（EDSR论文实现）pytorch复现1

EDSR

残差模块

常见的残差模块如下图左边，是由两个卷积层组成。右边是瓶颈残差模块。这里的1*1的卷积是用来进行升维和降维的。

class ResBlock(nn.Module):def __init__(self, in_channel=256, out_channel=256, kernel_size=3, stride=1, padding=1,bias=True):super(ResBlock, self).__init__()layers = [nn.Conv2d(in_channel, out_channel, kernel_size=kernel_size, stride=stride,padding=padding, bias=bias),nn.ReLU(inplace=True),nn.Conv2d(in_channel, out_channel, kernel_size=kernel_size, stride=stride,padding=padding, bias=bias)]self.body = nn.Sequential(*layers)self.res_scale = 0.1def forward(self, x):###残差缩放res = self.body(x).mul(self.res_scale)res += xreturn res

上采样模块。这里的上采样模块分别可以实现2，3，4倍上采样。通常上采样模块放在模型的尾端。这里使用的亚像素卷积层。

class Upsampler(nn.Sequential):def __init__(self, scale):layers = []if scale == 2 or scale == 4:layers.append(nn.Conv2d(256, 256 * 4, kernel_size=3, stride=1, padding=1, bias=True))layers.append(nn.PixelShuffle(2))if scale == 4:layers.append(nn.Conv2d(256, 256 * 4, kernel_size=3, stride=1, padding=1,bias=True))layers.append(nn.PixelShuffle(2))elif scale == 3:layers.append(nn.Conv2d(256, 256 * 9, kernel_size=3, stride=1, padding=1, bias=True))layers.append(nn.PixelShuffle(3))else:raise NotImplementedErrorsuper(Upsampler, self).__init__(*layers)

MeanShift(均值漂移)。这里这部分代码没有看懂，不知道是用来干什么的。

class MeanShift(nn.Conv2d):def __init__(self, rgb_range=255, rgb_mean=(0.4488, 0.4371, 0.4040), rgb_std=(1.0, 1.0, 1.0),sign=-1):super(MeanShift, self).__init__(3, 3, kernel_size=1)std = torch.Tensor(rgb_std)self.weight.data = torch.eye(3).view(3, 3, 1, 1) / std.view(3, 1, 1, 1)self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean) / stdfor p in self.parameters():p.requires_grad = False

EDSR模型结构


class EDSR(nn.Module):def __init__(self, scale):super(EDSR, self).__init__()head_layers = [nn.Conv2d(3, 256, kernel_size=3, stride=1, padding=1, bias=True)]body_layers = []for _ in range(32):body_layers.append(ResBlock())body_layers.append(nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=True))tail_layers = [Upsampler(scale=scale),nn.Conv2d(256, 3, kernel_size=3, stride=1, padding=1, bias=True)]self.sub_mean = MeanShift(sign=-1)self.add_mean = MeanShift(sign=1)self.head = nn.Sequential(*head_layers)self.body = nn.Sequential(*body_layers)self.tail = nn.Sequential(*tail_layers)def forward(self, x):x = self.sub_mean(x)x = self.head(x)res = self.body(x)res += xx = self.tail(res)x = self.add_mean(x)return x

