济南微网站建设,网站设计 版权,短裙怎么做视频网站,ui一般用什么软件做前言#xff1a;本篇主要偏向图像分类实战部分#xff0c;使用MMclassification工具进行代码应用#xff0c;最后对水果分类进行实战演示#xff0c;本次环境和代码配置部分省略#xff0c;具体内容建议参考前一篇文章#xff1a;计算机视觉框架OpenMMLab开源学习#x… 前言本篇主要偏向图像分类实战部分使用MMclassification工具进行代码应用最后对水果分类进行实战演示本次环境和代码配置部分省略具体内容建议参考前一篇文章计算机视觉框架OpenMMLab开源学习二图像分类 计算机视觉框架OpenMMLab开源学习三图像分类实战
一、安装OpenMMLab v2.0
Step 1. Install MMCV
mim install mmcv2.0.0rc0
Step 2. Install MMClassification and MMDetection
mim install mmcls1.0.0rc0 mmdet3.0.0rc0代码模版讲解
model dict(typeImageClassifier, # 分类器类型backbonedict(typeResNet, # 主干网络类型depth50, # 主干网网络深度 ResNet 一般有18, 34, 50, 101, 152 可以选择num_stages4, # 主干网络状态(stages)的数目这些状态产生的特征图作为后续的 head 的输入。out_indices(3, ), # 输出的特征图输出索引。越远离输入图像索引越大frozen_stages-1, # 网络微调时冻结网络的stage训练时不执行反相传播算法若num_stages4backbone包含stem 与 4 个 stages。frozen_stages为-1时不冻结网络 为0时冻结 stem 为1时冻结 stem 和 stage1 为4时冻结整个backbonestylepytorch), # 主干网络的风格pytorch 意思是步长为2的层为 3x3 卷积 caffe 意思是步长为2的层为 1x1 卷积。neckdict(typeGlobalAveragePooling), # 颈网络类型headdict(typeLinearClsHead, # 线性分类头num_classes1000, # 输出类别数这与数据集的类别数一致in_channels2048, # 输入通道数这与 neck 的输出通道一致lossdict(typeCrossEntropyLoss, loss_weight1.0), # 损失函数配置信息topk(1, 5), # 评估指标Top-k 准确率 这里为 top1 与 top5 准确率))
二、Pytorch图像分类任务
本次任务训练数据为FashionMNIST完整代码如下
# https://pytorch.org/tutorials/beginner/basics/quickstart_tutorial.htmlimport torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor# Training## Construct Dataset and Dataloader
training_data datasets.FashionMNIST(rootdata,trainTrue,downloadTrue,transformToTensor(),
)
test_data datasets.FashionMNIST(rootdata,trainFalse,downloadTrue,transformToTensor(),
)batch_size 64train_dataloader DataLoader(training_data, batch_sizebatch_size)
test_dataloader DataLoader(test_data, batch_sizebatch_size)## Define model
class NeuralNetwork(nn.Module):def __init__(self):super(NeuralNetwork, self).__init__()self.flatten nn.Flatten()self.linear_relu_stack nn.Sequential(nn.Linear(28*28, 512),nn.ReLU(),nn.Linear(512, 512),nn.ReLU(),nn.Linear(512, 10))def forward(self, x):x self.flatten(x)logits self.linear_relu_stack(x)return logitsdevice cuda if torch.cuda.is_available() else cpu
model NeuralNetwork().to(device)## Define loss function and Optimizer
loss_fn nn.CrossEntropyLoss()
optimizer torch.optim.SGD(model.parameters(), lr1e-3)## Inner loop for training
def train(dataloader, model, loss_fn, optimizer):size len(dataloader.dataset)model.train()for batch, (X, y) in enumerate(dataloader):X, y X.to(device), y.to(device)# Compute prediction errorpred model(X)loss loss_fn(pred, y)# Backpropagationoptimizer.zero_grad()loss.backward()optimizer.step()# Output Logsif batch % 100 0:loss, current loss.item(), batch * len(X)print(floss: {loss:7f} [{current:5d}/{size:5d}])## Inner loop for test
def test(dataloader, model, loss_fn):size len(dataloader.dataset)num_batches len(dataloader)model.eval()test_loss, correct 0, 0with torch.no_grad():for X, y in dataloader:X, y X.to(device), y.to(device)pred model(X)test_loss loss_fn(pred, y).item()correct (pred.argmax(1) y).type(torch.float).sum().item()test_loss / num_batchescorrect / sizeprint(fTest Error: \n Accuracy: {(100*correct):0.1f}%, Avg loss: {test_loss:8f} \n)## Launch training / test loops#
epochs 5
for t in range(epochs):print(fEpoch {t1}\n-------------------------------)train(train_dataloader, model, loss_fn, optimizer)test(test_dataloader, model, loss_fn)
print(Done!)## Saving Modelstorch.save(model.state_dict(), model.pth)
print(Saved PyTorch Model State to model.pth)# Deployment## Loading Models
model NeuralNetwork()
model.load_state_dict(torch.load(model.pth))# Predict new imagesclasses [T-shirt/top,Trouser,Pullover,Dress,Coat,Sandal,Shirt,Sneaker,Bag,Ankle boot,
