cifar_ESD_with_local_AGD.ipynb

{
  "cells": [
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "Pyd1rz813mEo"
      },
      "outputs": [],
      "source": [
        "## IMPORTS ##\n",
        "\n",
        "import os\n",
        "import networkx as nx\n",
        "import matplotlib.pyplot as plt\n",
        "import numpy as np\n",
        "import torch\n",
        "from torch import nn\n",
        "from torchvision import datasets, transforms\n",
        "from torch.utils.data import DataLoader\n",
        "import torch.nn.functional as F\n",
        "from scipy.linalg import kron"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 287
        },
        "id": "UMAVkFc33pSZ",
        "outputId": "7f573fcf-2abc-4623-c3d3-5bc5f259e01d"
      },
      "outputs": [
        {
          "data": {
            "image/png": "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",
            "text/plain": [
              "<Figure size 250x250 with 1 Axes>"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        }
      ],
      "source": [
        "## GENERATE RANDOM GRAPH ##\n",
        "\n",
        "N = 10  # number of nodes\n",
        "p = 1  # probability of creating edge\n",
        "\n",
        "# generate random graph\n",
        "G = nx.erdos_renyi_graph(N, p)\n",
        "\n",
        "# visualize graph\n",
        "plt.figure(figsize=(2.5, 2.5))\n",
        "nx.draw(G, with_labels=True)\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "c1Jo6jjc3sG7"
      },
      "outputs": [],
      "source": [
        "## CALCULATE A FROM GRAPH ##\n",
        "\n",
        "adj_matrix = nx.adjacency_matrix(G).todense()\n",
        "\n",
        "def get_A(adj_matrix):\n",
        "    \"\"\"\n",
        "    This function works given the following assumptions on the adjacency matrix:\n",
        "        1. square\n",
        "        2. symmetric (graph is undirected)\n",
        "        3. binary\n",
        "    \"\"\"\n",
        "\n",
        "    N_agents = adj_matrix.shape[0]  # num agents\n",
        "    A = np.zeros((2*N_agents, 2*N_agents)) # size of A will be 2*size of adj_matrix (to account for shared and local layers)\n",
        "\n",
        "    for i in range(N_agents):\n",
        "        agent_num = i + 1 # makes the math easier\n",
        "\n",
        "        # update local layer row (only depends on itself)\n",
        "        local_layer_index = (2 * agent_num) - 1\n",
        "        A[local_layer_index, local_layer_index] = 1\n",
        "\n",
        "        # update shared layer row\n",
        "        for j in range(N_agents):\n",
        "            constant = np.count_nonzero(adj_matrix[i, :]) + 1 # number of connections agent i has including itself\n",
        "            shared_layer_index = 2 * (agent_num - 1)\n",
        "            if adj_matrix[i, j] == 1 or j == i:\n",
        "                A[shared_layer_index, 2 * (j - 1)] = 1/constant\n",
        "    return A\n",
        "\n",
        "A = get_A(adj_matrix)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "Mcy26Zat3tce"
      },
      "outputs": [],
      "source": [
        "# Shared Layers\n",
        "class SharedNet(nn.Module):\n",
        "    def __init__(self):\n",
        "        super(SharedNet, self).__init__()\n",
        "        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1)\n",
        "        self.pool = nn.MaxPool2d(2, 2)\n",
        "        self.conv2 = nn.Conv2d(16, 32, kernel_size=3, padding=1)\n",
        "\n",
        "    def forward(self, x):\n",
        "        x = self.pool(F.relu(self.conv1(x)))\n",
        "        x = self.pool(F.relu(self.conv2(x)))\n",
        "        shape_before_flatten = x.shape\n",
        "        x = x.view(-1, 32 * shape_before_flatten[2] * shape_before_flatten[3])\n",
        "        return x, shape_before_flatten\n",
        "\n",
        "# Task-Specific Layers\n",
        "class Agent(nn.Module):\n",
        "    def __init__(self, shared_net, output_dim):\n",
        "        super(Agent, self).__init__()\n",
        "        self.shared = shared_net\n",
        "        self.task_specific = nn.Sequential(\n",
        "            nn.Conv2d(32, 64, kernel_size=1, padding=0),\n",
        "            nn.ReLU(),\n",
        "            nn.MaxPool2d(2, 2),\n",
