8C+-
z^fW$%PSDgMaw%c*uWOYE&`>#&7h^?~l_nzUQ~<yX&-h20ie~lb36?Ik+6JSxp*7fS
zG?oI#Rke2KL!ATKkF~-X7k^5PsR<tGpnGM|M4(l8Ln6>lWf?$8kloS`kAMZR8zU}x
z<^ZK5NCb2Y%MSzupy-DEBdMt^B83|Hy@!xO6JQfiA{%IMQBsKy;)?iZki=~4MO7o<
zrHH|hUvW-v8cT_$3sBpPV!+()YQ(hOxwkO+06KfB+P1pc+RN)}a%A=hQbAY`K9=^+
za;CzXpeZ2u%%75jlBA(S$O2eXWiaB@yceJw5NKeQ=>&;eh?BXubnelIzLg2Hs&_mn
z1j54)s%2^mbPOAd1f`ixn>r{oW*HVwK#)C#2%#<jJj5Q6XSjsSoT}P*8x%&p$gTj~
zy9EMhiSMT+jvKAzwXZjJn&aj;7`Viot==hI>+I>=seL>k^?&VW)a{I;E;0^~Gdi=~
znra^;gouE6BrMo03lK_{1qMt`$AKA2B#<g8RM5yyi=PXJYRxfYGW+1?)V1?Et2?)A
z+l<0fTdl255ufbjI;$r;>8s1~OTMF?GRN51NpXP`iqD;rFxZrq&G8v9WzH$$B*h6@
zSmUrk07_+?(D&eyp<JdsBt(&^*7nXF+LM9>Z_QBg6YxCRSCkgCgODtF$P0z*JXngE
zfu~oAfvHiD!iH$!07tr&Df=Esh+l%vD3(>guu}h!t=*!|%y)lLJ+IoGO`Y4c1M7tu
zR(wfqR%4VL?G6URR<W?_YDL5#H9)-K7>TxS0I(e=f+Im7Ggo<D1~AY?14sDcY&Q!+
z40-LidS~&>W-l%4T%$Q)p4<PM+6|5Of{!!Jd1=zwS0?Aa15@C2vW-t#M>A5gLZ;MB
z*~7=w*6hS9){g9|m(-10$Wems0L~zS%}5zFbTl+ep}DLw7-OrlhMSRithrzmK3nk!
z0tMzFi{d)ZDl(<_V&fv7Xpr<HW?b}iX@ob1R4bAe4x>s+EMN|ifMMozjC3=D@S|R3
zCrh{>mx4GJwzhTd(g%mC>iMGYs~gn>;sKe*D2-@@?kcW=G?+$2piy=(b=M$gfvcbi
z>Cv+IEG?p%Av;nMW0eDI^sVmPU51+EfBl6*u@f(kkqpL>5t^gu?6cz_^&-dUgbaYC
zMRcYtBTx#!4{uA0I2Q9dAAzMI1=&N*D&~Y6I-jV{*_YcdL&=-j9WhuCbBYXQR2t=H
zxClpBbjwzwYm^loBZ#bCpkJ9$=nfk52o5v~gv5FZ&E#lKZl^t+{e?~8<VODcAE=|k
z!E8*REv6ijH#HI?J;^eV0lAC?<w<3oBqa$tkbcFi)UC6Gr3OPZ@<<v*P|WMkzN$<v
zPzDWEyVYo{sNLEe%-g3`T}JrYynN2D)E2Xi3~_ZR3K_EWS0Y2`CMYnLKAe%DRtW&x
zpif3|8A;vJEx3SnG8-WqD3h2f<;0z7i7^@uXZX`8pF5d4Rq&@#U!<GUMZZz=WCg~K
zkGAGXbl5JI*?`R8U>Q+}0Ij)1Iq`#EWX{7a*=rYZfqpp(+(YU>BzpqkpZ@ykG7*RF
z%764*wbsl|T%+E|mYlL~%AWYW+AXs7?R@S#>P=+rAJaqLG-^*Z=eG>=P4i7S;#GRc
z_ljq9;`EX#e2x4UrggEJVC<IyK>Ue>tfGe;WF6(>K63>VK!qgKBl8Cd^r>}G{#_t}
