Peng Zheng comments

Results 38 comments of


                                            Peng Zheng

Dataset split & Query portrait pictures & resolution of key frames on Character Identification task.

Thanks for your quick reply👍.

Dataset split & Query portrait pictures & resolution of key frames on Character Identification task.

Besides, could you provide the **keyframes with 720P** as is stated in the original paper? Because all the annotations are made in 720P.... If trained on 240P, details of the...

img_size

Search the variable `keep_res` in dataset.py and you will know how to do it.

RuntimeError: Given groups=1, weight of size 48 160 1 1, expected input[2, 640, 3, 3] to have 160 channels, but got 640 channels instead

Have you fixed this error?

RuntimeError: Given groups=1, weight of size 48 160 1 1, expected input[2, 640, 3, 3] to have 160 channels, but got 640 channels instead

`print(model)` and search 160, you will find the only 160 is the input_channel of the decoder. It's the `low_level_inplanes` controls the input_channel, which defined by filter_multiplier. Then see the ln25...

为啥是蓝色效果

不懂什么意思, 请把视频, 修改代码的位置给出. 还有"模就是蓝色的通道"是什么意思...

为啥是蓝色效果

不是呀, 我之前跑出来的效果就像README里的gif图. 我觉得你要不去检查一下测试视频分解下来的frame? 如果不行, 可以试一下改下通道试试? 我最近太累了, 而且项目太久了. 我看看吧, 这几天如果有精力的话写个colab, 如果有的话我到时回复你.

为啥是蓝色效果

你可以去看下原论文对这个synthetic数据集的介绍, 这个数据集只针对运动的放大, 没有颜色相关的事. 检查代码的话我应该是真不太有时间去debug。。。你就先确保数据本身, 环境这些一致. 代码直接跑应该是没问题的.

run_temporal_on_test_videos.sh takes a lot of time to process.

So do I. I set the GPU device but found that it only used a bit of GPU memory during the test. Maybe that's the common time-consuming?...

求问如何选择的val数据集,有什么特殊的标准吗

You can just randomly selected about 10 or 20 images from the training set as validation set. But in my personal view, I prefer to train also on the validation...