SOLIDER-REID
SOLIDER-REID copied to clipboard
tinyvision
训练好的模型
您好，如果您是需要预训练模型，可以在SOLIDER上下载，用SOLIDER-REID的训练命令直接训练即可得到finetune后的ReID模型。如果您希望直接使用训练好的ReID模型，我们争取近期更新一版，把模型放上去。
`import torch import torch.nn as nn import torchvision.transforms as T import cv2 from config import cfg from model import make_model import torch.nn.functional as F import numpy as np from PIL import Image, ImageFile import os.path as osp
def read_image(img_path): got_img = False if not osp.exists(img_path): raise IOError("{} does not exist".format(img_path)) while not got_img: try: img = Image.open(img_path).convert('RGB') got_img = True except IOError: print("IOError incurred when reading '{}'. Will redo. Don't worry. Just chill.".format(img_path)) pass return img
@torch.no_grad() def get_feature(img,model,device,normalize=False): input = val_transforms(img).unsqueeze(0) input = input.to(device) output, _ = model(input) if normalize: output = F.normalize(output) return output
val_transforms = T.Compose([ T.Resize(cfg.INPUT.SIZE_TEST), T.ToTensor(), T.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) ]) config_file = "configs/msmt17/swin_small.yml" cfg.merge_from_file(config_file) cfg.freeze() model = make_model(cfg, num_class=1000, camera_num=0, view_num = 0, semantic_weight = 0.2) model.load_param("weights/swin_small_msmt17.pth") device = "cuda"
if device: if torch.cuda.device_count() > 1: print('Using {} GPUs for inference'.format(torch.cuda.device_count())) model = nn.DataParallel(model) model.to(device)
model.eval()
img1 = read_image("test_imgs/p5.jpg") img2 = read_image("test_imgs/p7.jpg") feature1 = get_feature(img1,model,device,normalize=True) feature2 = get_feature(img2,model,device,normalize=True) feature1,feature2 = F.normalize(feature1),F.normalize(feature2) sim = torch.mm(feature1,feature2.t()) print(sim)
@cwhgn 这是我参考test.py写的代码，能帮忙看看问题在哪儿吗？图片之前相似度高达0.9+
`
目前看我暂时也没找到您代码的问题。几点建议哈：1）可以先check下swin_small_msmt17.pth载入是否成功；2）分步check下图片特征和提供的test.py的输出是否一致；3）如果和test.py一致，则有没有可能两者相似度就是很高。
1）模型载入成功 2）对比了test.py提取的特征和我demo里面提取的特征，数值是一样的 3）使用欧式距离，相同图片和不同图片能区分开，但换成余弦相似度，任何图片的相似度都在0.9+
针对3，我理解是因为训练时采用的是欧式距离，见https://github.com/tinyvision/SOLIDER-REID/blob/c35c07671ceb16f72bb7c8fd94929ef436b9406b/loss/triplet_loss.py#L124
如果你希望余弦距离可分的话，可以尝试用余弦距离进行训练。
@deep-practice
A silly newbee question.
this p5 and p7 image is person cropped image. What should be the dimesion of the image ?
is 384 x 192 ?
Is there any newbee document to make inference and compare distance of two detected person image bow.
Detected person could be from yolo or from solider
Best