zHhLh4mNh{V>L#+Rajev@H>4M9+l)Wc0vQ*9%uc>+HL@?fW!)=W_2=0dkCyVI1=e3@
zll|84INc1BERJyo^-s#WHO77lX~I$#%Jj<k$}S0#khM6AR_MR<5%j_;Jc)NOTNE`0
z!FTd6G_1X9m^&;zeaaeWPG9FE>1yq9A%*5Ot(#c37mX6FA*!QlZS3mh8UYT0z?hkF
zP-_63kP9i<@iAgfNg-4VYC?nqv=Jw=+gsGSI~^(P$&P(S-e_5Os}Dnhf)y!ThTy5`
zg5lO<#O2`_W)oAQoA!qj5H36_%1>b5O%RSj&I!VYKFfZk=<<)QvFa`3hlcS(>+gD=
zX{~GSdSvO6zifB*cXm|1H1eTg>wub&`vev?2pXK8(!b2jFgwsC7u>E*DPWd&=v0X|
zl9>%f=?Up!g&{L9Rf9AxxOMK0ot+gvrl~;t-T3%W{^oq^JLUxGWxxTlG#=S~mx#a-
z1UeHTSd_I@S#rog_Ap3toTVj~{3q-uBp^ymM%ga?9SJ+PXlLe(cJ9pg&#`1%B{fUl
zQBPPF{n8zGXd9?1haoJIbazb(=};fCm|;F3U)|%gQ;>l`Hm)(PQRkuCE_pi{<Nq=|
zBmyICLhip<+VdDL1m&iu#1|D2BRhC28&LsC8RH@-Ty=82=R4V&Rn~BEQ8iAjFb%-g
z(%u=FrYs<b8Lq&Rj8?LW)d*%zVo(LyB#ezxRqr%L!VbI3WbPCI9{Nycw~m@jE0q#?
z<d?55%}%ei_T~p?S<jgh6pn1`*v%jfrVY%Tgy{kq%h7pq=>eLbcJ8wcL<q89s9*$b
zP2VX5#ZqA??(EV&&>^PzaQwB)G8TRB$#$P>t&k;0R}XD%)-KN0^w;-huMAm>$B@$q
zd<F@eOK5UscqmntX{bX`D_f{CI=og)1ia2H<yXGZz)3MzK|sX|ik;VBtNsl`sc#zU
zn@vyt<bzABFRIzS&sudnrsDP#j_Vt?zgR}~Zw_P$glH)*oJx1NVv_x&OvAu|mS6fK
zph8p7p_u}U*y`e^+W{-wVfYbEQ`CN~wWM}`{>(X+W$~kVAb)<jwT~uyb7!v(A=8kR
zN9S9+xMrGme-5Ws^A`;t4ugJi2gytTw8#aIvTu|2fzmBuNiL~3QG;*=QnI*hgMm-x
zXJ2T&F{}TCVV-EsZLV%?tldTDdaCI7s>)|Q-!e+e+QfEQx@SUt6ZooS$r1qx0_{ud
zsxDZ7x#(>K4*D7~QWhe0$OHgm7t8s~9IY>Ecdkyk`jZ_15U%Ght+mF~7?A{NazDY>
z6yoaW(jln`R34Nv_2mI(8~$EgO|yWV+Q`x^;cVH-%C^SQaBXc*CHdGZ=9y#bteuM8
zGZZDllN#ZIy-Yfrb|f@WLw++nmnqlKW6dP$h2~Z=0J}mwDYL`1YO!;hNNd_2S3qo$
z)>|Kj*C}G!p)4s-0KlZQC<vBsV90$Ifh}{(buuCfAE7-4x2nSgf&|(s<s$^CXY1CD
zsiiH_MQ7wYHqb@4pciF@Q><k9K*d6Gn+q#Kqi`MVsyZZXW%?jd9a1DM;}ng67R&yx
zm^~IedzEF{13kYrYSw9^?{^K;I^9y>>7><8YpK3f7GZ1Uy`QokP-AFrGQw_^{zkc#
z$ETG;dkoO1bx>6~gEo_OHg_aQVOd3XYmD5okCEG04i;!v=JcH^1<;PPxWk`$l(Hi9
