Skip to content

R
Resource

沒 / Resource
与 Fork 源项目一致

从无法访问的项目Fork

通知 1
Star 0
Fork 0
R
Resource

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

未验证 提交 35e06ddc 编写于 作者: F flashrunrun 提交者： GitHub
浏览文件
操作

Merge pull request #9 from MegEngine/courses 

Courses
上级 15fce5e7 75a1a388
展开全部 隐藏空白更改
内联 并排
Showing with 524 addition and 5 deletion
零基础入门旷视天元MegEngine/1.旷视天元MegEngine介绍和快速上手.md
浏览文件 @ 35e06ddc
......@@ -12,7 +12,7 @@
## PPT地址
[1.旷视天元MegEngine介绍和快速上手.pptx](./PPT合集/1.旷视天元MegEngine介绍和快速上手.pptx)
[1.旷视天元MegEngine介绍和快速上手.pptx](./slides/1.旷视天元MegEngine介绍和快速上手.pptx)
## 视频地址
......
零基础入门旷视天元MegEngine/2.模型构建和训练入门.md
浏览文件 @ 35e06ddc
......@@ -12,8 +12,8 @@
## PPT地址
[2.模型构建和训练入门课件.pdf](./PPT合集/2.模型构建和训练入门课件.pdf)<br/>
[2.模型构建和训练入门.ipynb](./PPT合集/2.模型构建和训练入门.ipynb)
[2.模型构建和训练入门课件.pdf](./slides/2.模型构建和训练入门课件.pdf)<br/>
[2.模型构建和训练入门.ipynb](./notebooks/2.模型构建和训练入门.ipynb)
## 视频地址
......
零基础入门旷视天元MegEngine/3.模型构建和训练进阶 I：分类问题.md
浏览文件 @ 35e06ddc
......@@ -10,7 +10,12 @@
## PPT地址
[3.模型构建和训练进阶 I：分类问题.pptx](./PPT合集/3.模型构建和训练进阶%20I：分类问题.pptx)
[3.模型构建和训练进阶 I：分类问题.pptx](./slides/3.模型构建和训练进阶%20I：分类问题.pptx)<br/>
[3.模型构建和训练进阶 I：分类问题1.ipynb](./notebooks/3.模型构建和训练进阶%20I：分类问题1.ipynb)<br/>
[3.模型构建和训练进阶 I：分类问题2.ipynb](./notebooks/3.模型构建和训练进阶%20I：分类问题2.ipynb)<br/>
[3.模型构建和训练进阶 I：分类问题3.ipynb](./notebooks/3.模型构建和训练进阶%20I：分类问题3.ipynb)<br/>
[3.模型构建和训练进阶 I：分类问题4.ipynb](./notebooks/3.模型构建和训练进阶%20I：分类问题4.ipynb)<br/>
## 视频地址
......
零基础入门旷视天元MegEngine/PPT合集/README.md 已删除 100644 → 0
浏览文件 @ 15fce5e7
包含了「零基础入门旷视天元MegEngine」系列课程的PPT文件。
零基础入门旷视天元MegEngine/notebooks/3.模型构建和训练进阶 I：分类问题3.ipynb 0 → 100644
浏览文件 @ 35e06ddc
{
"cells": [
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"import megengine\n",
"import megengine.data as data\n",
"import megengine.data.transform as T\n",
"import megengine.module as M\n",
"import megengine.functional as F\n",
"import megengine.optimizer as optimizer\n",
"import megengine.jit as jit\n",
"import megengine.distributed as dist\n",
"import multiprocessing as mp\n",
"\n",
"F.sync_batch_norm\n",
"M.SyncBatchNorm"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"class BasicBlock(M.Module):\n",
" \"\"\"每个ResNet18的Block都包含两层卷积\"\"\"\n",
" def __init__(self, in_channels, out_channels, stride=1):\n",
" super(BasicBlock, self).__init__()\n",
" # 第一层卷积，接 BN 和 ReLU\n",
" self.conv1 = M.ConvBnRelu2d(\n",
" in_channels=in_channels, out_channels=out_channels,\n",
" kernel_size=3, stride=stride, padding=1)\n",
" # 第二层卷积，只接 BN\n",
" self.conv2 = M.ConvBn2d(\n",
" in_channels=out_channels, out_channels=out_channels,\n",
" kernel_size=3, stride=1, padding=1)\n",
" # 残差连接，当输入输出不一致/需要下采样时，用 ConvBn 实现变换\n",
" if in_channels == out_channels and stride == 1:\n",
" self.res_conn = M.Identity()\n",
" else:\n",
" self.res_conn = M.ConvBn2d(\n",
