fix bugs in modeldict_cossim

Author: WwZzz

flgo/__init__.py

@@ -1,4 +1,4 @@
-from .utils.fflow import init, gen_task, gen_task_from_para, gen_benchmark_from_file, gen_decentralized_benchmark, gen_hierarchical_benchmark, convert_model,tune, run_in_parallel, multi_init_and_run
+from .utils.fflow import init, gen_task, gen_task_from_para, gen_benchmark_from_file, gen_decentralized_benchmark, gen_hierarchical_benchmark, convert_model,tune, run_in_parallel, module2fmodule, multi_init_and_run
 
 communicator = None
 

flgo/algorithm/fednova.py

@@ -6,6 +6,9 @@
 from flgo.utils import fmodule
 
 class Server(BasicServer):
+    def initialize(self, *args, **kwargs):
+        self.sample_option = 'uniform'
+
     def iterate(self):
         self.selected_clients = self.sample()
         # training

flgo/utils/fflow.py

@@ -985,3 +985,11 @@ def init_global_module(self, object):
     else:
         raise NotImplementedError('The current version only support converting model for horizontalFL and DecentralizedFL.')
     return AnonymousModel()
+
+def module2fmodule(Model):
+    class TempFModule(Model, flgo.utils.fmodule.FModule):
+        def __init__(self, *args, **kwargs):
+            super(TempFModule, self).__init__(*args, **kwargs)
+    return TempFModule
+
+

flgo/utils/fmodule.py

@@ -852,7 +852,7 @@ def _modeldict_cossim(md1, md2):
     l1 = torch.tensor(0.).to(md1[list(md1)[0]].device)
     l2 = torch.tensor(0.).to(md1[list(md1)[0]].device)
     for layer in md1.keys():
-        if md1[layer] is None or md1[layer].requires_grad==False:
+        if md1[layer] is None:
             continue
         res += (md1[layer].view(-1).dot(md2[layer].view(-1)))
         l1 += torch.sum(torch.pow(md1[layer], 2))

tutorial/2.1_Try_Your_Own_Algorithms - Local_Training.ipynb

@@ -3,31 +3,47 @@
   {
    "cell_type": "markdown",
    "id": "c48d75ca",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "### 介绍 \\ Introduction"
    ]
   },
   {
    "cell_type": "markdown",
    "id": "e5ee136c",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "这一节主要介绍如何使用FLGo来实现自己的idea。这里首先关注的是本地训练阶段有所变化的算法（如FedProx，MOON等），这类算法在联邦学习中是极为常见的，因为联邦学习中用户数据分布通常是non-IID的，导致本地训练容易越走越偏。下面从FedProx入手，讲解如何使用FLGo复现该算法。"
    ]
   },
   {
    "cell_type": "markdown",
    "id": "088c1728",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "### FedProx简介 "
    ]
   },
   {
    "cell_type": "markdown",
    "id": "4e264286",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "FedProx是Li Tian等人于2018年([论文链接](https://arxiv.org/abs/1812.06127))所提出的一种针对系统异构性鲁棒的联邦优化算法，发表于MLSys 2020上。它相较于FedAvg主要做出了两点改进：\n",
     "\n",
@@ -43,7 +59,11 @@
   {
    "cell_type": "markdown",
    "id": "b2df44b4",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "### 加入算法超参数\n",
     "所有的横向联邦中的超参数都在Server的initialize方法中加入，加入的方法是调用Server.init_algo_para方法，并传入超参数字典。例如，对于fedprox来说，超参数是mu，因此只要通过该方法传入{'mu':0.01}，Server和Client就会被添加额外的属性Server.mu和Client.mu来访问该超参数，其中0.01是默认值。"
@@ -53,7 +73,11 @@
    "cell_type": "code",
    "execution_count": 1,
    "id": "08a33678",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   },
    "outputs": [],
    "source": [
     "import flgo.algorithm.fedbase as fedbase\n",
@@ -72,7 +96,11 @@
   {
    "cell_type": "markdown",
    "id": "a04dd293",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "### 修改本地训练阶段"
    ]
@@ -81,7 +109,11 @@
    "cell_type": "code",
    "execution_count": 2,
    "id": "6bf59686",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   },
    "outputs": [],
    "source": [
     "class Client(fedbase.BasicClient):\n",
@@ -112,7 +144,11 @@
   {
    "cell_type": "markdown",
    "id": "e10e0698",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "### 构造算法fedprox\n",
     "在FLGo中，每个算法既可以用一个类表示，也可以用一个文件表示，唯一的要求是算法必须具备 algorithm_name.Server和algorithm_name.Client这两个属性（对于横向联邦是这俩）。"
@@ -122,7 +158,11 @@
    "cell_type": "code",
    "execution_count": 3,
    "id": "8cd43e21",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   },
    "outputs": [],
    "source": [
     "class fedprox:\n",
@@ -133,7 +173,11 @@
   {
    "cell_type": "markdown",
    "id": "a3ea442a",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   },
    "source": [
     "### 测试fedprox"
    ]
@@ -142,7 +186,11 @@
    "cell_type": "code",
    "execution_count": 4,
    "id": "0936a8f9",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   },
    "outputs": [
     {
      "name": "stderr",
@@ -1004,7 +1052,11 @@
    "cell_type": "code",
    "execution_count": null,
    "id": "00b4061b",
-   "metadata": {},
+   "metadata": {
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   },
    "outputs": [],
    "source": []
   }
@@ -1030,4 +1082,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 5
-}
+}