zhppXe9P$7^`sNHgm6k&rAkz#Ok0|{qB+JzUp0F6^9+46+O>_H-8sMgml>#0W%5l-`
z@*cCE${4{{Za8v1zo%oN<+%a}`!q$G2P1Noas)Xsl~p!lD)*0N&ZVtN#`ppZZ<&%*
zvI7LJBC@3~b*J?GY1ek~Z2|K^;WE6_q6yW@PgVOQgu+X}no5K;OWtQBl`r7%AL;?~
z1V_hN`joT|&_>-kTF1byZ<Fo|jGpJM3tFo0v-#k|)&i@|CCd^Gzv34E(nsjz<J=q)
zTijsn0m)ZkXBQxM_0oT!`HJOn5H`bQ`R)xztvo?C^AOq+F`)j6D{H<DWW17{K4yKN
zT|v57+g>kN`|o4cW_2~YE2cIiiX?q%S`_t7=YzMA7buc4?lQN@HHKU@@~4ae#3Q(%
zbX+fc{bTjT?evCpnKrg_^u7c6zkc4jz|2p)$2!~0j<4WS8Ue<vDA*)rRly;J(2jD6
z$R@ftSy#6*$fK4eVSGk}ZWrUIQBj=87XWA%C9bv0a=w$pvW6uLxWct6zsO{{7{fLT
zJT^?(Kd-JY8+g|O*7u70sWZFZ_t>8<)p#w?)A$nT;i?e1L9M`OxSCXWm3toG6QUeo
zXCX^{00fux2l$Hx1B6r_fG5z(p-PBOcFD!&cUG$vvs-zNFDseP)xO%ir|~#!@m0Fd
zj&zL<YEiS>zD6sY5G9UGl5dXCXrwr`NMoBK3n1oik$zE?!hlq<ITSt?Ef0<ASPH$!
z>@AxR_KDpU!@hKFb@+OI^H=5Y2s^nQqBOxgSST&Z;!{RTzzqBZOuG5ff<$E!Z@Ijm
zw&#QASv)gN&_Ec2^ipFiy<DFWls57gz6?so3-v=!z$OyCYC$Z~Wv2&(<bIa{zYMob
ztxWVDt20#DG^tON1<;tJsJx;==5@mKGHnr!_O4M-t|0SMk6Wksa$G<Sic|&SaDU|j
zK!78-ALR>1+yubB`pnEEL>+TXr%baIHKGK-DXLe70%!VKWhA{+dr>gM6;F!BC!hGr
z%>p)1DOD@L5;0?fU}B*}E2}s`<B$qLA8|iW<`$VK1tJnbG@R}QdYAXK_0B~D&BUnH
zHX6;z#+9{=t@q@EC$0Tvt#yuiIsf$Q);HCX);Wg0fe<+;cXxX_`_hZ1A!GF4K?Z{{
zoJ$83HUoYYGGyfsT_Lr=2mvgkB+E0sW8@Z{4paKeu9j5<Buvy{Rjn1arz>@U3T?9D
zq;<WSKmBv7MRMd~l>S2&1eYZqNr4DrdbKQ@q@C$v6t>gti9Hz?7o}Cv+DzzBD^V1|
zAUiMfylmtx>#TNXTY8~3U#$1mzg3LiT)k53Xh^0!alg9<5EUR7-TY$uB0y$tR0eHS
z7PF$6x~0b`M=>(e(Bimo;!AeJd}Gqq_CRGl|NQILDmulxcUa`NYY>Sd-`t}B5tgC}
zH3|_D2t|oK?8sKCG^q6_+{2rJb-_B14%T{(^qjWxp3}8Y-lxXKvd`UC_u9cKKV0Ry
zE6EN!t^CKg)i1iGX_+q=wTJkc$P0P<Wc~Kq0sz##JS91N>vWxV{lMs<19|H|>vyP)
z)G8Q+7i7y&0a@0LREIPVnz=ZrQzSi)(V)1a<y=#11M&jB%Jp&qiu~Yz)SpwVoa!&<
KBX8ATJobN+wPQj6

delta 11172
zcmd5?eRNgTm4El#_wtgLyblZ{1cbaqApwEh?++xB1mx4m)Xr$NQb3F%0tujEsx*M!