" in_channels=in_channels, out_channels=out_channels,\n",
" kernel_size=1, stride=stride)\n",
" \n",
" def forward(self, x):\n",
" identity = x\n",
" x = self.conv1(x)\n",
" x = self.conv2(x)\n",
" x = x + self.res_conn(identity)\n",
" return F.relu(x)"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"class ResNet18(M.Module):\n",
" def __init__(self):\n",
" self.conv1 = M.ConvBnRelu2d(in_channels=3, out_channels=64,\n",
" kernel_size=3, padding=1)\n",
" # 8 个 BasicBlock，3 次下采样(stride=2)，共 8x2=16 层卷积\n",
" self.blocks = M.Sequential(\n",
" BasicBlock(64, 64),\n",
" BasicBlock(64, 64),\n",
" BasicBlock(64, 128, stride=2),\n",
" BasicBlock(128, 128),\n",
" BasicBlock(128, 256, stride=2),\n",
" BasicBlock(256, 256),\n",
" BasicBlock(256, 512, stride=2),\n",
" BasicBlock(512, 512),\n",
" )\n",
" # 全连接分类器，输出维度为 10 类的预测\n",
" self.classifier = M.Sequential(\n",
" M.Dropout(0.2),\n",
" M.Linear(512, 10)\n",
" )\n",
" \n",
" def forward(self, x):\n",
" # 1. 特征提取，输入为 Nx3x32x32 的图片，输出为 Nx512x4x4的张量(Tensor)\n",
" x = self.conv1(x)\n",
" x = self.blocks(x)\n",
" # 2. 4x4平均池化(Average Pooling)\n",
" x = F.avg_pool2d(x, 4)\n",
" x = F.flatten(x, 1)\n",
" # 3. 分类预测\n",
" x = self.classifier(x)\n",
" return x"
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
"def training():\n",
" # megengine内置CIFAR10的数据集\n",
" dataset = data.dataset.CIFAR10(root=\"/data\", train=True)\n",
" \n",
" # 构造数据生产线\n",
" dataloader = data.DataLoader(\n",
" dataset,\n",
" sampler=data.RandomSampler(dataset, batch_size=64, drop_last=True),\n",
" transform=T.Compose([\n",
" T.RandomHorizontalFlip(),\n",
" T.Normalize(mean=0., std=255.), # f(x) = (x - mean) / std\n",
" T.ToMode(\"CHW\"),\n",
" ])\n",
" )\n",
" \n",
" # 构造网络与输入\n",
" model = ResNet18()\n",
" image = megengine.tensor(dtype=\"float32\")\n",
" label = megengine.tensor(dtype=\"int32\")\n",
" \n",
" # 构造网络优化器\n",
" opt = optimizer.SGD(model.parameters(), lr=0.005, momentum=0.9, weight_decay=1e-4)\n",
" \n",
" # 构造静态的计算图以充分发挥性能\n",
" @jit.trace\n",
" def train_func(image, label):\n",
" # 前传\n",
" loglikelihood = model(image)\n",
" loss = F.cross_entropy_with_softmax(loglikelihood, label)\n",
" accuracy = F.accuracy(loglikelihood, label)\n",
"\n",
" # 反传并更新网络参数\n",
" opt.zero_grad()\n",
" opt.backward(loss)\n",
" opt.step()\n",
" return loss, accuracy\n",
" \n",
" for epoch in range(90):\n",
" # 训练一个epoch == 遍历一次训练数据集\n",
" for i, batch_data in enumerate(dataloader):\n",
" # 进行一次迭代\n",
" image.set_value(batch_data[0])\n",
" label.set_value(batch_data[1])\n",
" \n",
" loss, acc1 = train_func(image, label)\n",
"\n",
" if i % 50 == 0:\n",
" print(\"epoch\", epoch, \"step\", i, \"loss\", loss, \"acc@1\", acc1)"
]
},
{
"cell_type": "code",
"execution_count": 15,
"metadata": {},
"outputs": [],
"source": [
"def worker(rank):\n",