        "            nn.Flatten(),\n",
        "            nn.Linear(64 * 4 * 4, output_dim)\n",
        "        )\n",
        "\n",
        "    def forward(self, x):\n",
        "      x, shape_before_flatten = self.shared(x)\n",
        "      x = x.view(shape_before_flatten)\n",
        "      #print(\"Shape before task-specific layers:\", x.shape)  # Debugging line\n",
        "      x = self.task_specific(x)\n",
        "      return x"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "fxD7tRON3u-f",
        "outputId": "fe2c267e-8bc6-425f-ed8e-bcb9cae01ba1"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to ./data/cifar-10-python.tar.gz\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "100%|██████████| 170498071/170498071 [00:03<00:00, 48618709.92it/s]\n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Extracting ./data/cifar-10-python.tar.gz to ./data\n",
            "Files already downloaded and verified\n"
          ]
        }
      ],
      "source": [
        "# Data Loading and Preprocessing for CIFAR-10\n",
        "transform = transforms.Compose([\n",
        "    transforms.Resize(32),\n",
        "    transforms.ToTensor(),\n",
        "    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))\n",
        "])\n",
        "\n",
        "# CIFAR-10 dataset\n",
        "trainset = datasets.CIFAR10(root='./data', train=True, download=True, transform=transform)\n",
        "testset = datasets.CIFAR10(root='./data', train=False, download=True, transform=transform)\n",
        "\n",
        "# Splitting into two tasks\n",
        "trainset_agent = [[(x, y) for x, y in trainset if y < 5], [(x, y - 5) for x, y in trainset if y >= 5]]\n",
        "\n",
        "# Data Loaders\n",
        "trainloader_agent = []\n",
        "for trainset in trainset_agent:\n",
        "    trainloader_agent.append(DataLoader(trainset, batch_size=32, shuffle=True))\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "q-Rt0Olq3w0n"
      },
      "outputs": [],
      "source": [
        "import math\n",
        "from torch.nn.init import orthogonal_\n",
        "\n",
        "class CustomOpt(torch.optim.Optimizer):\n",
        "    def __init__(self, params, A, E, lr=0.01):\n",
        "        defaults = dict(lr=lr)\n",
        "        self.A = A\n",
        "        self.E = E\n",
        "        self.lr = lr/E\n",
        "\n",
        "        self.weights_store = None\n",
        "        self.psi = None\n",
        "        self.old_psi = None\n",
        "        self.phi = None\n",
        "\n",
        "        self.param_info = None\n",
        "        self.setup_weights = False\n",
        "\n",
        "        super(CustomOpt, self).__init__(params, defaults)\n",
        "\n",
        "    def singular_value(self, p):\n",
        "        try:\n",
        "            sv = math.sqrt(p.shape[0] / p.shape[1])\n",
        "            if p.dim() == 4:\n",
        "                sv /= math.sqrt(p.shape[2] * p.shape[3])\n",
        "            return sv\n",
        "        except:\n",
        "            print()\n",
        "            #print(p.shape)\n",
        "\n",
        "    @torch.no_grad()\n",
        "    def init_weights(self, params):\n",
        "        for group in params:\n",
        "            for p in group['params']:\n",
        "                if p.dim() == 2: orthogonal_(p)\n",
        "                if p.dim() == 4:\n",
        "                    for kx in range(p.shape[2]):\n",
        "                        for ky in range(p.shape[3]):\n",
        "                            orthogonal_(p[:,:,kx,ky])\n",
        "                if p.dim() == 1:\n",
        "                    continue # skip biases\n",
        "                p *= self.singular_value(p)\n",
        "\n",
        "    def setup_params(self):\n",
        "        param_vector = []\n",
        "        grad_vector = []\n",
        "        #param_info = []\n",
        "        all_param_info = []\n",
        "\n",
        "        depth = 0\n",
        "        for group in self.param_groups:\n",
        "            for param in group['params']:\n",
        "                depth += param.numel()\n",
        "\n",
        "        G = 0\n",
        "\n",
        "        lr_vector = []\n",
