SwanLab icon indicating copy to clipboard operation
SwanLab copied to clipboard

[FEATURE] record model info

Open Zeyi-Lin opened this issue 1 year ago • 3 comments

根据 #166 中提到的打印模型权重信息的需求,计划很快上线相关功能

👾 What does the proposed API look like

import swanlab
import torchvision
from torchvision.models import ResNet50_Weights

# load pretrain resnet50
model = torchvision.models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)

run = swanlab.init()
run.watch_model_graph(model)

Zeyi-Lin avatar Jan 07 '24 15:01 Zeyi-Lin

UI like:

image

Zeyi-Lin avatar Jan 13 '24 07:01 Zeyi-Lin

  1. 感觉api叫做watch_model_graph 容易跟梯度图关联起来
  2. 好像更喜欢的是torchinfo那样的model的结构打印
  3. 如果对我来说,我可能更喜欢netron这种的可视化
  4. 可能需要检查下如果是多卡包裹模型的情况

ZhikangNiu avatar Jan 13 '24 07:01 ZhikangNiu

  1. 感觉api叫做watch_model_graph 容易跟梯度图关联起来
  2. 好像更喜欢的是torchinfo那样的model的结构打印
  3. 如果对我来说,我可能更喜欢netron这种的可视化
  4. 可能需要检查下如果是多卡包裹模型的情况

❤️nice advice

Zeyi-Lin avatar Jan 13 '24 10:01 Zeyi-Lin