Meta-rPPG icon indicating copy to clipboard operation
Meta-rPPG copied to clipboard

This model is very difficult to train. What training skills are there?

Open Z-Xiong opened this issue 4 years ago • 5 comments

Thanks a lot for releasing the code of Meta-rPPG paper! We use the Deap dataset train Meta-rPPG model, but the model is difficult to converge. Do you have any training skills? Thaks a lot!

Z-Xiong avatar Nov 09 '20 04:11 Z-Xiong

Thanks a lot for releasing the code of Meta-rPPG paper! We use the Deap dataset train Meta-rPPG model, but the model is difficult to converge. Do you have any training skills? Thaks a lot!

Hello, How do you process the DEAP dataset?

kite323 avatar Dec 16 '20 12:12 kite323

Hi,

A simple guideline is to pay attention to the partitioning of meta-training set into query and support sets. The rest of the training is straightforward and can be directly derived using our released code. Good luck!

Eugene

eugenelet avatar Jan 17 '21 06:01 eugenelet

Thanks a lot for releasing the code of Meta-rPPG paper! We use the Deap dataset train Meta-rPPG model, but the model is difficult to converge. Do you have any training skills? Thaks a lot!

Hello, How do you process the DEAP dataset?

I use PFLD to extract facial key points, and use the key points to generate facial ROI and mask. Then I downsampling the image and plethysmograph signal to 30fps. However, members of our group found that DEAPS dataset is difficult to converge. I recommend that you use UBFC dataset. Its video data and PPG signal are better, so you can use them directly without preprocessing.

Z-Xiong avatar Feb 07 '21 09:02 Z-Xiong

Hi,

A simple guideline is to pay attention to the partitioning of meta-training set into query and support sets. The rest of the training is straightforward and can be directly derived using our released code. Good luck!

Eugene

I use the data example “exampl.pth” provided by you to train 30 iterations with default parameters. The loss is as follows. It seems that there is no convergence. I drew the prediction label and the real label, and found that the network didn't seem to be trained. Is it because there are too few samples?

Total number of parameters : 0.380 M
---------------------end----------------------
dataset [rPPGDataset-train] was created
dataset [rPPGDataset-test] was created
Data Size: 650 ||||| Batch Size: 3 ||||| initial lr: 0.001000
Epoch 1/30 ||||| Time: 9 sec ||||| Lr: 0.0009141 ||||| Loss: 27.127/27.025
Epoch 2/30 ||||| Time: 10 sec ||||| Lr: 0.0006891 ||||| Loss: 26.496/26.333
Epoch 3/30 ||||| Time: 10 sec ||||| Lr: 0.0004109 ||||| Loss: 25.989/25.749
Epoch 4/30 ||||| Time: 9 sec ||||| Lr: 0.0001859 ||||| Loss: 25.520/25.350
Epoch 5/30 ||||| Time: 8 sec ||||| Lr: 0.0001000 ||||| Loss: 25.441/25.150
Epoch 6/30 ||||| Time: 8 sec ||||| Lr: 0.0001859 ||||| Loss: 25.147/25.039
Epoch 7/30 ||||| Time: 8 sec ||||| Lr: 0.0004109 ||||| Loss: 24.989/24.829
Epoch 8/30 ||||| Time: 8 sec ||||| Lr: 0.0006891 ||||| Loss: 24.463/24.353
Epoch 9/30 ||||| Time: 12 sec ||||| Lr: 0.0009141 ||||| Loss: 23.767/23.556
Epoch 10/30 ||||| Time: 9 sec ||||| Lr: 0.0010000 ||||| Loss: 22.900/22.494
Epoch 11/30 ||||| Time: 9 sec ||||| Lr: 0.0009141 ||||| Loss: 22.128/21.585
Epoch 12/30 ||||| Time: 8 sec ||||| Lr: 0.0006891 ||||| Loss: 21.419/21.028
Epoch 13/30 ||||| Time: 8 sec ||||| Lr: 0.0004109 ||||| Loss: 21.424/20.806
Epoch 14/30 ||||| Time: 8 sec ||||| Lr: 0.0001859 ||||| Loss: 21.671/20.743
Epoch 15/30 ||||| Time: 12 sec ||||| Lr: 0.0001000 ||||| Loss: 20.855/20.712
Epoch 16/30 ||||| Time: 9 sec ||||| Lr: 0.0001859 ||||| Loss: 21.125/20.700
Epoch 17/30 ||||| Time: 10 sec ||||| Lr: 0.0004109 ||||| Loss: 21.850/20.678
Epoch 18/30 ||||| Time: 8 sec ||||| Lr: 0.0006891 ||||| Loss: 21.629/20.645
Epoch 19/30 ||||| Time: 9 sec ||||| Lr: 0.0009141 ||||| Loss: 21.301/20.597
Epoch 20/30 ||||| Time: 8 sec ||||| Lr: 0.0010000 ||||| Loss: 20.735/20.559
Epoch 21/30 ||||| Time: 8 sec ||||| Lr: 0.0009141 ||||| Loss: 21.225/20.634
Epoch 22/30 ||||| Time: 9 sec ||||| Lr: 0.0006891 ||||| Loss: 21.797/20.564
Epoch 23/30 ||||| Time: 9 sec ||||| Lr: 0.0004109 ||||| Loss: 21.015/20.575
Epoch 24/30 ||||| Time: 8 sec ||||| Lr: 0.0001859 ||||| Loss: 21.235/20.549
Epoch 25/30 ||||| Time: 8 sec ||||| Lr: 0.0001000 ||||| Loss: 21.467/20.539
Epoch 26/30 ||||| Time: 8 sec ||||| Lr: 0.0001859 ||||| Loss: 21.035/20.538
Epoch 27/30 ||||| Time: 10 sec ||||| Lr: 0.0004109 ||||| Loss: 21.793/20.538
Epoch 28/30 ||||| Time: 11 sec ||||| Lr: 0.0006891 ||||| Loss: 21.228/20.547
Epoch 29/30 ||||| Time: 11 sec ||||| Lr: 0.0009141 ||||| Loss: 21.193/20.504
Epoch 30/30 ||||| Time: 9 sec ||||| Lr: 0.0010000 ||||| Loss: 21.438/20.281

Process finished with exit code 0```

![Uploading 图片.png…]()

Z-Xiong avatar Feb 07 '21 09:02 Z-Xiong

I also encountered the same problem, did you solve this problem?

DaJun666666 avatar Mar 30 '23 08:03 DaJun666666