z%g0h{tMOQ;RG49=*48O4OK(Tp4n^10X)9A_DYa{6S~~4=rY@G~)Yc^n>-^3=xi|N`
z5dC+OweAX+_wLze@8ACI-`?l$+BId@t|<>sA=4iJ(#1Wv?q2)sTIt9~L$_6(O%=}+
z-zdIRJX-u}abIy;@y6nkqEehvOc&lRyiqt&_*T`~!efP9g^h*3EL>4&FI<#ApMNv|
z^ZfJqBl(B&cjs@)UzNA=<$SL0y}Do3y<GR*y06yxbpv%b)Ge-?S2wwi)Sjt5ReQYl
z>DotYzfilO_L|xSwKHo=xpTSS<%V~V(v0c3>YD$p`Hz}^t~pxsM9uv*TWUU2(^E62
zrlBTP{Z{pB)jzC$s`|^rhsdPjq3nU|=d#PPTK3{>RkrF_OS!&^o~<GtNu}srSu#64
zDNWkwkt}(e|Mgfk*>&vM-)yU*`-)^RUwtx_NgYgOUa$Kgx0n`6<cn!f^WETZiTpsK
zr3F=cXIYkRDwg3!*Ey~02V3VW^E6v?Oj|YLEB#hKX_=4z?%#6DZ8!IAS>G3aOfwwY
zu{7M@v~K%$+}gSN{xA4WE=AfV>@W6Cs7ZMPx82&;zh(Gnx`#e9J97*D_4{NF9q1#+
z=%I(?SLyLL<@(HO?@r$gemRNUDbb@d$e~A+H<r>}ACTrD%~4F%HAb##WS?s#jk96E
z4rY&H>xQA4rZaNY=d^AJ?+bgeRkLy<rf!*<ZYTGVN=bL+g5h_`R;l2rzFo`qw6C^z
z_h4}FU2>O1o9ohl&y1Wt#ne?N@yi|)dQsQu*ZN5VHBQNsXBm}i+Lq<my2I|r&hus8
z3O>A;q~!^zR4OA~p31wqdD-@KhjcmOKTAq`1AZGVZ6*_go6BS!X)}`HQ><ud#}&5h
z(5KI2TIe6WOnDAgRa}>SGF}$$pUL;Pqj%4wIB07uD%~x21rJ|JW|AQ{0SE5FaFdQ6
zgM`Hdn-4>ws=~*OI22;!)G=;xEL~G97NOWYpZGIsYeq1&r!lyqjoeFyOw+JbMR#0&
zjlFJKB6K5zEKRX(SBtMPHUv-ED)Wqha!1o$*LDSqvVOy*p~v=5>fJMS)MF-H_hSN&
z%_b)4vXLE%s=55i9uRuW5MvX0OwsL;PmX-cqAzyHvsw+?bacaUlIwem{7ZPi8hiIl
z2zuv|ZUiWD1#?Ken&}Af!F?Vd!*U#5Rii7c;HhE>A9u!j!`4+>2v0tQ>I=0Dt25HO
zr5Ca*fYomxZF>C6xRnzCtKe0Ybdn*<Q8ir$_(!ghR}jwR%D_u#r=|))b3g&yPJsY`
zH9SX*nxkNM0yTM8`m=Mf+a?818l;OXv#p3;LaK>zWSIhjv1M)$J{5(O1uC|O{^xSi
z-0GN`tvH65Uf20^__CKXKdiOMHDm~19pzmP_z)egnskGPu=9A02Vct-f5!U3Q5;KT
znUX(?`(b|@&>M8lC;JKAeKym$(H@iK;p$_t8Espu$euCzY$M^oMgn30-*(#BN%jOK
zvud?O2Oc3S=|BgW*6t{ZX1cZ!T_bepA}n3Hu3c`qvSMwvhJXQ6cTFTcZpU20!Fqiv
zKOtAjR;RC!t_N0^0;}u%CLW{SUQkSbH@T6tiO*+P8cAS*VJ!Ooo=jQNIRHS1s;yN3
zZ;3xS#$97@-D2RbXew^BOSWVB_$j$u(o9`5EIYo^mru#Fvo<hgt6Fe=37IITy`6po
zow$PdkIf`=f>kTY33zZ=u32O$sN0F($X>9Au#+ga6va;Z8h@K1<y?j>Xn2^IJh%Ik