图像增强操作。包括随机旋转和随机翻转等操作。但是我在测试的过程中发现裁剪操作好像是随机裁剪的，但是看源码的过程中是固定尺寸裁剪的呀，没有看明白是哪里不对。

import torch
import cv2 as cv
from matplotlib import pyplot as plt
import numpy as np
import glob
import osdef tensor_to_numpy(img):"""将tensor转换为numpy个数convert tensor image to numpy image (0-1 scale to 0-255 scale):param img: the input tensor image:return: the output numpy image"""return 255 * img.permute(1, 2, 0).numpy()def show_tensor_img(img, img_name):"""展现Tensor格式的图像display the input tensor image:param img: the input image:param img_name: the image name"""# convert the tensor image to numpy imagenumpy_img = tensor_to_numpy(img)numpy_img[numpy_img < 0] = 0numpy_img[numpy_img > 255.] = 255.numpy_img = numpy_img.astype(np.uint8)# convert the image from BGR representation to RGB representationcorrected_img = cv.cvtColor(numpy_img, cv.COLOR_BGR2RGB)    ####注意plt.show()展现图像和cv.imshow()展现图像的格式不一样plt.imshow(corrected_img)plt.title(img_name)plt.show()def save_tensor_img(img, img_name, output_path):"""save the tensor image to the output path:param img: the input image:param img_name: the image name:param output_path: the output path"""numpy_img = tensor_to_numpy(img)  # convert the tensor image to numpy imageoutput_p = '{}/{}.jpg'.format(output_path, img_name)cv.imwrite(output_p, numpy_img)def crop(lr_img, hr_img, data_type='tensor', hr_crop_size=192, scale=4):"""crop the same region in low-resolution and high-resolution images:param lr_img: the low resolution image:param hr_img: the high resolution image:param data_type: flag indicating whether the input images are tensor images or numpy images:param hr_crop_size: the cropped size of the high resolution image, default is 192:param scale: scale of the high-resolution image compared to the low-resolution image:return: the cropped low-resolution and high-resolution images"""lr_crop_size = hr_crop_size // scale  # compute the crop size of the low-resolution imageif data_type == 'tensor':lr_img_shape = lr_img.shape[1:3]elif data_type == 'array':lr_img_shape = lr_img.shape[0:2]else:raise NotImplementedError# determine the width and height of cropped low-resolution and high-resolution imageslr_width = torch.randint(low=0, high=lr_img_shape[1] - lr_crop_size + 1, size=(1, 1)).item()lr_height = torch.randint(low=0, high=lr_img_shape[0] - lr_crop_size + 1, size=(1, 1)).item()hr_width = lr_width * scalehr_height = lr_height * scale# extract the corresponding cropped partsif data_type == 'tensor':lr_img_cropped = lr_img[:, lr_height:lr_height + lr_crop_size,lr_width:lr_width + lr_crop_size]hr_img_cropped = hr_img[:, hr_height:hr_height + hr_crop_size,hr_width:hr_width + hr_crop_size]elif data_type == 'array':lr_img_cropped = lr_img[lr_height:lr_height + lr_crop_size,lr_width:lr_width + lr_crop_size, :]hr_img_cropped = hr_img[hr_height:hr_height + hr_crop_size,hr_width:hr_width + hr_crop_size, :]else:raise NotImplementedErrorreturn lr_img_cropped, hr_img_croppeddef random_flip(lr_img, hr_img, data_type='tensor'):"""randomly flip the images left-right:param lr_img: the low-resolution image:param hr_img: the high-resolution image:param data_type: flag indicating whether input images are tensor images or numpy images:return: original images or flipped images"""random = torch.rand(1).item()  # generate a random number between 0 and 1if random < 0.5:  # only flip if the generated random is greater or equal to 0.5return lr_img, hr_imgelse:if data_type == 'tensor':return torch.flip(lr_img, (2,)), torch.flip(hr_img, (2,))elif data_type == 'array':return cv.flip(lr_img, 1), cv.flip(hr_img, 1)else:raise NotImplementedErrordef random_rotate(lr_img, hr_img, data_type='tensor'):"""randomly rotate the images by 0, 90, 180, or 270 degrees clockwise:param lr_img: the low-resolution image:param hr_img: the high-resolution image:param data_type: flag indicating whether input images are tensor images or numpy images:return: original images or rotated images"""# generate a random integer among 0, 1, 2, and 3random = torch.randint(low=0, high=4, size=(1, 1)).item()# rotate according to the generated random numberif data_type == 'tensor':return torch.rot90(lr_img, random, (1, 2)), torch.rot90(hr_img, random, (1, 2))elif data_type == 'array':out_lr = lr_imgout_hr = hr_imgfor _ in range(random):out_lr = cv.rotate(out_lr, cv.ROTATE_90_CLOCKWISE)out_hr = cv.rotate(out_hr, cv.ROTATE_90_CLOCKWISE)return out_lr, out_hrelse:raise NotImplementedErrordef augment_image(train_img_file, target_img_file, train_output_path, target_output_path,aug_num, hr_crop_size, scale):"""augment a single image for some times  :param train_img_file: the path to the training image:param target_img_file: the path to the target image:param train_output_path: the output path of the augmented training image:param target_output_path: the output path of the augmented target image:param aug_num: the repeat number of the augmentation process:param hr_crop_size: the cropped size of the high resolution image:param scale: scale of the high-resolution image compared to the low-resolution image"""train_img = cv.imread(train_img_file)target_img = cv.imread(target_img_file)# repeat the augmentation process for aug_num timesfor i in range(aug_num):count = 0  # record the number of times when no operation appliedc_train, c_target = crop(train_img, target_img, data_type='array',hr_crop_size=hr_crop_size, scale=scale)random = torch.torch.rand(1).item()if random < 0.5:f_c_train, f_c_target = random_flip(c_train, c_target, data_type='array')else:f_c_train, f_c_target = c_train, c_targetcount += 1random = torch.torch.rand(1).item()if random < 0.5:r_f_c_train, r_f_c_target = random_rotate(f_c_train, f_c_target, data_type='array')else:r_f_c_train, r_f_c_target = f_c_train, f_c_targetcount += 1# output the augmented image only if some operation has been appliedif count < 2:train_out = '{}/{}_aug_{}.png'.format(train_output_path,train_img_file.split('/')[-1].split('\\')[-1].split('.')[0], i + 1)target_out = '{}/{}_aug_{}.png'.format(target_output_path,target_img_file.split('/')[-1].split('\\')[-1].split('.')[0], i + 1)cv.imwrite(train_out, r_f_c_train)cv.imwrite(target_out, r_f_c_target)def augment_dir(train_root, target_root, train_output_path, target_output_path, aug_num=5,hr_crop_size=192, scale=4):"""augment the whole directory for some times:param train_root: the path to the training dataset directory:param target_root: the path to the training dataset directory:param train_output_path: the output path of the augmented training dataset:param target_output_path: the output path of the augmented target dataset:param aug_num: the repeat number of the augmentation process:param hr_crop_size: the cropped size of the high resolution image:param scale: scale of the high-resolution image compared to the low-resolution image"""train_files = np.array(glob.glob(train_root + '/*'))target_files = np.array(glob.glob(target_root + '/*'))if not os.path.exists(train_output_path):os.makedirs(train_output_path)if not os.path.exists(target_output_path):os.makedirs(target_output_path)for i in range(len(train_files)):print('augmenting image {} ...'.format(i + 1))augment_image(train_files[i], target_files[i], train_output_path, target_output_path,aug_num=aug_num, hr_crop_size=hr_crop_size, scale=scale)

