Skip to content

feat: model_config 支持 summaries_set(PCOC / scalars)训练期 TensorBoard 监控#555

Closed
fooSynaptic wants to merge 1 commit into
alibaba:masterfrom
fooSynaptic:feat/summaries-set-pcoc
Closed

feat: model_config 支持 summaries_set(PCOC / scalars)训练期 TensorBoard 监控#555
fooSynaptic wants to merge 1 commit into
alibaba:masterfrom
fooSynaptic:feat/summaries-set-pcoc

Conversation

@fooSynaptic

@fooSynaptic fooSynaptic commented Jul 8, 2026

Copy link
Copy Markdown
Collaborator

背景

修复 #532:训练过程中希望在 TensorBoard 中监控 PCOC 等指标,且支持按特征切片查看预估均值。

改动说明

model_config.summaries_set(不是 eval_config)新增训练期 TensorBoard summary 配置:

类型 作用 TensorBoard tag
pcoc 全局 PCOC = mean(pred) / mean(label) summary/pcocsummary/predicted_ctrsummary/observed_ctr
scalars 特征切片上的 mean(pred)(非 PCOC) summary/<name>
  • pred_name 对应 prediction_dict 的 key;若为 logits 会先 sigmoid
  • 多塔模型在顶层通过 pred_name: "probs_ctr" 等区分任务,无需在 task_towers 重复配置
  • scalarsfeature_value 兼容 RawFeature 数值型特征(如 c1=1005

配置示例

model_config {
  summaries_set {
    pcoc {
      # pred_name: "probs"   # default
      # epsilon: 1e-7
    }
  }
  summaries_set {
    scalars {
      name: "c1_1005_pred"
      pred_name: "probs"
      feature_name: "c1"
      feature_value: "1005"
    }
  }
}

参见 samples/model_config/deepfm_combo_on_avazu_ctr.config 末尾示例。

测试

  • easy_rec/python/test/eval_summary_test.py(proto 解析、PCOC 图、数值特征切片)
  • deepfm_combo_on_avazu_ctr.config 样本 50 step 短训,events 中 PCOC / scalars tag 与数值校验通过

@CLAassistant

CLAassistant commented Jul 8, 2026

Copy link
Copy Markdown

CLA assistant check
All committers have signed the CLA.

@fooSynaptic fooSynaptic force-pushed the feat/summaries-set-pcoc branch 3 times, most recently from 9594e7e to 095a04d Compare July 8, 2026 16:12
训练期 TensorBoard summary 配置在 model_config.summaries_set(非 eval_config)。
- PCOC:全局 mean(pred)/mean(label),写入 summary/pcoc 等 tag
- scalars:按特征取值切片统计 mean(pred),写入 summary/<name>
- 兼容 RawFeature 数值型特征的 feature_value 匹配

Fixes alibaba#532
@fooSynaptic fooSynaptic force-pushed the feat/summaries-set-pcoc branch from 095a04d to d580f50 Compare July 8, 2026 16:14
@fooSynaptic fooSynaptic changed the title feat: add configurable summaries_set (PCOC) for TensorBoard monitoring feat: model_config 支持 summaries_set(PCOC / scalars)训练期 TensorBoard 监控 Jul 8, 2026
@fooSynaptic fooSynaptic closed this Jul 8, 2026
@fooSynaptic fooSynaptic reopened this Jul 8, 2026
@fooSynaptic fooSynaptic closed this Jul 8, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants