先後端對接之json

結構梳理詳見：Pytorch&先後端工做梳理

如下介紹先後端對接的json文件

base.json

描述：包含全部基礎的網絡層，已經in和out節點

{
  "type": "base",
  //sequential表示嵌套模型，base表示單個網絡層
  "name": "base_1",
  //對於base層爲網絡層對應的名字，默認按序號排列
  "attribute": {
    "layer_type": "pool_layer",
    //對於base attribute的結構,以pool_layer爲例
    //對於輸入和輸出層，"layer_type":"in"/"out"，僅有"left"/"right"屬性

    "attribute": {
      "layer_type": "max_pool",
      "attribute": {
        "kernel_size": 2,
        "stride": 2,
        "padding": 0
      }
    },
    "left": "XXXpx",
    //繪製時的位置,Sequential可缺省該屬性
    "Right": "XXXpx"
    //繪製時的位置,Sequential可缺省該屬性
  }
}

connection.json

描述：在sequential中用以描述canvas鏈接的文件

{
  "source": {
    "id": "canvas_%d",
    "anchor_position": "Bottom"
    //("Bottom"/"Up"/"Left"/"Right"),
    //對於type=base表示箭頭鏈接位置,對於Sequential可缺省
  },
  "target": {
    "id": "canvas_%d",
    "anchor_position": "Up"
    //("Bottom"/"Up"/"Left"/"Right"),
    //對於type=base表示箭頭鏈接位置,對於Sequential可缺省
  }
}

sequential.json

描述：核心的封裝結構，有明確且單一的輸入和輸出節點

{
  "type": "sequential",
  //sequential表示嵌套模型，base表示單個網絡層
  "name": "sequential 01",
  //對於Sequential爲用戶在保存網絡層時爲網絡層取的名字，默認按照sequential_%d來排序
  "attribute": {
    "in": "canvas_%d",
    //表示每一個Sequential開始節點，即入度爲0的節點，該節點必定是type="base" && attribute.layer_type = "in"
    "out": "canvas_%d",
    //表示每一個Sequential結束節點，即出度爲0的節點，該節點必定是type="base" && attribute.layer_type = "out"
    //對於Sequential attribute的結構
    "nets": {
      "canvas_%d": "sequential1.json",
      //這裏能夠是sequential.json或者base.json,modulelist.json,moduledict.json，能夠有多個
      "canvas_2": "base1.json"
    },
    "nets_conn": [
      //描述每一個Sequential內部的連通狀況,base層沒有該屬性
      "connection1.json",
      "connection2.json"
    ]
  }
}

modulelist.json

描述：一種封裝的網絡結構，多個相同的層封裝在一塊兒，注意其中的canvas只有一個

{
  "type": "modulelist",
  "name": "multiple layers",
  //對於modulelist爲用戶在保存網絡層時爲網絡層取的名字
  "attribute": {
    //對於modulelist不須要指定in
    "canvas_%d": "sequential1.json",
    //這裏能夠是sequential.json或者base.json等，只能是一個
    "num": 10
  }
}

moduledict.json

描述：一種封裝的網絡結構，至關於一個多路選擇器，從衆多canvas中選擇一個

{
  "type": "moduledict",
  "name": "moduledict_1",
  //對於moduledict爲用戶在保存網絡層時爲網絡層取的名字，默認在後面表序號
  "attribute": {
    "default": "canvas_1",
    "choose": "canvas_2",
    //moduledict至關與一個多路選擇器，有一個default路，和可選的canvas
    "nets": {
      "canvas_%d": "sequential1.json",
      //這裏能夠是sequential.json或者base.json等，能夠有多個
      "canvas_2": "base1.json"
    }
  }
}

static.json

描述：靜態變量。後續可能添加數據模塊（數據加強，打亂等）

{
  "epoch": 100,
  //全數據集訓練次數 非0正數
  "learning_rate": 0.01,
  //學習率 大於0的實數

  "learning_rate_scheduler": {
    "name": "StepLR",
    "attribute": {
      "step_size": 50,
      "gamma": 0.1
    }
  },
  "device": "gpu",
  "data": "svhn",
  //mnist, cifar10, stl10, svhn等

  "optimizer": {
    "name": "Adam",
    "attribute": {
      "momentum": 0.9
    }
  },
  //SGD, RMSprop, Adam

  "loss": {
    "name": "CrossEntropyLoss",
    "attribute": {
      "reduction":"mean"
    }
  },

  "batch_size": 16
}

現修改成data.json

描述：前端最後給後端傳的全部數據

static = {
    "epoch": epoch,
    "learning_rate": learning_rate,
    "batch_size": batch_size,
    "learning_rate_scheduler":learning_rate_scheduler,
    "device":platform,
    "data":dataset,
    "optimizer":optimizer,
    "loss":loss
};

structure = {
    "canvas": sequential,
    "static": static
};

ret = {
    "name" : $("#model_name").val(),
    "structure":structure
}
//這個ret是傳回後端的json格式，爲了後端的向下兼容

原data.json 現已棄用

{
  "canvas": "sequential.json",
  "static": "static.json"
}