计算两幅图像的之间的PSNR

def get_psnr(img1, img2, shave_border=0, tensor=False):"""compute the PSNR score between two images:param img1: the first image:param img2: the second image:param shave_border: the amount of border to be shaved:param tensor: flag indicating whether input images are tensor images or numpy images:return: the PSNR score"""if tensor:img1 = util.tensor_to_numpy(img1)else:img1 = img1.astype(int)if tensor:img2 = util.tensor_to_numpy(img2)else:img2 = img2.astype(int)height, width = img1.shape[0:2]# crop both images to the region we should focus onimg1 = img1[shave_border:height - shave_border,shave_border:width - shave_border]img2 = img2[shave_border:height - shave_border,shave_border:width - shave_border]im_dff = img1 - img2rmse = math.sqrt(np.mean(im_dff ** 2))if rmse == 0:return 100return 20 * math.log10(255.0 / rmse)

计算两幅图像之间的SSIM

def ssim(img1, img2):"""helper function of get_ssim to compute the SSIM score of a single channel:param img1: the first image:param img2: the second image:return: SSIM score of a single channel"""C1 = (0.01 * 255) ** 2C2 = (0.03 * 255) ** 2img1 = img1.astype(np.float64)img2 = img2.astype(np.float64)kernel = cv.getGaussianKernel(11, 1.5)window = np.outer(kernel, kernel.transpose())mu1 = cv.filter2D(img1, -1, window)[5:-5, 5:-5]  # validmu2 = cv.filter2D(img2, -1, window)[5:-5, 5:-5]mu1_sq = mu1 ** 2mu2_sq = mu2 ** 2mu1_mu2 = mu1 * mu2sigma1_sq = cv.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sqsigma2_sq = cv.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sqsigma12 = cv.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *(sigma1_sq + sigma2_sq + C2))return ssim_map.mean()def get_ssim(img1, img2, tensor=False):"""compute the SSIM score between two images:param img1: the first image:param img2: the second image:param tensor: flag indicating whether input images are tensor images or numpy images:return: the SSIM score"""if tensor:img1 = util.tensor_to_numpy(img1)else:img1 = img1.astype(int)if tensor:img2 = util.tensor_to_numpy(img2)else:img2 = img2.astype(int)if not img1.shape == img2.shape:raise ValueError('Input images must have the same dimensions.')if img1.ndim == 2:  # in case both images are single-channel imagesreturn ssim(img1, img2)elif img1.ndim == 3:  # in case both images are multi-channel imagesif img1.shape[2] == 3:  # for three-channel imagessims = []for i in range(3):ssims.append(ssim(img1, img2))return np.array(ssims).mean()elif img1.shape[2] == 1:  # for single-channel imagereturn ssim(np.squeeze(img1), np.squeeze(img2))else:raise ValueError('Wrong input image dimensions.')