z_#W2u?rDk{k9K!v7SUC2rmqiB*^fx{+l$D>3&$ylsv{bLEaC7n^kkV=!d<snmJ99^
zfWxErR4Sjzb!P8MZ<U^^e)W--vR(k9F6P#OY8rzFuOpw9nvn0)W49;Dnn@3z&UDkG
zPmra->1X8S5?$L$ZXN=CCNmOV;^B4`!Fy3g>RO_hps>KukeO8B)1WClM?j^k*qUw#
zsPrX&+DJyu?cKd2a6d&JAd^jPw0bx(FS%N%DpqP*vYOd0*5S-Jh6-)N(k#8QfK>Ra
zlXO|JaN;5=a)2dnDRKaqIf*vSp)bECx6gHubt;q%z8z<sx7eQ=N1{Fk&l;0<A0%D$
zcw455zGILrRQVtIAL-e{nf3H9W|F=owkxg?Sp-HAem<U?=}$Z4mKF!11^Q4hYllCD
zn+4s|6ii!92FUX9D9F?<SPzs6^#||pFzED^j!g3m#kDmTg?)5kb?54QO%B@ps#MkD
zd|j?fzJ`{{WI+Ksrx=<8=?oB;*FolVIr=yXuOg{q7mSwHu*}kxugmSNHcFD^7(z06
zTl@mweIMKXlQ)wMWTBBzETfo1VhBG`-j0<lC(dMMwVH-u8K#)J5CgS*wNMN^ek;Vl
za@%C(A;xj6?;QjyAqRlyk=k3axyaWp^L4phw{_hS_6QYkh26-$X}zf!cl1*aW)9HP
za|vx&=I6Op(~`<J{6K&YEE_3KyB;9Zh9CqJG!-gBE`i02kE~T~W3-Sla*YT=No)<<
zWSK1xgoC1Gx?%zi_;o>PGkKTr#^AYBc5`|#e@RZ2o<jjUf&#YCpU8Jo%S6V{wvzb}
zTZ%H;Jprjv{y>-}{R_Wykzpj-7_>Cl3CHp$iFVsj4MS()8%_BEe*$k*p16ZGD!N<I
znQ?N3g~=b474~wEA$o{Wm*64YgoXw=3JZ4mrPu+D!Tuc=Z~*&47_xE1ovJv3h>q2r
zC{DGoYZEMTpfwdslzm(LhI($%?Y(;vtY^^fMVXNGR@45!k>><pJyi;u-iPlc&25&@
z#WFl|WoQdtG3X1+NtwQSN^WU}<Yd80c>j9;BEB>vXH&3XCnP7WdtGj78>a%JO-Skv
zgBI`0%$jK!j)5{KY>U_6t*QLnoSmJSUMp<{0M`J3xBC)bk_m*MV=w6=O=_(385zY}
zQ4QW1fA>tLiN4ZI%5zm0<phw6HvyIeWu)SWCYmgb)$cBfwzNnUqw0p|rjpjMhA*cB
zUnEm|96M?*LI#XkMHTB9D^Xm2R8}%OltFWJv{8y8Hhy_;C^U&F=Y*ut!jJ}Q53m-;
zhMZ59(l~$>D2kSMbdf-!tTV(U9U~+h4UN1h1jpOvePN`<*~(~O`h#RMX|k1AP&2_7
z*<&g!3XR3e(ErORc~%R=mxE2j@xW-uTjlNK@#yZ|JrsQAesV3Dsta4SVv5l%l^*Pn
z8%tsZ*KE4#7jn}~C@s}c!e%?B434a;Q~Bw+YqRUqE2XP}tOd|<gWe9l?oQOjn8nwq
zRP9Bxee_FY7N{;c3<%|Sm^2kGj6%)^2Z`?2!F|qjVkE{+il9i)jf>>$PxS7hs~#mc
z(t&fNkG}8<CR}9?6Rx8-kt%xVB-zV4oDE=OK`+IM4&y4J=ZkPL29}YL6*wMirK-4&
z<pS?<_>fKouiv|q`w@IJDHx<=BWbs6D;@$cmqg&uD;5cfS<*sL(}+T=nN0B}7-z9}
zM{vUv<Rqbg8kP$|*Fmz1v{|F`7FOk>hnt2_eoT7qw{p3?!VP0sAM+H;v1~nf?@5v`
