chenhao2345
chenhao2345
跨域时需要设定目标域，不需要设定源域，因为我们的算法本身是不需要两个域的。只是有的baseline预热算法需要两个域。
Our encoder has two branches after CNN layers: one branch outputs 2048x4x1 for generative part with **part average pooling**, the other one outputs 2048x1x1 for ReID training with **normal average...
To get FID and SSIM, first generate images with our examples/generate_data.py. Then, use following project code FID: https://github.com/layumi/TTUR SSIM: https://github.com/layumi/PerceptualSimilarity
Thanks for your attention. f_{id} is the 2048x4x1 feature after part average polling designed for the generative module, which is described in the implementation details.
Yes, it helps to train generator. This part is traditional reconstruction in GAN and AE. Reconstruction allows for better adapting the decoder to the person distribution.
你好，感谢你的关注。应该是路径没有设置好的问题，我在examples文件夹下加了data，mesh，logs三个文件夹分别用来存放数据集，mesh和预训练的权重。希望能方便你设置路径。
第一个问题应该是你的驱动版本太老。或许你可以升级下你的驱动。 第二个问题应该是你没有在另外一个环境运行python setup.py develop 另外，推荐使用一块显存较大的gpu >20G。
我重新上传了mesh，加入了test测试集。如果显存比较小，可以尝试调小batchsize到8或者用半精度。我们还没有调试到多GPU训练。
这还是多GPU训练没匹配好得问题。我在.sh文件里暂时加入了CUDA_VISIBLE_DEVICES=0，先强制使用第一个GPU。main.py里做了相应修改。 如果你想尝试用多GPU，可以去掉CUDA_VISIBLE_DEVICES=0，结果不太能保证。
Please first check you are working on Pytorch 1.2. Some people said newer Pytorch version might raise a problem, because of the updated syntax. In stage 2, we pre-train generator...