model

from model.edsr import EDSR
import torch
import os
from utils.trainer import Trainer
import torchvision.transforms as transforms
import cv2 as cv
from utils import image_util
from utils import evaluation_util
import numpy as npdef create_model(scale):"""create an EDSR model of a specific scale:param scale: scale for the model:return: the EDSR model"""return EDSR(scale=scale)def save_checkpoint(model, lr=-1, epoch=-1, batch=-1, path=None, final=False,cuda=False):"""save model checkpoint:param model: the EDSR model:param lr: learning rate:param epoch: epoch number:param batch: batch number:param path: checkpoint output path:param final: flag indicating whether this checkpoint is the last one for the training:param cuda: flag indicating whether GPU will be used"""if final:  # if the checkpoint is the last one for the entire trainingmodel_out_path = path + '/final_model_weights.pt'if not os.path.exists(path):os.makedirs(path)state = {'epoch': epoch, 'model': model.state_dict(), 'lr': lr, 'batch': batch,'cuda': cuda}torch.save(state, model_out_path)print('Final checkpoint saved to {}'.format(model_out_path))else:model_out_path = path + '/model_epoch_{}.pt'.format(epoch)if not os.path.exists(path):os.makedirs(path)state = {'epoch': epoch, 'model': model.state_dict(), 'lr': lr, 'batch': batch,'cuda': cuda}torch.save(state, model_out_path)print('Checkpoint saved to {}'.format(model_out_path))def load_checkpoint(model, checkpoint_load_path):"""load checkpoint into provided model:param model: the model we want the checkpoint to load into:param checkpoint_load_path: the path to the checkpoint:return: the loaded model together with related information"""if os.path.isfile(checkpoint_load_path):print('loading checkpoint \'{}\' ...'.format(checkpoint_load_path))checkpoint = torch.load(checkpoint_load_path, map_location=lambda storage, loc: storage)state = checkpoint['model']start_epoch = checkpoint['epoch']lr = checkpoint['lr']batch_update = checkpoint['batch']use_cuda = checkpoint['cuda']model.load_state_dict(state)print('load checkpoint successfully')return model, start_epoch, lr, batch_update, use_cudaelse:print('=> no checkpoint found at \'{}\''.format(checkpoint_load_path))return None, None, Nonedef train_model(scale, train_dataset, epoch, lr, batch_size, checkpoint_save_path, checkpoint=False,checkpoint_load_path=None, cuda=False):"""train the model:param scale: model scale:param train_dataset: the dataset object used for training:param epoch: epoch number for the training:param lr: learning rate for the training:param batch_size: batch size for the dataloader:param checkpoint_save_path: the output path for the checkpoint:param checkpoint: flag indicating whether a pretrained checkpoint will be used:param checkpoint_load_path: the path to the checkpoint:param cuda: flag indicating whether GPU will be used"""print('initializing model ...')sr_model = create_model(scale=scale)if cuda:sr_model.cuda()# create a Trainer objecttrainer = Trainer(train_dataset, sr_model, cuda=cuda, lr=lr, batch_size=batch_size)trainer.set_checkpoint_saving_path(checkpoint_save_path)trainer.train(epoch=epoch, checkpoint_load_path=checkpoint_load_path,checkpoint=checkpoint)def evaluate_model(model, input_path, cuda=False):"""predict the super-resolution image:param model: EDSR model provided:param input_path: the path to the input image:param cuda: flag indicating whether GPU will be used:return: input and predicted images"""# convert the tensor image from 0-1 to 0-256input_image = 255 * transforms.ToTensor()(cv.imread(input_path))if cuda:  # in case GPU will be usedinput_image = input_image.cuda()output_image = model(input_image)return input_image.detach(), output_image.detach()def enhance(scale, image_path, pre_train=False, weight_path=None, display=False, save=False,output_path=None, cuda=False):"""predict(display/save) the super-resolution image given the input image provided:param scale: model scale:param image_path: path to the input image:param pre_train: flag indicating whether a pretrained model will be used:param weight_path: path to the pretrained weight:param display: flag indicating whether the output will be displayed:param save: flag indicating whether the output will be saved:param output_path: output path of the super-resolution image:param cuda: flag indicating whether GPU will be used"""print('initializing model ...')sr_model = create_model(scale=scale)  # create model of given scaleif pre_train:  # use pretrained modelif scale == 2:sr_model, _, _, _, _ = load_checkpoint(sr_model, './weights/EDSRx2.pt')elif scale == 3:sr_model, _, _, _, _ = load_checkpoint(sr_model, './weights/EDSRx3.pt')elif scale == 4:sr_model, _, _, _, _ = load_checkpoint(sr_model, './weights/EDSRx4.pt')else:raise NotImplementedErrorelse:  # use self-trained checkpointsr_model, _, _, _, _ = load_checkpoint(sr_model, weight_path)if cuda:  # in case GPU will be usedsr_model.cuda()# set the model to evaluation modesr_model.eval()print('enhancing image ...')lr_img, sr_img = \evaluate_model(sr_model, image_path, cuda=cuda)# convert the tensor image from 0-256 to 0-1lr_img /= 255sr_img /= 255print('getting evaluation scores ...')resize_lr_img = image_util.tensor_to_numpy(lr_img).astype(np.uint8)resize_lr_img = cv.resize(resize_lr_img, dsize=(sr_img.shape[2], sr_img.shape[1]),interpolation=cv.INTER_CUBIC)resize_lr_img = transforms.ToTensor()(resize_lr_img)psnr_score = evaluation_util.get_psnr(resize_lr_img, sr_img, tensor=True)ssim_score = evaluation_util.get_ssim(resize_lr_img, sr_img, tensor=True)if display:image_util.show_tensor_img(lr_img, 'low resolution image')image_util.show_tensor_img(sr_img, 'super resolution image')if save:print('saving image ...')if not os.path.exists(output_path):os.makedirs(output_path)image_util.save_tensor_img(sr_img,'{}_sr_x{}'.format(image_path.split('/')[-1].split('.')[0],scale),output_path)print('PSNR score is: {}'.format(round(psnr_score, 2)))print('SSIM score is: {}'.format(round(ssim_score, 2)))