" # 每个子进程需要初始化分布式进程组\n",
" dist.init_process_group(\n",
" master_ip=\"localhost\", # 主节点的IP地址。单机多卡的情况下可以简单设为localhost\n",
" master_port=2233, # 进行通信的可用的端口号，0-65535，注意不能被其他进程占用\n",
" world_size=8, # 参与任务的进程总数（即总进程个数，也等于显卡个数）\n",
" rank=rank, # 当前进程的进程号（即第几个进程）\n",
" dev=rank, # 第几个进程用第几块显卡\n",
" )\n",
" print(\"init process\", rank)\n",
" # 开始训练\n",
" training()\n",
"\n",
"def main():\n",
" # 在一台机器上启动 8 个子进程（因为总共有 8 块显卡）\n",
" # 关于 multiprocessing 的使用方法参见python官方文档\n",
" for rank in range(8):\n",
" p = mp.Process(target=worker, args=(rank,))\n",
" p.start()"
]
},
{
"cell_type": "code",
"execution_count": 16,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"init process 0\n",
"init process 1\n",
"init process 2\n",
"init process 3\n",
"init process 4\n",
"init process 5\n",
"init process 6\n",
"init process 7\n",
"epoch 0 step 0 loss Tensor([2.5735], device=gpu6:0) acc@1 Tensor([0.0781], device=gpu6:0)\n",
"epoch 0 step 0 loss Tensor([2.4922], device=gpu1:0) acc@1 Tensor([0.0625], device=gpu1:0)\n",
"epoch 0 step 0 loss Tensor([2.5264], device=gpu3:0) acc@1 Tensor([0.1406], device=gpu3:0)\n",
"epoch 0 step 0 loss Tensor([2.7726], device=gpu5:0) acc@1 Tensor([0.], device=gpu5:0)\n",
"epoch 0 step 0 loss Tensor([2.6011], device=gpu2:0) acc@1 Tensor([0.0938], device=gpu2:0)\n",
"epoch 0 step 0 loss Tensor([2.5178], device=gpu7:0) acc@1 Tensor([0.0312], device=gpu7:0)\n",
"epoch 0 step 0 loss Tensor([2.554], device=gpu4:0) acc@1 Tensor([0.1094], device=gpu4:0)\n",
"epoch 0 step 0 loss Tensor([2.5615], device=gpu0:0) acc@1 Tensor([0.1406], device=gpu0:0)\n",
"epoch 0 step 50 loss Tensor([1.7417], device=gpu6:0) acc@1 Tensor([0.4219], device=gpu6:0)\n",
"epoch 0 step 50 loss Tensor([1.521], device=gpu5:0) acc@1 Tensor([0.5312], device=gpu5:0)\n",
"epoch 0 step 50 loss Tensor([1.6334], device=gpu1:0) acc@1 Tensor([0.3281], device=gpu1:0)\n",
"epoch 0 step 50 loss Tensor([1.6283], device=gpu4:0) acc@1 Tensor([0.375], device=gpu4:0)\n",
"epoch 0 step 50 loss Tensor([1.5762], device=gpu2:0) acc@1 Tensor([0.375], device=gpu2:0)\n",
"epoch 0 step 50 loss Tensor([1.5677], device=gpu0:0) acc@1 Tensor([0.4062], device=gpu0:0)\n",
"epoch 0 step 50 loss Tensor([1.6616], device=gpu3:0) acc@1 Tensor([0.3594], device=gpu3:0)\n",
"epoch 0 step 50 loss Tensor([1.6297], device=gpu7:0) acc@1 Tensor([0.4531], device=gpu7:0)\n",
"epoch 1 step 0 loss Tensor([1.2938], device=gpu7:0) acc@1 Tensor([0.5], device=gpu7:0)\n",
"epoch 1 step 0 loss Tensor([1.5492], device=gpu1:0) acc@1 Tensor([0.4219], device=gpu1:0)\n",
"epoch 1 step 0 loss Tensor([1.3215], device=gpu0:0) acc@1 Tensor([0.4531], device=gpu0:0)\n",
"epoch 1 step 0 loss Tensor([1.5797], device=gpu5:0) acc@1 Tensor([0.5], device=gpu5:0)\n",
"epoch 1 step 0 loss Tensor([1.5253], device=gpu6:0) acc@1 Tensor([0.4219], device=gpu6:0)\n",
