pytorch2paddle.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# X2Paddle快速上手——PyTorch迁移至PaddlePaddle\n",
    "***X2Paddle简介***：X2Paddle支持将Caffe/TensorFlow/ONNX/PyTorch深度学习框架训练得到的模型，迁移至PaddlePaddle模型。   \n",
    "***X2Paddle代码GitHub链接***：[https://github.com/PaddlePaddle/X2Paddle](https://github.com/PaddlePaddle/X2Paddle)  \n",
    "***【注意】***前往GitHub给[X2Paddle](https://github.com/PaddlePaddle/X2Paddle)点击Star，关注项目，即可随时了解X2Paddle的最新进展。  \n",
    "本教程用于帮助用户学习将PyTorch训练后的预测模型迁移至PaddlePaddle框架，以PyTorch版本的[AlexNet](https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py)为例进行详细介绍。  \n",
    "\n",
    "## 安装及准备\n",
    "### 1. 安装X2Paddle\n",
    "***方式一：（推荐）***"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "! git clone https://github.com/PaddlePaddle/X2Paddle.git\n",
    "! cd X2Paddle\n",
    "! git checkout develop\n",
    "! python setup.py install"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "***方式二：***"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "! pip install x2paddle --index https://pypi.Python.org/simple/"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. 安装PyTorch"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "! pip install torch==1.6.0 torchvision==0.7.0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 3. 安装PaddlePaddle"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "! pip install paddlepaddle==2.2.0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 模型迁移\n",
    "### 1. 获取AlexNet模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from torchvision.models import AlexNet\n",
    "from torchvision.models.utils import load_state_dict_from_url\n",
    "\n",
    "torch_model = AlexNet()\n",
    "torch_state_dict = load_state_dict_from_url('https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth')\n",
    "torch_model.load_state_dict(torch_state_dict)\n",
    "torch_model.eval()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. 转换\n",
    "PyTorch到PaddlePaddle的转换需要传入输入的示例，才可以进行转换，以下为构建输入的过程（输入也可为值随机初始化的Tensor）："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "from PIL import Image\n",
    "img = Image.open(\"dog_pt.png\")\n",
    "img = np.array(img).astype(\"float32\") / 255.0\n",
    "img -= [0.485, 0.456, 0.406]\n",
    "img /= [0.229, 0.224, 0.225]\n",
    "img = np.transpose(img, (2, 0, 1))\n",
    "img = np.expand_dims(img, 0)\n",
    "\n",
    "import torch\n",
    "input_tensor = torch.tensor(img)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "***方式一：*** Trace方式  \n",
    "模型输入的shape固定，PyTorch模型基本均支持此方式转换。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from x2paddle.convert import pytorch2paddle\n",
    "save_dir = \"pd_model_trace\"\n",
    "jit_type = \"trace\"\n",
    "pytorch2paddle(module=torch_model, \n",
    "               save_dir=save_dir, \n",
    "               jit_type=jit_type, \n",
    "               input_examples=[input_tensor])\n",
    "# module (torch.nn.Module): PyTorch的Module。\n",
    "# save_dir (str): 转换后模型的保存路径。\n",
    "# jit_type (str): 转换方式，此时为\"trace\"。\n",
    "# input_examples (list[torch.tensor]): torch.nn.Module的输入示例，list的长度必须与输入的长度一致。默认为None。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "***方式二：*** Script方式  \n",
    "模型输入的shape可不固定，由于PyTorch的Script方式可识别的代码格式有限，所以PyTorch模型在此方式下转换的支持度较低。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from x2paddle.convert import pytorch2paddle\n",
    "save_dir = \"pd_model_script\"\n",
    "jit_type = \"script\"\n",
    "pytorch2paddle(module=torch_model, \n",
    "               save_dir=save_dir, \n",
    "               jit_type=jit_type, \n",
    "               input_examples=[input_tensor])\n",
    "# module (torch.nn.Module): PyTorch的Module。\n",
    "# save_dir (str): 转换后模型的保存路径。\n",
    "# jit_type (str): 转换方式，此时为\"script\"。\n",
    "# input_examples (list[torch.tensor]): torch.nn.Module的输入示例，list的长度必须与输入的长度一致。默认为None。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## PaddlePaddle模型使用\n",
    "使用方式一转换的PaddlePaddle预测模型进行预测：  \n",
    "（1）下载ImageNet类别文件"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "! wget https://raw.githubusercontent.com/Lasagne/Recipes/master/examples/resnet50/imagenet_classes.txt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "（2）预测"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import paddle\n",
    "paddle.enable_static()\n",
    "exe = paddle.static.Executor(paddle.CPUPlace())\n",
    "[prog, inputs, outputs] = paddle.static.load_inference_model(path_prefix=\"pd_model_trace/inference_model\", \n",
    "                                                            executor=exe, \n",
    "                                                            model_filename=\"model.pdmodel\",\n",
    "                                                            params_filename=\"model.pdiparams\")\n",
    "print(img.shape)\n",
    "result = exe.run(prog, feed={inputs[0]: img}, fetch_list=outputs)\n",
    "max_index = np.argmax(result)\n",
    "with open('imagenet_classes.txt') as f:\n",
    "    classes = [line.strip() for line in f.readlines()]\n",
    "print(\"The category of dog.jpg is: {}\".format(classes[max_index]))"
   ]
  }
 ],
 "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": "3.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}