docs/source/tutorials/EasyCV图像自监督训练-MAE.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "69debbc2-30e8-4390-805a-cb6ebad35f3f",
   "metadata": {
    "tags": []
   },
   "source": [
    "## EasyCV图像自监督训练-MAE\n",
    "本文将介绍如何利用EasyCV使用自监督算法[MAE](https://arxiv.org/pdf/2111.06377.pdf)进行图像自监督模型的训练\n",
    " \n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b9ef9614-cb63-487f-bafb-c62c90ae607b",
   "metadata": {},
   "source": [
    "## 运行环境要求\n",
    "\n",
    "PAI-Pytorch镜像 or 原生Pytorch1.5+以上环境 GPU机器， 内存32G以上"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f855f736-c08a-44c2-b1c1-33a8e7043864",
   "metadata": {
    "tags": []
   },
   "source": [
    "## 安装依赖包\n",
    "\n",
    "注: 在PAI-DSW docker中无需安装相关依赖，可跳过此部分 在本地notebook环境中执行\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "43e60ee5-f52f-4045-bc69-4a123731a5a5",
   "metadata": {},
   "source": [
    "1、 首先，安装pytorch和对应版本的torchvision，支持Pytorch1.5.1以上版本"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f97b2295-4d67-4a83-8637-ecde48fa3001",
   "metadata": {},
   "outputs": [],
   "source": [
    "# install pytorch and torch vision\n",
    "! conda install --yes pytorch==1.10.0 torchvision==0.11.0 -c pytorch"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ff3ed9f8-703d-416e-8b70-10a8ce029e8d",
   "metadata": {},
   "source": [
    "2、获取torch和cuda版本，安装对应版本的mmcv和nvidia-dali"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c95571d5-3065-40b3-ad53-dc3063db8604",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import os\n",
    "os.environ['CUDA']='cu' + torch.version.cuda.replace('.', '')\n",
    "os.environ['Torch']='torch'+torch.version.__version__.replace('+PAI', '')\n",
    "!echo \"cuda version: $CUDA\"\n",
    "!echo \"pytorch version: $Torch\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ce3dca16-9baf-42f8-94c4-5ea7f087d61e",
   "metadata": {},
   "outputs": [],
   "source": [
    "# install some python deps\n",
    "! pip install mmcv-full==1.4.4 -f https://download.openmmlab.com/mmcv/dist/${CUDA}/${Torch}/index.html\n",
    "! pip install http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/third_party/nvidia_dali_cuda100-0.25.0-1535750-py3-none-manylinux2014_x86_64.whl"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3dee3a99-191a-4515-97dd-80841db43775",
   "metadata": {},
   "source": [
    "3、  安装EasyCV算法包"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7f03f9e9-6029-4918-b8dc-533db9c7fae3",
   "metadata": {},
   "outputs": [],
   "source": [
    "pip install pai-easycv"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3483904d-91dd-44c4-a486-246aa38d4124",
   "metadata": {},
   "source": [
    "4、 简单验证"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4d382373-76a0-4fbc-b09f-5d082aab5104",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "from easycv.apis import *"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "29bb3d55-d00b-453b-9522-c686260e325c",
   "metadata": {},
   "source": [
    "## 数据准备\n",
    "\n",
    "自监督训练只需要提供无标注图片即可进行， 你可以下载[ImageNet](http://www.image-net.org/download-images) 数据，或者使用你自己的图片数据。需要提供一个包含若干图片的文件夹路径`p`，以及一个文件列表，文件列表中是每个图片相对图片目录`p`的路径\n",
    "\n",
    "图片文件夹结构示例如下, 文件夹路径为`./images`\n",
    "\n",
    "```shell\n",
    "images/\n",
    "├── 0001.jpg\n",
    "├── 0002.jpg\n",
    "├── 0003.jpg\n",
    "|...\n",
    "└── 9999.jpg\n",
    "```\n",
    "\n",
    "文件列表内容示例如下\n",
    "```text\n",
    "0001.jpg\n",
    "0002.jpg\n",
    "0003.jpg\n",
    "...\n",
    "9999.jpg\n",
    "```\n",
    "\n",
    "为了快速走通流程，我们也提供了一个小的示例数据集，执行如下命令下载解压"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9492b324-17d9-4963-b7dd-a49f684c54c3",
   "metadata": {},
   "outputs": [],
   "source": [
    "! wget http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/data/imagenet_raw_demo/imagenet_raw_demo.tar.gz && tar -zxf imagenet_raw_demo.tar.gz"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5fdc9dc3-0886-493f-be5b-f858eff72164",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 重命名文件夹\n",
    "! mv imagenet_raw_demo  imagenet_raw"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c9bd7101-3072-417c-ac54-5d7d254b31b4",
   "metadata": {},