"epoch 1 step 0 loss Tensor([1.5287], device=gpu4:0) acc@1 Tensor([0.375], device=gpu4:0)\n",
"epoch 1 step 0 loss Tensor([1.4993], device=gpu3:0) acc@1 Tensor([0.4062], device=gpu3:0)\n",
"epoch 1 step 0 loss Tensor([1.5306], device=gpu2:0) acc@1 Tensor([0.4219], device=gpu2:0)\n",
"epoch 1 step 50 loss Tensor([1.0871], device=gpu6:0) acc@1 Tensor([0.5938], device=gpu6:0)\n",
"epoch 1 step 50 loss Tensor([1.2019], device=gpu5:0) acc@1 Tensor([0.5625], device=gpu5:0)\n",
"epoch 1 step 50 loss Tensor([1.449], device=gpu1:0) acc@1 Tensor([0.5], device=gpu1:0)\n",
"epoch 1 step 50 loss Tensor([1.3241], device=gpu2:0) acc@1 Tensor([0.5625], device=gpu2:0)\n",
"epoch 1 step 50 loss Tensor([1.3581], device=gpu7:0) acc@1 Tensor([0.4531], device=gpu7:0)\n",
"epoch 1 step 50 loss Tensor([1.3815], device=gpu3:0) acc@1 Tensor([0.5], device=gpu3:0)\n",
"epoch 1 step 50 loss Tensor([1.5198], device=gpu4:0) acc@1 Tensor([0.375], device=gpu4:0)\n",
"epoch 1 step 50 loss Tensor([1.2055], device=gpu0:0) acc@1 Tensor([0.5156], device=gpu0:0)\n",
"epoch 2 step 0 loss Tensor([1.313], device=gpu1:0) acc@1 Tensor([0.5156], device=gpu1:0)\n",
"epoch 2 step 0 loss Tensor([1.0568], device=gpu6:0) acc@1 Tensor([0.6094], device=gpu6:0)\n",
"epoch 2 step 0 loss Tensor([1.4929], device=gpu5:0) acc@1 Tensor([0.5625], device=gpu5:0)\n",
"epoch 2 step 0 loss Tensor([1.1126], device=gpu7:0) acc@1 Tensor([0.5938], device=gpu7:0)\n",
"epoch 2 step 0 loss Tensor([1.1693], device=gpu4:0) acc@1 Tensor([0.5781], device=gpu4:0)\n",
"epoch 2 step 0 loss Tensor([1.1453], device=gpu2:0) acc@1 Tensor([0.5312], device=gpu2:0)\n",
"epoch 2 step 0 loss Tensor([1.235], device=gpu0:0) acc@1 Tensor([0.4844], device=gpu0:0)\n",
"epoch 2 step 0 loss Tensor([1.1699], device=gpu3:0) acc@1 Tensor([0.5156], device=gpu3:0)\n",
"epoch 2 step 50 loss Tensor([1.3884], device=gpu6:0) acc@1 Tensor([0.5], device=gpu6:0)\n",
"epoch 2 step 50 loss Tensor([0.9665], device=gpu7:0) acc@1 Tensor([0.5938], device=gpu7:0)\n",
"epoch 2 step 50 loss Tensor([1.0038], device=gpu1:0) acc@1 Tensor([0.5938], device=gpu1:0)\n",
"epoch 2 step 50 loss Tensor([1.1243], device=gpu0:0) acc@1 Tensor([0.5781], device=gpu0:0)\n",
"epoch 2 step 50 loss Tensor([1.014], device=gpu5:0) acc@1 Tensor([0.7188], device=gpu5:0)\n",
"epoch 2 step 50 loss Tensor([1.077], device=gpu4:0) acc@1 Tensor([0.5156], device=gpu4:0)\n",
"epoch 2 step 50 loss Tensor([1.1881], device=gpu2:0) acc@1 Tensor([0.6562], device=gpu2:0)\n",
"epoch 2 step 50 loss Tensor([1.0908], device=gpu3:0) acc@1 Tensor([0.5312], device=gpu3:0)\n",
"epoch 3 step 0 loss Tensor([1.1245], device=gpu1:0) acc@1 Tensor([0.5625], device=gpu1:0)\n",
"epoch 3 step 0 loss Tensor([1.0347], device=gpu5:0) acc@1 Tensor([0.6875], device=gpu5:0)\n",