训练模型

import torch.autograd
from torch import optim, nn
from torch.utils.data import DataLoaderfrom utils import model_util as utilclass Trainer:def __init__(self, dataset, model, cuda, lr=1e-4, batch_size=16):self.dataset_ = datasetself.cuda_ = cudaself.model_ = modelself.step_ = 2e5self.lr_ = lrself.batch_size_ = batch_sizeself.batch_update_ = 1self.checkpoint_path_ = Nonedef set_checkpoint_saving_path(self, path):"""set the output path for checkpoints:param path: the output path for checkpoints"""self.checkpoint_path_ = pathdef train(self, epoch, num_worker=1, weight_decay=1e-4, checkpoint_load_path=None,checkpoint=False):"""train the model:param epoch: the epoch number for the training:param num_worker: number of workers for the dataloader:param weight_decay: weight decay for the optimizer:param checkpoint_load_path: the output path for the checkpoints:param checkpoint: flag indicating whether a pretrained checkpoint will be used"""train_loader = DataLoader(self.dataset_, num_workers=num_worker,batch_size=self.batch_size_, shuffle=True)start_epoch = 0if checkpoint: # load the checkpoint if neededself.model_, start_epoch, self.lr_, self.batch_update_, _ = \util.load_checkpoint(self.model_, checkpoint_load_path)# set up the optimizeroptimizer = optim.Adam(filter(lambda p: p.requires_grad, self.model_.parameters()),lr=self.lr_, weight_decay=weight_decay, betas=(0.9, 0.999),eps=1e-08)criterion = nn.L1Loss()if self.cuda_:criterion = criterion.cuda()print('training ...')for epoch_i in range(start_epoch, start_epoch + epoch):self.model_.train()for i, (input_img, target_img) in enumerate(train_loader):# update the learning rate and optimizer for every "step"-size updatesif self.batch_update_ % self.step_ == 0:self.lr_ /= 2optimizer = optim.Adam(filter(lambda p: p.requires_grad,self.model_.parameters()),lr=self.lr_, weight_decay=weight_decay,betas=(0.9, 0.999), eps=1e-08)# convert the tensor image from 0-1 to 0-255input_img *= 255target_img *= 255if self.cuda_:input_img = input_img.cuda()target_img = target_img.cuda()predict_img = self.model_(input_img)target_img = torch.autograd.Variable(target_img, requires_grad=False)loss = criterion(predict_img, target_img)optimizer.zero_grad()loss.backward()optimizer.step()self.batch_update_ += 1print('===> Epoch[{}/{}]({}/{}): Loss: {}'.format(epoch_i + 1, start_epoch + epoch,i + 1, len(train_loader),round(loss.item(), 4)))# save the checkpoint for every epochutil.save_checkpoint(model=self.model_, epoch=epoch_i + 1, lr=self.lr_,batch=self.batch_update_, path=self.checkpoint_path_)util.save_checkpoint(model=self.model_, epoch=epoch, batch=self.batch_update_, lr=self.lr_,path=self.checkpoint_path_, final=True)