zm%_2b!Bn<2eM{ZjxspUDl*j{rtCT@SS9<sIjbc19DYzCR8E6eh>~<ns99_T&Y~)(Q
zP?9-pB3dCrR0S@Oo9D2(m!gX<aJ<22+1>XfaoVypp>liZRl^ZNj-Qk_($b6QJ`8Pi
zmc^hPndf4O&>$TOBGDPMyypg)0(izwY`|eEYOk2e8@;=^sUT~XQbGAA<R-)m<G0b@
z7@=bK?O3ezw9ucP$&_1ANnO=tpN#X^yTQ9BdN`VytuK+A+4#^++8OW9v4oUE(7F53
z2`5W&yjuoPq4J_|V8S2-!Y_EA4-H@}%?TsRrqF9to3<Pv-wcqVuS)GUOiIen*yW<M
zw&)zxSo%soDQCgBnqj!Xd;baq=BrPn(sF$&b4sqyPYeq0kPN!>g}xRv{(;;8<+x4$
zSvVmcf_+I9A14_emUIe2#5a%3PnZ@Dy;MvAl3!q(7JT8jhNf%m@uR`m;Hx7Q0imuB
z=D&|<EJNu?%+-~+B)FQeLX-@sgvnr_ByzTN>B(l&T*fw_yafzEcwfzr0or;~!Sr*a
zTT1&W-}w5(a|BM>;OoKU56SOG<m~BGb~ycax!Sr9q|@O1lStui{|b&F$l0dgeM!EY
zv_Xm_<V$R{sU>ft5B+8{(E020q@~3|g~T{caL6YAa{gqrJ(Gg(r{!Bg4(NLlK92(g
zO_GEsP#6XX6P_e+w77iOp+g!S^dZmy$z|`~=Fbnll9jiSg16CM5akE@y!M>jM;4;l
zzmNf#6eKr?g+c=qv{%t72h-Iq(`&QKpfU`Rlgr{np_8WqddjAtkeBZz06v<*ME%1T
z!eWWRt`L59>Qkogd70LkhN(DOqKSnZ;N|*MDqo$mvYXOJq*K+y6Xc$Px5V4Tw+HrZ
z41UxgZ$~^=cw6{QXoow4d#1>LMJ}_@PbGH2luT#`Mz_gbGX|s&eMB0%IqFzAMcAus
zUL^=MZ?U(TuLt}z$6Wl#p*MO$;rLi>+M=Z|leg*5r^`Q>;#$d>ZcI)Xz>FhePMpnL
z)}reQR5LSsG~LE*zKdX-Cge{HTv5<#(QIBbO!_M?)6gUI-1x_sePcMFpnMAkT<zV?
zH{RE~`}*LQ^?>&f#whVfk;!~~fwvS&5?t}HnVlZ++QJ@P;cC}H?<K5whqp0Y5o3#F
z*;CYc8wYM(ciZpx24_q1ZKP?u!$vo32Cg=o(c=gP&N!)9I0Pd!6<sZ65=kra)z78U
z<mpuA+vMqdIX5dPmE|W$!I!*&$T(#CKeYn?%Pq}JvR~{DUEmkECdxKfuNaGPAo3AU
zqX`1CL**O8z?Z%L;MfdVmeXLu?ZNZy@}QK4BySA5XUm%*R}F#7S<Z6kk3%>ZctwY_
ztdbxuBMdACaBRF3HUp?A{x^Ek*5J~)@_$`4?|Z5IRk<6pgXz7}_o|<Pq^kJw*7#`-
zC(Om`gB2^}`;p8m{VIMF$-E=jzgWH&7#r0M?5}7;rdZ;liZbrljA=w6*tU!A#S%yj
zldBA0f3&d<`j&!5hcG!xw4AY9F@#GHIgH~hY9%=rDkAb~l6~;^7_S;QG=io|;*UG3
znjI2Hp870VW@?FHKBtN>%%&wU92kMu6|jA>4#f6^Wd$Ql%sg~vO$;<|lP_~upt?^D
zX7<7fQ$#0>Hhii|55Fci(C?hdOrfXGlX452(rM!9kf->BKM{xBJA#wlsJ!!C;n0i4
zh@&+!T&pO-F+T0^RHO>yK<qwvZ<U-^>Z|WVrVgeu17xu7T+JWk*Q5tg^zV@-R{iN?