   "source": [
    "## 模型训练\n",
    "\n",
    "这个Demo中我们采用[MAE](https://arxiv.org/pdf/2111.06377.pdf)自监督算法训练vit-base主干网络， 下载示例配置文件"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ab669385-8525-4694-8388-148ba1c2753a",
   "metadata": {},
   "outputs": [],
   "source": [
    "! rm -rf mae_vit_base_patch16_8xb64_1600e.py\n",
    "! wget http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/release/doc/easycv/configs/selfsup/mae/mae_vit_base_patch16_8xb64_1600e.py"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2890d267-6b95-47e6-9b51-83402446fa7f",
   "metadata": {},
   "source": [
    "为了缩短训练时间，打开配置文件 `mae_vit_base_patch16_8xb64_1600e.py`，修改`total_epoch`参数为5， 每隔1次迭代打印一次日志。\n",
    "\n",
    "```python\n",
    "# runtime settings\n",
    "total_epochs = 5\n",
    "\n",
    "# log config\n",
    "log_config=dict(interval=1)\n",
    "```\n",
    "\n",
    "正式训练时，建议使用`多机8卡`或`单机8卡`配合该配置文件使用，修改update_interval参数进行梯度累积，确保有效batch_size一致"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "aaf88f46-a578-4dfc-a33f-afa0357f734a",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 查看easycv安装位置\n",
    "import easycv\n",
    "print(easycv.__file__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d01f27fb-b373-4e88-9ac2-01bc35a2f5f2",
   "metadata": {},
   "outputs": [],
   "source": [
    "!python -m torch.distributed.launch --nproc_per_node=1 --master_port=29930 \\\n",
    "/home/pai/lib/python3.6/site-packages/easycv/tools/train.py mae_vit_base_patch16_8xb64_1600e.py --work_dir work_dir/selfsup/jpg/mae --launcher pytorch"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e3aa0e8b-330c-45b2-be84-43f90260588c",
   "metadata": {},
   "source": [
    "## 模型导出\n",
    "对模型的字段进行修改,以便用于fintune任务"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "db2d0dbc-b04e-4244-803b-c03561006c5c",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch \n",
    "weight_path = 'work_dir/selfsup/jpg/mae/epoch_5.pth'\n",
    "state_dict = torch.load(weight_path)['state_dict']\n",
    "state_dict_out = {}\n",
    "for key in state_dict:\n",
    "    state_dict_out[key.replace('encoder.','')] = state_dict[key]\n",
    "torch.save(state_dict_out,weight_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "aad27bb4-e602-4d66-bba5-e379926559f3",
   "metadata": {},
   "source": [
    "## 使用自监督模型进行图像分类fintune\n",
    "下载分类任务示例配置文件"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "78dc417a-2c16-4d7b-a239-e150283e9cac",
   "metadata": {},
   "outputs": [],
   "source": [
    "! rm -rf mae_vit_base_patch16_8xb64_100e_lrdecay065_fintune.py\n",
    "! wget http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/release/doc/easycv/configs/selfsup/mae/mae_vit_base_patch16_8xb64_100e_lrdecay065_fintune.py"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8db1ffd1-6788-41f4-a422-2709903290ad",
   "metadata": {},
   "source": [
    "修改配置文件 `mae_vit_base_patch16_8xb64_1600e.py`，修改`model.pretrained`参数为`work_dir/selfsup/jpg/mae/epoch_5.pth`，为了缩短训练时间，修改`total_epoch`参数为5，每隔1次迭代打印一次日志。\n",
    "\n",
    "```python\n",
    "# runtime settings\n",
    "total_epochs = 5\n",
    "\n",
    "# log config\n",
    "log_config=dict(interval=1)\n",
    "\n",
    "# \n",
    "pretrained='work_dir/selfsup/jpg/mae/epoch_5.pth'\n",
    "```\n",
    "\n",
    "正式训练时，建议使用`多机8卡`或`单机8卡`配合该配置文件使用，修改update_interval参数进行梯度累积，确保有效batch_size一致"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "68e72360-c054-4835-9643-ad0698a97d65",
   "metadata": {},
   "source": [
    "### 分类模型训练\n",
    "这里提供了单卡进行训练和验证集评估的命令"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "defc60df-e8ca-4eac-9923-2cfb1f9fa7f1",
   "metadata": {},
   "outputs": [],
   "source": [
    "!python -m torch.distributed.launch --nproc_per_node=1 --master_port=29930 \\\n",
    "/home/pai/lib/python3.6/site-packages/easycv/tools/train.py mae_vit_base_patch16_8xb64_100e_lrdecay065_fintune.py --work_dir work_dir/selfsup/jpg/mae --launcher pytorch"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fc43e194-d8e7-4796-af3a-1f64663b9744",
   "metadata": {},
   "source": [
    "### 预测\n",
    "参考EasyCV图像分类的demo，对训练好的模型导出并预测"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}