        "\n",
        "        # lr calculation\n",
        "        for group in self.param_groups:\n",
        "            for param in group['params']:\n",
        "                check = self.singular_value(param)\n",
        "                if check != None:\n",
        "                    G += self.singular_value(param) * param.grad.norm().sum()\n",
        "\n",
        "        # add\n",
        "        #G /= depth\n",
        "\n",
        "        log = math.log(0.5 * (1 + math.sqrt(1 + 4*G)))\n",
        "\n",
        "        for group in self.param_groups:\n",
        "            group_lr_vector = []\n",
        "            for param in group['params']:\n",
        "                #print(param.numel())\n",
        "                check = self.singular_value(param)\n",
        "                if check == None:\n",
        "                    group_lr_vector.append([self.lr] * param.numel())\n",
        "                else:\n",
        "                    factor = self.singular_value(param) / param.grad.norm(keepdim=True)\n",
        "                    temp_lr = 1.0 * log / depth * torch.nan_to_num(factor) # set self.gain to 1.0 for now\n",
        "                    group_lr_vector.append([temp_lr.item()] * param.numel())\n",
        "            group_lr_vector = [item for sublist in group_lr_vector for item in sublist]\n",
        "            lr_vector.append(group_lr_vector)\n",
        "\n",
        "        # original param/grad setup\n",
        "        for group in self.param_groups:\n",
        "            flat_params = [param.view(-1) for param in group['params']]\n",
        "            flat_grads = [param.grad.view(-1) for param in group['params']]\n",
        "            param_info = []\n",
        "            start_idx = 0\n",
        "\n",
        "            for param, flat_param in zip(group['params'], flat_params):\n",
        "                end_idx = start_idx + flat_param.size(0)\n",
        "                param_info.append((param, start_idx, end_idx, param.size()))\n",
        "                start_idx = end_idx\n",
        "\n",
        "            try:\n",
        "                concat_params = torch.cat(flat_params)\n",
        "            except:\n",
        "                print(len(flat_params))\n",
        "                print(flat_params[0])\n",
        "            concat_grads = torch.cat(flat_grads)\n",
        "            param_vector.append(concat_params)\n",
        "            grad_vector.append(concat_grads)\n",
        "            all_param_info.append(param_info)\n",
        "\n",
        "        max_length = max([tensor.size(0) for tensor in param_vector])\n",
        "\n",
        "        param_vector = [torch.nn.functional.pad(tensor, (0, max_length - tensor.size(0))) for tensor in param_vector]\n",
        "        grad_vector = [torch.nn.functional.pad(tensor, (0, max_length - tensor.size(0))) for tensor in grad_vector]\n",
        "\n",
        "        param_matrix = torch.stack(param_vector)\n",
        "        grad_matrix = torch.stack(grad_vector)\n",
        "\n",
        "        return param_matrix, grad_matrix, all_param_info, lr_vector\n",
        "\n",
        "    def putback_params(self, updated_params_matrix, all_param_info):\n",
        "        for group_idx, param_info_list in enumerate(all_param_info):\n",
        "          updated_values_group = updated_params_matrix[group_idx]\n",
        "\n",
        "          for param_info in param_info_list:\n",
        "              param, start_idx, end_idx, original_shape = param_info\n",
        "              updated_values = updated_values_group[start_idx:end_idx]\n",
        "              param.data = updated_values.view(original_shape)\n",
        "\n",
        "    def local_step(self):\n",
        "        # set up params\n",
        "        param_matrix, grad_matrix, param_info, lr_matrix = self.setup_params()\n",
        "\n",
        "        max_len = max(len(sublist) for sublist in lr_matrix)\n",
        "        padded_lists = [sublist + [0] * (max_len - len(sublist)) for sublist in lr_matrix]\n",
        "        lr_matrix = torch.tensor(padded_lists)\n",
        "\n",
        "        if self.param_info == None:\n",
