{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "ede95fbe",
   "metadata": {},
   "source": [
    "# DeepSolo\n",
    "\n",
    "## 1. 模型简介\n",
    "DeepSolo是一个简洁的类似DETR的基线模型，允许一个具有显式点的解码器同时进行检测和识别。\n",
    "\n",
    "**模型结构**\n",
    "\n",
    "DeepSolo中，编码器在接收到图像特征后，生成由四个Bezier控制点表示的Bezier中心曲线候选和相应的分数，然后，选择前K个评分的候选。对于每个选定的曲线候选，在曲线上均匀采样N个点，这些点的坐标被编码为位置query并将其添加到内容query中形成复合query。接下来，将复合query输入deformable cross-attention解码器收集有用的文本特征。在解码器之后，采用了几个简单的并行预测头（线性层或MLP）将query解码为文本的中心线、边界、script和置信度，从而同时解决检测和识别问题。\n",
    "\n",
    "<div align=center>\n",
    "    <img src=\"./doc/DeepSolo.jpg\"/>\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "aa7b7b67-9d3b-4926-b039-ba9840eefa4d",
   "metadata": {},
   "source": [
    "¶\n",
    "## 2. 环境检查及依赖补全\n",
    "\n",
    "### 2.1 环境检查\n",
    "\n",
    "推荐环境：pytorch=1.13.1 py38\n",
    "推荐环境：dcu=23.04.1\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "925e2b69",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 检查torch版本\n",
    "\n",
    "import torch\n",
    "import torch.utils.cpp_extension\n",
    "version = torch.__version__\n",
    "num = float(version[:version.rfind('.')])\n",
    "assert num >= 1.10\n",
    "device = \"cpu\"\n",
    "\n",
    "# 检查硬件环境\n",
    "if torch.utils.cpp_extension.HIP_HOME:\n",
    "    device = \"dtk\"\n",
    "    !rocm-smi\n",
    "elif torch.utils.cpp_extension.CUDA_HOME:\n",
    "    device = \"cuda\"\n",
    "    !nvidia-smi\n",
    "print(\"pytorch version:\", version)\n",
    "print(\"device =\", device)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "aa7b7b67-9d3b-4926-b039-ba9840eefacc",
   "metadata": {},
   "source": [
    "### 2.2 依赖安装\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9e7d7059",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install -r requirements.txt -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5bf42297",
   "metadata": {},
   "outputs": [],
   "source": [
    "!bash make.sh"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0a7b0e93",
   "metadata": {},
   "outputs": [],
   "source": [
    "!git clone https://github.com/facebookresearch/detectron2.git\n",
    "!python -m pip install -e detectron2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "15f59780",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip3 install -r requirements.txt"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "64c534dd-ad8f-493d-a5e4-435676d4f162",
   "metadata": {},
   "source": [
    "## 3. 素材准备\n",
    "### 3.1 数据集准备\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b1caaec4",
   "metadata": {},
   "source": [
    "项目已经预制了轻量数据simple进行训练验证,请确保当前项目中包含datasets目录且结构如下:\n",
    "\n",
    "```\n",
    "├── datasets\n",
    "│   ├── simple\n",
    "│       ├── test_images\n",
    "│       ├── train_images\n",
    "│       ├── test.json\n",
    "│       └── train.json\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ef0faf15-d9ab-454f-9368-e026372752ad",
   "metadata": {},
   "source": [
    "## 4 训练\n",
    "### 4.1 开始训练\n",
    "\n",
    "根据需求选择单卡或多卡训练"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c305213b",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 单卡训练\n",
    "!export HIP_VISIBLE_DEVICES=4\n",
    "!python tools/train_net.py --config-file configs/simple/train_simple.yaml --num-gpus 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2d95c0e6-4143-40de-b90c-3978c02cb169",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 多卡训练（需要2个加速卡）\n",
    "!export HIP_VISIBLE_DEVICES=4,5,6,7\n",
    "!python tools/train_net.py --config-file configs/simple/train_simple.yaml --num-gpus 4"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5f1986b2-1ea1-45be-8f56-dfaacfba7694",
   "metadata": {},
   "source": [
    "## 5. 推理\n",
    "## 5.1 开始推理\n",
    "\n",
    "提供了一个推理脚本来测试模型，执行下面的脚本来测试模型输出"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "de9d22f1-4764-4c16-a4c9-52e9581669ce",
   "metadata": {},
   "outputs": [],
   "source": [
    "!python demo/demo.py --config-file configs/simple/test_simple.yaml --input datasets/simple/test_images\n",
    "# 推理结果默认保存在test_results文件夹下"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "cfe3acde",
   "metadata": {},
   "source": [
    "## 6. 相关文献和引用\n",
    "https://github.com/ViTAE-Transformer/DeepSolo.git\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.8.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}