demo？

[lunalulu]

I implemented a skeleton extraction, and then motion recognition demo；
I modified the use of a smaller number of key points （14）for motion recognition, retraining，there are 6 classes。The training accuracy was 0.90102261。
also I modified the network parameters, window size is 16, person num is 1。
But demo's action recognition is poor.
I accumulated 16 frames of input into the HCN, the next time from the previous accumulation of half the beginning of the accumulation。

Please give me some advice, thank you

[huguyuehuhu]

Maybe you should check if your model is overfiting , beacause the training accuracy is seems good. It really hard to provide any useful suggestion without any specific information about training /testing curves.

[lunalulu]

@huguyuehuhu OK，thanks
another problem：
When I only take the data of the previous 49 classes as the training set and test set, the training effect is very poor，using the two-dimensional coordinates of the skeleton, of course. However, it is good to select 9 types of training results, although it is possible to overfit.
188 2019-04-26 14:20:59,997:INFO: Epoch 30/600 189 2019-04-26 14:21:08,408:INFO: - Train metrics: accuracytop5: 0.096 ; loss: 3.892 ; ls_CE: 3.892 ; ls_all: 3.892 ; accuracytop1: 0.018 190 2019-04-26 14:21:10,102:INFO: - Eval metrics : accuracytop5: 0.102 ; loss: 3.892 ; ls_CE: 3.892 ; ls_all: 3.892 ; accuracytop1: 0.020 191 2019-04-26 14:21:10,158:INFO: Epoch 31/600 192 2019-04-26 14:21:18,724:INFO: - Train metrics: accuracytop5: 0.095 ; loss: 3.893 ; ls_CE: 3.893 ; ls_all: 3.893 ; accuracytop1: 0.019 193 2019-04-26 14:21:20,352:INFO: - Eval metrics : accuracytop5: 0.103 ; loss: 3.892 ; ls_CE: 3.892 ; ls_all: 3.892 ; accuracytop1: 0.021

[huguyuehuhu]

1. You can see the training accuracy is about 1/49~=0.02, that means your model learned nothing but random guess the results. It is impossible, even a one layer LSTM could perform better than random guess, so it is surely that there is something wrong in your data preparation or somewhere else.

2. Maybe you can try:
   (1) use visdom to see the accuracy and loss curves during all the training process, it obviously your loss did not decrease. NOT just focusing on the final accuracy at the terminal.
   (2) try to use one layer LSTM to replace the HCN model, once you can make your loss decreasing, then you can sure your code processes right. After that you can use HCN to pursuit a better performance.