        "            self.param_info = param_info # should always be the same, therefore only assign it once\n",
        "\n",
        "        # if it is the first loop of E, save the current iteration's weights to be used to update phi at the end of the E loop\n",
        "        if self.weights_store == None:\n",
        "            self.weights_store = param_matrix.clone()\n",
        "\n",
        "        # update one E step\n",
        "        # param_matrix = param_matrix - self.lr * grad_matrix\n",
        "        param_matrix = param_matrix - lr_matrix * grad_matrix\n",
        "        self.psi = param_matrix.clone()\n",
        "\n",
        "        # put back params\n",
        "        self.putback_params(param_matrix, param_info)\n",
        "\n",
        "    def correct(self):\n",
        "        if self.old_psi == None:\n",
        "            phi = self.psi.clone()\n",
        "        else:\n",
        "            phi = self.weights_store + self.psi - self.old_psi\n",
        "\n",
        "        # store new old_psi\n",
        "        self.old_psi = self.psi.clone()\n",
        "\n",
        "        # reset weights_store for next epoch\n",
        "        self.weights_store = None\n",
        "\n",
        "        # store phi to update weights\n",
        "        self.phi = phi.clone()\n",
        "\n",
        "    def step(self):\n",
        "        # if it is the first step, intialize the weights\n",
        "        if self.setup_weights == False:\n",
        "            self.init_weights(self.param_groups)\n",
        "            self.setup_weights = True\n",
        "\n",
        "        # update weights+biases\n",
        "        torch_A = torch.from_numpy(self.A).float()\n",
        "        torch_A = 0.5 * (torch_A + torch.eye(torch_A.size(0)))\n",
        "        updated_params_matrix = torch.matmul(torch_A, self.phi)\n",
        "\n",
        "        # put back params\n",
        "        self.putback_params(updated_params_matrix, self.param_info)\n",
        "        return 1"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "SJpfLrp_36lI"
      },
      "outputs": [],
      "source": [
        "# initialize full task-specific networks (N copies of shared + local network architecture)\n",
        "shared_networks = []\n",
        "agent_networks = []\n",
        "for n in range(N):\n",
        "    shared_net = SharedNet()\n",
        "    shared_networks.append(shared_net)\n",
        "    agent_networks.append(Agent(shared_net, 5)) # 5 output classes to demonstrate with CIFAR-10 data"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "h9te0X0p37y9"
      },
      "outputs": [],
      "source": [
        "param_list = []\n",
        "E = 2\n",
        "\n",
        "for network in agent_networks:\n",
        "    param_list.append({'params': network.shared.parameters()})\n",
        "    param_list.append({'params': network.task_specific.parameters()})\n",
        "\n",
        "optimizer = CustomOpt(param_list, A, E, lr=0.05)\n",
        "criterion = nn.CrossEntropyLoss()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "background_save": true,
          "base_uri": "https://localhost:8080/"
        },
        "id": "RNrK2KGx389d",
        "outputId": "d7c1f2f2-4c57-4ed7-b68f-4f3216ce576e"
      },
      "outputs": [],
      "source": [
        "agent1_losses = []\n",
        "agent2_losses = []\n",
        "agent3_losses = []\n",
        "agent4_losses = []\n",
        "\n",
        "agent1_gradients = []\n",
        "agent2_gradients = []\n",
        "agent3_gradients = []\n",
        "agent4_gradients = []\n",
        "\n",
        "num_epochs = 75\n",
        "\n",
        "for epoch in range(num_epochs):\n",
        "    agent1_loss = 0.0\n",
        "    agent2_loss = 0.0\n",
        "    agent3_loss = 0.0\n",
        "    agent4_loss = 0.0\n",
        "    num_batches = 0\n",
        "\n",
        "    val_losses = []  # To store validation losses\n",
        "\n",
        "\n",
        "    for (inputs1, labels1), (inputs2, labels2) in zip(trainloader_agent[0], trainloader_agent[1]):\n",
        "        agent1 = agent_networks[0]\n",
        "        agent2 = agent_networks[1]\n",
        "        agent3 = agent_networks[2]\n",
        "        agent4 = agent_networks[3]\n",
        "\n",