"epoch 3 step 0 loss Tensor([0.9063], device=gpu7:0) acc@1 Tensor([0.6875], device=gpu7:0)\n",
"epoch 3 step 0 loss Tensor([1.046], device=gpu3:0) acc@1 Tensor([0.6562], device=gpu3:0)\n",
"epoch 3 step 0 loss Tensor([0.961], device=gpu6:0) acc@1 Tensor([0.7031], device=gpu6:0)\n",
"epoch 3 step 0 loss Tensor([0.7993], device=gpu0:0) acc@1 Tensor([0.7188], device=gpu0:0)\n",
"epoch 3 step 0 loss Tensor([0.8532], device=gpu2:0) acc@1 Tensor([0.6406], device=gpu2:0)\n",
"epoch 3 step 0 loss Tensor([1.0973], device=gpu4:0) acc@1 Tensor([0.6094], device=gpu4:0)\n",
"epoch 3 step 50 loss Tensor([0.8748], device=gpu7:0) acc@1 Tensor([0.7344], device=gpu7:0)\n",
"epoch 3 step 50 loss Tensor([0.9524], device=gpu6:0) acc@1 Tensor([0.6875], device=gpu6:0)\n",
"epoch 3 step 50 loss Tensor([1.0182], device=gpu1:0) acc@1 Tensor([0.7031], device=gpu1:0)\n",
"epoch 3 step 50 loss Tensor([0.9418], device=gpu4:0) acc@1 Tensor([0.7188], device=gpu4:0)\n",
"epoch 3 step 50 loss Tensor([0.9911], device=gpu2:0) acc@1 Tensor([0.6719], device=gpu2:0)\n",
"epoch 3 step 50 loss Tensor([1.0507], device=gpu5:0) acc@1 Tensor([0.6562], device=gpu5:0)\n",
"epoch 3 step 50 loss Tensor([0.9237], device=gpu3:0) acc@1 Tensor([0.6562], device=gpu3:0)\n",
"epoch 3 step 50 loss Tensor([0.722], device=gpu0:0) acc@1 Tensor([0.8281], device=gpu0:0)\n",
"epoch 4 step 0 loss Tensor([0.8674], device=gpu7:0) acc@1 Tensor([0.7188], device=gpu7:0)\n",
"epoch 4 step 0 loss Tensor([1.0817], device=gpu5:0) acc@1 Tensor([0.5938], device=gpu5:0)\n",
"epoch 4 step 0 loss Tensor([1.0578], device=gpu1:0) acc@1 Tensor([0.6094], device=gpu1:0)\n",
"epoch 4 step 0 loss Tensor([0.8786], device=gpu6:0) acc@1 Tensor([0.7031], device=gpu6:0)\n",
"epoch 4 step 0 loss Tensor([1.2348], device=gpu2:0) acc@1 Tensor([0.5156], device=gpu2:0)\n",
"epoch 4 step 0 loss Tensor([0.9356], device=gpu4:0) acc@1 Tensor([0.7188], device=gpu4:0)\n",
"epoch 4 step 0 loss Tensor([0.933], device=gpu0:0) acc@1 Tensor([0.6562], device=gpu0:0)\n",
"epoch 4 step 0 loss Tensor([0.7652], device=gpu3:0) acc@1 Tensor([0.7188], device=gpu3:0)\n",
"epoch 4 step 50 loss Tensor([0.7796], device=gpu6:0) acc@1 Tensor([0.7344], device=gpu6:0)\n",
"epoch 4 step 50 loss Tensor([0.8349], device=gpu5:0) acc@1 Tensor([0.7031], device=gpu5:0)\n",
"epoch 4 step 50 loss Tensor([0.9181], device=gpu4:0) acc@1 Tensor([0.7344], device=gpu4:0)\n",
"epoch 4 step 50 loss Tensor([0.9559], device=gpu3:0) acc@1 Tensor([0.5781], device=gpu3:0)\n",
"epoch 4 step 50 loss Tensor([0.5912], device=gpu2:0) acc@1 Tensor([0.8281], device=gpu2:0)\n",
"epoch 4 step 50 loss Tensor([0.9764], device=gpu0:0) acc@1 Tensor([0.6562], device=gpu0:0)\n",
"epoch 4 step 50 loss Tensor([0.9125], device=gpu7:0) acc@1 Tensor([0.6875], device=gpu7:0)\n",