zW_4E;yzPE9w=7Jg;$C?RI|g=fkP`n!mR0ySjyQ-u9129&$its_j)gu6p%i@Y7yazW
z)E+vsDaam>_pzB<qK=1{8sS>;*cVvi7#z3zuVDYNAk<wP?TM!}sZ^~r5WRkp3Q7;k
z>!kuTz=)Zl2EfgG*-YTVLkf!?#_J07*?uy)M-T%5ESxK(rs`r`W`}3Z9FXWTF~qJ3
zg9Po3A;F*TllK#dm?U{6>R;ppO{6{3p}#m!W?uq*0I##p$AD+AwWMV=l{q7;c(J7_
zC_OAsmFd@}ld6Je`(+N0#@^k7f%}Nug?Nl3GS~^sNYV)nFm8{?3^7H+DVVT*a6VH^
zN$b4N`OVydrQmudJLsQ><?mDFkK~6``G$NaUAutvEw7li#RaIc&x$NmL~h0BqXNXT
zxQy{E=HBcnv71nJ_-i=(ZioIp^cDFs(00Gy%8d*Af85+i7FH-<JcDPpo8Zfk0XbSL
zPU<Xr#>-$>gb1VcNE!gpwQwA8&>aNmx@@r+#6`joo4_bCIJg#+lS*IeC$nZc29D1(
zQFn1MdMuT9a-YtAHvO1%43u(MY6LAWiUX23wmfXeJ;8@xmkWd*I3+g>i8A|RHXONL
zY?LmL946XzOwfd=psVTO*08Q7R;6p4fIcmEjL#io6Oki#yeXDAUNOc(*MWq>Lj^kX
z?x^#-qmb>sCwTS<LWcA)lgFYA?IA!cRKJn*xp2)g#a`k9@736Bbf7a3#IVINd>41n
z*52LQ*<9i$hk;W1Z@-lrTH_Th;msd5?BnTRhohMmAlIyTM+FCv3;A==lhg;j&xFh%
zzG}hbB;OiFlL@(i2=a|;#%P^ke?`lF(g4OW4fK)8Z8Ef01(~Dr`Ho_KMk*~GNaft@
z;`Bo4Ku|g@e;)67slF0C^c(p*Ov3sdQCJ{h*ZdA)p>Df8c|8twlTJo7Lqo?c#-DhM
zgQ=k!wi$6ig6XH_O%gqRIMYGr{<VC9y|6d6-4Kr`VnTO~q&=n;4h<`23KvKf%fNU6
zbGW37Fd~>sJKvHoq5tq>*$s}rCD)VAabyTPe!5V=bb;9Eiq|pMwk^C&DE7#y6!~`!
zj{hI;=P-2aO}`@t>37l(-AC51qmORS=mmYgY0NjRdDsCGAb5RyrejksE&V7}_YuG|
z6ckQp?npObGavHJfaDX0CT=(^XGiU;0~|Wxts9)MuL#50%|`Lb*KXo?q76$r7Wa83
rRKDC7R)MKu`-Zu597cAA)9EjPIl<6>XPzf%oakH`biS2&arplL3Qzhk

diff --git a/server/training/train_server.py b/server/training/train_server.py
index d44c028..44c9f20 100644
--- a/server/training/train_server.py
+++ b/server/training/train_server.py
@@ -6,9 +6,10 @@
     python train_server.py --task-json tasks/<taskId>.json
 
 任务参数（Go 侧写入，字段相对训练机 workdir）：
-    workdir       训练机工作目录（脚本 / yolov8n.pt / venv 所在），启动即 chdir
+    workdir       训练机工作目录（脚本 / yolov8s.pt / venv 所在），启动即 chdir
     yolo          训练集目录（含 dataset.yaml），相对 workdir
-    imgsz         训练/导出分辨率（默认 704，与端侧推理对齐）
+    model         训练基座权重文件名（workdir 下，默认 yolov8s.pt）
+    imgsz         训练/导出分辨率（默认 1280，与端侧推理对齐）
     epochs        训练轮数
     batch         批大小
     device        GPU 编号或 cpu
@@ -236,7 +237,7 @@ def main():
     except Exception:
         pass  # 日志不可写（如 workdir 缺失）不阻断错误路径，traceback 仍走 stderr
 
-    imgsz = int(task.get("imgsz") or 704)
+    imgsz = int(task.get("imgsz") or 1280)
     epochs = int(task.get("epochs") or 150)
     batch = int(task.get("batch") or 16)
     device = task.get("device") or "0"
@@ -245,7 +246,7 @@ def main():
 
     try:
         from ultralytics import YOLO
-        model = YOLO("yolov8n.pt")
+        model = YOLO(task.get("model") or "yolov8s.pt")
         register_callback(model)
         model.train(
             data=data, imgsz=imgsz, epochs=epochs,
diff --git a/server/技术设计.md b/server/技术设计.md