        "        optimizer.zero_grad()\n",
        "\n",
        "        # e loop\n",
        "        for e in range(E):\n",
        "\n",
        "            optimizer.zero_grad()\n",
        "            outputs1 = agent1(inputs1)\n",
        "            outputs2 = agent2(inputs2)\n",
        "            outputs3 = agent3(inputs1)\n",
        "            outputs4 = agent4(inputs2)\n",
        "\n",
        "            loss1 = criterion(outputs1, labels1)\n",
        "            loss2 = criterion(outputs2, labels2)\n",
        "            loss3 = criterion(outputs3, labels1)\n",
        "            loss4 = criterion(outputs4, labels2)\n",
        "\n",
        "            total_loss = loss1 + loss2 + loss3 + loss4\n",
        "            total_loss.backward()\n",
        "\n",
        "            optimizer.local_step()\n",
        "\n",
        "        # correction\n",
        "        optimizer.correct()\n",
        "\n",
        "        # update all weights_i\n",
        "        optimizer.step()\n",
        "\n",
        "        agent1_loss += loss1.item()\n",
        "        agent2_loss += loss2.item()\n",
        "        agent3_loss += loss3.item()\n",
        "        agent4_loss += loss4.item()\n",
        "        num_batches += 1\n",
        "\n",
        "    agent1_loss /= num_batches\n",
        "    agent2_loss /= num_batches\n",
        "    agent1_losses.append(agent1_loss)\n",
        "    agent2_losses.append(agent2_loss)\n",
        "\n",
        "    agent3_loss /= num_batches\n",
        "    agent4_loss /= num_batches\n",
        "    agent3_losses.append(agent3_loss)\n",
        "    agent4_losses.append(agent4_loss)\n",
        "\n",
        "    print(f'Epoch {epoch+1}, Agent1 Loss: {agent1_loss}, Agent2 Loss: {agent2_loss}, Agent3 Loss: {agent3_loss}, Agent4 Loss: {agent4_loss}')\n",
        "\n",
        "# After training is done, plot the losses\n",
        "# plt.figure(figsize=(10, 7))\n",
        "# plt.plot(range(num_epochs), agent1_losses, label='Agent 1 Loss')\n",
        "# plt.plot(range(num_epochs), agent2_losses, label='Agent 2 Loss')\n",
        "# plt.plot(range(num_epochs), agent3_losses, label='Agent 3 Loss')\n",
        "# plt.plot(range(num_epochs), agent4_losses, label='Agent 4 Loss')\n",
        "# plt.xlabel('Epochs')\n",
        "# plt.ylabel('Loss')\n",
        "# plt.legend()\n",
        "# plt.show()\n",
        "\n",
        "# plot log loss instead:\n",
        "plt.figure(figsize=(10, 7))\n",
        "\n",
        "# Take 10*log10 of the losses\n",
        "agent1_losses_log = 10 * np.log10(agent1_losses)\n",
        "agent2_losses_log = 10 * np.log10(agent2_losses)\n",
        "agent3_losses_log = 10 * np.log10(agent3_losses)\n",
        "agent4_losses_log = 10 * np.log10(agent4_losses)\n",
        "\n",
        "# Plot the transformed losses\n",
        "plt.plot(range(num_epochs), agent1_losses_log, label='Agent 1 Loss (10*log10)')\n",
        "plt.plot(range(num_epochs), agent2_losses_log, label='Agent 2 Loss (10*log10)')\n",
        "plt.plot(range(num_epochs), agent3_losses_log, label='Agent 3 Loss (10*log10)')\n",
        "plt.plot(range(num_epochs), agent4_losses_log, label='Agent 4 Loss (10*log10)')\n",
        "\n",
        "plt.xlabel('Epochs')\n",
        "plt.ylabel('Loss (10*log10)')\n",
        "plt.legend()\n",
        "plt.show()\n",
        "\n",
        "# Plotting validation losses\n",
        "# plt.figure(figsize=(10, 7))\n",
        "# agent1_val_losses, agent2_val_losses = zip(*val_losses)\n",
        "# plt.plot(range(num_epochs), agent1_val_losses, label='Validation Agent 1 Loss')\n",
        "# plt.plot(range(num_epochs), agent2_val_losses, label='Validation Agent 2 Loss')\n",
        "# plt.xlabel('Epochs')\n",
        "# plt.ylabel('Validation Loss')\n",
        "# plt.legend()\n",
        "# plt.show()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "oY3unOSqgHD6"
      },
      "outputs": [],
      "source": [
        "import pickle\n",
        "\n",
        "with open(\"local_agd_cifar.pkl\", \"wb\") as file:\n",
        "  pickle.dump((agent1_losses, agent2_losses, agent3_losses, agent4_losses), file)\n"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "provenance": [],
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}