]model.eval()
x, y test_data[0][0], test_data[0][1]
with torch.no_grad():pred model(x)predicted, actual classes[pred[0].argmax(0)], classes[y]print(fPredicted: {predicted}, Actual: {actual})
三、利用MMClassification提供的预训练模型推理
安装环境
pip install openmim, mmengine
mim install mmcv-full mmcls
Inference using high-level API
from mmcls.apis import init_model, inference_modelmodel init_model(mobilenet-v2_8xb32_in1k.py, mobilenet_v2_batch256_imagenet_20200708-3b2dc3af.pth, devicecuda:0)
load checkpoint from local path: mobilenet_v2_batch256_imagenet_20200708-3b2dc3af.pth
result inference_model(model, banana.png)
result
{pred_label: 954, pred_score: 0.9999284744262695, pred_class: banana}
from mmcls.apis import show_result_pyplotshow_result_pyplot(model, banana.png, result) PyTorch codes under the hood
Let write some raw PyTorch codes to do the same thing.
These are actual codes wrapped in high-level APIs.
construct an ImageClassifier
Note: current implementation only allow configs of backbone, neck and classification head instead of Python objects.
from mmcls.models import ImageClassifierclassifier ImageClassifier(backbonedict(typeMobileNetV2, widen_factor1.0),neckdict(typeGlobalAveragePooling),headdict(typeLinearClsHead,num_classes1000,in_channels1280)
)Load trained parameters
import torchckpt torch.load(mobilenet_v2_batch256_imagenet_20200708-3b2dc3af.pth)
classifier.load_state_dict(ckpt[state_dict])Construct data preprocessing pipeline
Important: A models work only if image preprocessing pipelines is correct.
from mmcls.datasets.pipelines import Composetest_pipeline Compose([dict(typeLoadImageFromFile),dict(typeResize, size(256, -1), backendpillow),dict(typeCenterCrop, crop_size224),dict(typeNormalize,mean[123.675, 116.28, 103.53],std[58.395, 57.12, 57.375],to_rgbTrue),dict(typeImageToTensor, keys[img]),dict(typeCollect, keys[img])
])data dict(img_infodict(filenamebanana.png), img_prefixNone)
data test_pipeline(data)
data{img_metas: DataContainer({filename: banana.png, ori_filename: banana.png, ori_shape: (403, 393, 3), img_shape: (224, 224, 3), img_norm_cfg: {mean: array([123.675, 116.28 , 103.53 ], dtypefloat32), std: array([58.395, 57.12 , 57.375], dtypefloat32), to_rgb: True}}),img: tensor([[[ 0.3309, 0.2967, 0.3138, ..., 2.0263, 2.0092, 1.9920],[ 0.3481, 0.3309, 0.2282, ..., 2.0263, 2.0092, 1.9920],[ 0.2796, 0.2967, 0.2967, ..., 1.9920, 2.0263, 1.9749],...,[ 0.1939, 0.1768, 0.2282, ..., 0.3994, 0.3309, 0.3823],[ 0.1426, 0.1254, 0.2111, ..., 0.5878, 0.5364, 0.5536],[-0.0116, -0.0801, 0.1597, ..., 0.5707, 0.5536, 0.5364]],[[ 0.3803, 0.3803, 0.3803, ..., 2.1660, 2.1485, 2.1134],[ 0.4153, 0.4153, 0.3102, ..., 2.1835, 2.1310, 2.1134],[ 0.3452, 0.3803, 0.3803, ..., 2.1134, 2.1485, 2.1134],...,[ 0.2752, 0.2577, 0.3102, ..., 0.5028, 0.4328, 0.4328],[ 0.2227, 0.1877, 0.3102, ..., 0.6604, 0.6254, 0.5728],[ 0.0301, -0.0049, 0.2402, ..., 0.6604, 0.6254, 0.5728]],[[ 0.5485, 0.5485, 0.5485, ..., 2.3437, 2.3263, 2.2914],[ 0.5834, 0.5834, 0.4788, ..., 2.3611, 2.3088, 2.2914],[ 0.5136, 0.5485, 0.5485, ..., 2.3088, 2.3437, 2.3088],...,[ 0.4091, 0.3916, 0.4439, ..., 0.5834, 0.5136, 0.5311],[ 0.3568, 0.3045, 0.4265, ..., 0.7576, 0.7228, 0.7054],[ 0.1651, 0.1128, 0.3742, ..., 0.7576, 0.7402, 0.7054]]])}
equivalent in torchvision
from PIL import Image
from torchvision.transforms import Compose, Resize, CenterCrop, Normalize, ToTensortv_transform Compose([Resize(256), CenterCrop(224), ToTensor(),Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])image Image.open(banana.png).convert(RGB)
tv_data tv_transform(image)Forward through the model