index 437a620..6fbcbd2 100644
--- a/server/技术设计.md
+++ b/server/技术设计.md
@@ -96,7 +96,7 @@ CREATE TABLE IF NOT EXISTS model_training (
   name          TEXT NOT NULL,            -- 任务名（默认「数据集+时间」）
   status        TEXT NOT NULL DEFAULT 'running',  -- running | success | failed
   dataset       TEXT NOT NULL,            -- 训练机数据集名（datasetDir 下子目录名）
-  imgsz         INTEGER NOT NULL DEFAULT 704,
+  imgsz         INTEGER NOT NULL DEFAULT 1280,
   epochs        INTEGER NOT NULL DEFAULT 150,
   batch         INTEGER NOT NULL DEFAULT 16,
   device        TEXT NOT NULL DEFAULT '0',
@@ -360,12 +360,16 @@ training:
   datasetDir: datasets        # 训练机数据集根目录（相对 workdir，数据集为子目录）
   concurrency: 1              # GPU 独占：同时仅一个 running，新任务排队
   timeoutMinutes: 240         # 超时判死（started_at 起算；100 张实测约 3 分钟）
-  imgsz: 704                  # 训练/导出分辨率（须与 App 端推理输入对齐）
+  model: yolov8s.pt           # 训练基座权重（须已同步到训练机 workdir）
+  imgsz: 1280                 # 训练/导出分辨率（须与 App 端推理输入对齐）
   epochs: 150                 # 训练轮数（patience 30 早停，设大可自动停）
   batch: 16                   # 批大小（按训练机显存调整）
   device: "0"                 # GPU 编号（cpu 用 cpu）
 ```
 
+- **模型与分辨率升级（2026-09-01，用户定案）**：基座 `yolov8n.pt` → `yolov8s.pt`、`imgsz` 704 → **1280**——场景为远距离小目标 + 遮挡多，nano 容量不足且 704 输入下小目标仅剩 5~15 像素；s@1280 是端侧延迟可接受内的最优效果（中端机 70-120ms/帧，预览不卡、框 8-14fps 更新；高端机 30-45ms 流畅）。改动链路：`config.yml`（training.model/imgsz）→ task.json 传 model → `train_server.py` 读 task model → 导出 litert 输入 NCHW [1,3,1280,1280] → App 端 `TfliteDetector` 输入尺寸从模型形状动态读取（`defaultInputSize` 兜底 1280）。相机侧无需改：Android 分析帧固定 16:9 ≤1920（典型 1920x1080）、拍照原图 4000px+，letterbox 后均满足 1280。显存评估：12G 跑 s@1280 batch 8~16 可行（yolov8s 11.2M 参数）。端侧量化保持 fp16（int8 对 <10px 小目标掉点明显）。
+- **预标注模型固定 700x700（2026-09-01 实测确认）**：local-ai 的 rfdetr-xlarge 模型包 `inference_config.json` 声明固定输入 700x700（`training_input_size` 700、`dynamic_spatial_size_supported:false`、`resize_mode:stretch`），提交任何分辨率最终都拉伸到 700x700 推理——`localAi.inputSize` 只是提交前等比缩放，改大无增益（曾误改 1280，实测后回退 700）。**标注框精度天花板 = 700 分辨率**：极小目标（<5px）框不准是标注侧固有噪声；训练 imgsz=1280 与标注 700 并存可行（框坐标按比例映射到 1280 图），突破标注精度需换支持高分辨率输入的检测模型（RT-DETR / YOLO 动态输入 ONNX），未实施。
+
 - **训练参数默认走配置（2026-08-26）**：imgsz/epochs/batch/device 不随管理端请求传（界面一键开始），由 `training` 节点统一配置——device 取决于训练机硬件、imgsz 必须与端侧推理对齐、epochs 取决于算力预期，均为部署级参数；任务记录仍存各值（`model_training.imgsz/epochs/batch/device`）供列表展示
 
 - service 内 `Runner` 接口：`Start(ctx, *TrainingJob) (pid, error)` / `FetchLogTail(ctx, job)` / `IsAlive(ctx, job) bool` / `Cancel(ctx, job)` / `FetchArtifacts(ctx, job, destDir)`；`subprocess` 与 `ssh` 两个实现，按 config `mode` 选择；**ssh 凭据直接读本节点 `training.ssh` 配置**（见「全局训练配置」节）