"epoch 4 step 50 loss Tensor([0.7488], device=gpu1:0) acc@1 Tensor([0.7969], device=gpu1:0)\n",
"epoch 5 step 0 loss Tensor([0.5704], device=gpu5:0) acc@1 Tensor([0.7812], device=gpu5:0)\n",
"epoch 5 step 0 loss Tensor([0.9098], device=gpu7:0) acc@1 Tensor([0.6719], device=gpu7:0)\n",
"epoch 5 step 0 loss Tensor([0.9084], device=gpu2:0) acc@1 Tensor([0.7188], device=gpu2:0)\n",
"epoch 5 step 0 loss Tensor([0.8209], device=gpu3:0) acc@1 Tensor([0.7188], device=gpu3:0)\n",
"epoch 5 step 0 loss Tensor([0.8147], device=gpu4:0) acc@1 Tensor([0.6719], device=gpu4:0)\n",
"epoch 5 step 0 loss Tensor([0.8523], device=gpu6:0) acc@1 Tensor([0.7031], device=gpu6:0)\n",
"epoch 5 step 0 loss Tensor([0.9633], device=gpu0:0) acc@1 Tensor([0.7031], device=gpu0:0)\n",
"epoch 5 step 0 loss Tensor([1.0272], device=gpu1:0) acc@1 Tensor([0.6406], device=gpu1:0)\n"
]
}
],
"source": [
"main()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "8GPU",
"language": "python3",
"name": "8gpu"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.9"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
零基础入门旷视天元MegEngine/notebooks/3.模型构建和训练进阶 I：分类问题4.ipynb 0 → 100644
浏览文件 @ 35e06ddc
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import megengine\n",
"import megengine.data as data\n",
"import megengine.data.transform as T\n",
"import megengine.module as M\n",
"import megengine.functional as F\n",
"import megengine.optimizer as optimizer\n",
"import megengine.jit as jit"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"class BasicBlock(M.Module):\n",
" \"\"\"每个ResNet18的Block都包含两层卷积\"\"\"\n",
" def __init__(self, in_channels, out_channels, stride=1):\n",
" super(BasicBlock, self).__init__()\n",
" # 第一层卷积，接 BN 和 ReLU\n",
" self.conv1 = M.ConvBnRelu2d(\n",
" in_channels=in_channels, out_channels=out_channels,\n",
" kernel_size=3, stride=stride, padding=1)\n",
" # 第二层卷积，只接 BN\n",
" self.conv2 = M.ConvBn2d(\n",
" in_channels=out_channels, out_channels=out_channels,\n",
" kernel_size=3, stride=1, padding=1)\n",
" # 残差连接，当输入输出不一致/需要下采样时，用 ConvBn 实现变换\n",
" if in_channels == out_channels and stride == 1:\n",
" self.res_conn = M.Identity()\n",
" else:\n",
" self.res_conn = M.ConvBn2d(\n",
" in_channels=in_channels, out_channels=out_channels,\n",
" kernel_size=1, stride=stride)\n",
" \n",
" def forward(self, x):\n",
" identity = x\n",
" x = self.conv1(x)\n",
" x = self.conv2(x)\n",
" x = x + self.res_conn(identity)\n",
" return F.relu(x)"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"class ResNet18(M.Module):\n",
" def __init__(self):\n",
" self.conv1 = M.ConvBnRelu2d(in_channels=3, out_channels=64,\n",
" kernel_size=3, padding=1)\n",
" # 8 个 BasicBlock，3 次下采样(stride=2)，共 8x2=16 层卷积\n",
" self.blocks = M.Sequential(\n",
" BasicBlock(64, 64),\n",
" BasicBlock(64, 64),\n",
" BasicBlock(64, 128, stride=2),\n",