## IMPORTANT: set the classifier to eval mode
classifier.eval()imgs data[img].unsqueeze(0)
imgs tv_data.unsqueeze(0)with torch.no_grad():# class probabilitiesprob classifier.forward_test(imgs)[0]# featuresfeat classifier.extract_feat(imgs, stageneck)[0]print(len(prob))
print(prob.argmax().item())
print(feat.shape)1000
954
torch.Size([1, 1280])3.使用MMClassificaiton完整进行水果分类实战
数据集下载
GitHub - TommyZihao/MMClassification_Tutorials: Jupyter notebook tutorials for MMClassificationJupyter notebook tutorials for MMClassification. Contribute to TommyZihao/MMClassification_Tutorials development by creating an account on GitHub.https://github.com/TommyZihao/MMClassification_Tutorials
代码框架 def main():model build_classifier(cfg.model)model.init_weights()datasets [build_dataset(cfg.data.train)]train_model(model,datasets,cfg,distributeddistributed,validate(not args.no_validate),timestamptimestamp,devicecfg.device,metameta)
mmcls/apis/train_model.pydef train_model(model,dataset,cfg):data_loaders [build_dataloader(ds, **train_loader_cfg) for ds in dataset]optimizer build_optimizer(model, cfg.optimizer)runner build_runner(cfg.runner,default_argsdict(modelmodel,optimizeroptimizer))runner.register_training_hooks(cfg.lr_config,optimizer_config,cfg.checkpoint_config,cfg.log_config,cfg.get(momentum_config, None),custom_hooks_configcfg.get(custom_hooks, None))runner.run(data_loaders, cfg.workflow)
mmcv/runner/epoch_based_runner.pyclass EpochBasedRunner(BaseRunner):def run_iter(self, data_batch: Any, train_mode: bool, **kwargs) - None:if train_mode:outputs self.model.train_step(data_batch, self.optimizer, **kwargs)else:outputs self.model.val_step(data_batch, self.optimizer, **kwargs)self.outputs outputsdef train(self, data_loader, **kwargs):self.model.train()self.data_loader data_loaderfor i, data_batch in enumerate(self.data_loader):self.run_iter(data_batch, train_modeTrue, **kwargs)self.call_hook(after_train_iter)
mmcls/models/classifiers/base.pyclass BaseClassifier(BaseModule, metaclassABCMeta):def forward(self, img, return_lossTrue, **kwargs):Calls either forward_train or forward_test depending on whetherreturn_lossTrue.Note this setting will change the expected inputs. Whenreturn_lossTrue, img and img_meta are single-nested (i.e. Tensor andList[dict]), and when resturn_lossFalse, img and img_meta should bedouble nested (i.e. List[Tensor], List[List[dict]]), with the outerlist indicating test time augmentations.if return_loss:return self.forward_train(img, **kwargs)else:return self.forward_test(img, **kwargs)def train_step(self, data, optimizerNone, **kwargs):losses self(**data)loss, log_vars self._parse_losses(losses)outputs dict(lossloss, log_varslog_vars, num_sampleslen(data[img].data))return outputs
mmcls/models/classifiers/image.pyclass ImageClassifier(BaseClassifier):def __init__(self,backbone,neckNone,headNone,pretrainedNone,train_cfgNone,init_cfgNone):super(ImageClassifier, self).__init__(init_cfg)if pretrained is not None:self.init_cfg dict(typePretrained, checkpointpretrained)self.backbone build_backbone(backbone)if neck is not None:self.neck build_neck(neck)if head is not None:self.head build_head(head)def extract_feat(self, img):x self.backbone(img)if self.with_neck:x self.neck(x)return xdef forward_train(self, img, gt_label, **kwargs):x self.extract_feat(img)losses dict()loss self.head.forward_train(x, gt_label)losses.update(loss)return losses
mmcv/runner/hooks/optimizer.pyclass OptimizerHook(Hook):def after_train_iter(self, runner):runner.optimizer.zero_grad()runner.outputs[loss].backward()runner.optimizer.step()
总结本篇主要偏向图像分类实战部分使用MMclassification工具进行代码应用熟悉其框架应用为后续处理不同场景下分类问题提供帮助。
本文参考GitHub - wangruohui/sjtu-openmmlab-tutorial