（EDSR论文实现）pytorch复现1相关推荐

亚像素卷积网络(ESPCN)学习与Pytorch复现
论文内容论文地址:Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolution ...
NVIDIA新作解读：用GAN生成前所未有的高清图像（附PyTorch复现） | PaperDaily #15
在碎片化阅读充斥眼球的时代,越来越少的人会去关注每篇论文背后的探索和思考. 在这个栏目里,你会快速 get 每篇精选论文的亮点和痛点,时刻紧跟 AI 前沿成果. 点击本文底部的「阅读原文」即刻加入社区 ...
导师：顶会论文3天都复现不出来？你退学吧！
(文末招聘论文讲师) 说起国内的AI算法大赛平台,你首先会想到哪一个? 我猜你第一反应会是:阿里云天池--国内第一人工智能竞赛平台. 天池的比赛打的多了,但天池的论文复现课程你体验过吗?11月23日, ...
我用 PyTorch 复现了 LeNet-5 神经网络（自定义数据集篇）！
大家好,我是红色石头! 在上三篇文章: 这可能是神经网络 LeNet-5 最详细的解释了! 我用 PyTorch 复现了 LeNet-5 神经网络(MNIST 手写数据集篇)! 我用 PyTorch ...
我用 PyTorch 复现了 LeNet-5 神经网络（CIFAR10 数据集篇）！
大家好,我是红色石头! 在上两篇文章: 这可能是神经网络 LeNet-5 最详细的解释了! 我用 PyTorch 复现了 LeNet-5 神经网络(MNIST 手写数据集篇)! 详细介绍了卷积神经网络 ...
【深度学习】我用 PyTorch 复现了 LeNet-5 神经网络（自定义数据集篇）！
在上三篇文章: 这可能是神经网络 LeNet-5 最详细的解释了! 我用 PyTorch 复现了 LeNet-5 神经网络(MNIST 手写数据集篇)! 我用 PyTorch 复现了 LeNet-5 ...
Pytorch复现FCN网络
Pytorch复现FCN网络详解(可复现) 1.环境配置 windows10,pytorch=1.3,python=3.6 参考博客:https://github.com/wkentaro/pytor ...
InfoGAN论文理解及复现
InfoGAN论文理解及复现 by AHU Random_Walker 主要从四个方面来了解相关工作,即 Motivation Related Work-Method Experiments Conc ...
YOLO V1~V7论文及Pytorch实现详解
YOLO~V1论文及Pytorch实现详解论文地址:https://paperswithcode.com/paper/you-only-look-once-unified-real-time-obj ...
CV+Deep Learning——网络架构Pytorch复现系列——Detection(一：SSD:Single Shot MultiBox Detector 4.推理Detect)
上一话 CV+Deep Learning--网络架构Pytorch复现系列--Detection(一:SSD:Single Shot MultiBox Detector 3.loss)https:// ...

（EDSR论文实现）pytorch复现1

（EDSR论文实现）pytorch复现1相关推荐

最新文章

热门文章