" BasicBlock(128, 128),\n",
" BasicBlock(128, 256, stride=2),\n",
" BasicBlock(256, 256),\n",
" BasicBlock(256, 512, stride=2),\n",
" BasicBlock(512, 512),\n",
" )\n",
" # 全连接分类器，输出维度为 10 类的预测\n",
" self.classifier = M.Sequential(\n",
" M.Dropout(0.2),\n",
" M.Linear(512, 10)\n",
" )\n",
" \n",
" def forward(self, x):\n",
" # 1. 特征提取，输入为 Nx3x32x32 的图片，输出为 Nx512x4x4的张量(Tensor)\n",
" x = self.conv1(x)\n",
" x = self.blocks(x)\n",
" # 2. 4x4平均池化(Average Pooling)\n",
" x = F.avg_pool2d(x, 4)\n",
" x = F.flatten(x, 1)\n",
" # 3. 分类预测\n",
" x = self.classifier(x)\n",
" return x"
]
},
{
"cell_type": "code",
"execution_count": 22,
"metadata": {},
"outputs": [],
"source": [
"def testing():\n",
" # megengine内置CIFAR10的数据集(train=False)\n",
" dataset = data.dataset.CIFAR10(root=\"/data\", train=False)\n",
" \n",
" # 构造数据生产线\n",
" dataloader = data.DataLoader(\n",
" dataset,\n",
" # Random换成Sequential, drop_last=False不漏掉任何一张图片\n",
" sampler=data.SequentialSampler(dataset, batch_size=64, drop_last=False),\n",
" transform=T.Compose([\n",
" # T.RandomHorizontalFlip(), 测试的时候不进行数据增广\n",
" T.Normalize(mean=0., std=255.), # f(x) = (x - mean) / std\n",
" T.ToMode(\"CHW\"),\n",
" ])\n",
" )\n",
" \n",
" # 构造网络与输入\n",
" model = ResNet18()\n",
" image = megengine.tensor(dtype=\"float32\")\n",
" label = megengine.tensor(dtype=\"int32\")\n",
" \n",
" state_dict = megengine.load(\"checkpoint.pkl\")\n",
" model.load_state_dict(state_dict)\n",
" \n",
" # 构造静态的计算图以充分发挥性能\n",
" @jit.trace\n",
" def test_func(image, label):\n",
" # 前传\n",
" loglikelihood = model(image)\n",
" accuracy = F.accuracy(loglikelihood, label)\n",
" return accuracy\n",
" \n",
" # 遍历一次测试数据集\n",
" correct = 0\n",
" for i, batch_data in enumerate(dataloader):\n",
" image.set_value(batch_data[0])\n",
" label.set_value(batch_data[1])\n",
"\n",
" acc1 = test_func(image, label)\n",
" correct += acc1.item() * label.shape[0]\n",
" \n",
" if i % 50 == 0:\n",
" print(\"step\", i, \"acc@1\", acc1)\n",
" print(\"Final accuracy =\", correct / 10000 * 100, \"%\")"
]
},
{
"cell_type": "code",
"execution_count": 23,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"step 0 acc@1 Tensor([0.5938])\n",
"step 50 acc@1 Tensor([0.6719])\n",
"step 100 acc@1 Tensor([0.5938])\n",
"step 150 acc@1 Tensor([0.7188])\n",
"Final accuracy = 63.21 %\n"
]
}
],
"source": [
"testing()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "1GPU",
"language": "python",
"name": "1gpu"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.9"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
零基础入门旷视天元MegEngine/notebooks/README.md 0 → 100644
浏览文件 @ 35e06ddc
包含了零基础入门旷视天元MegEngine课程的notebooks资源
零基础入门旷视天元MegEngine/slides/README.md 0 → 100644
浏览文件 @ 35e06ddc
包含零基础入门旷视天元MegEngine课程的PPT资源
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或