python -m dataset.video2frame './dataset/breakfast/original/' './dataset/breakfast/rgb_frame/' --nw=12
python -m dataset.annotation_gather './dataset/breakfast/original/' './dataset/breakfast/'
- When processing the stereo video, it seems we just need to choose the ch0 video;
- Add data augmentation;
- Sample the clips based on the GT labels to guarantee the integrity of the video;
- Implement the positional encoding in Transformer;
- Which normalization we should use in Transformer;
- Shall we need the ReLU activation function when calculating the (Q, K, V) in attention;
- shall we need a Conv layer in I3D head instead of average pooling?
- Transformer
- 对子动作segment的长度回归可能**采用范围值而非确定值(或学习一个offset)**会好点,因为子动作持续时间是浮动的;
- 多尺度未来特征生成,因为小物体在浅层特征才具有一定分辨率;
- 在transformer中的attention使用多尺度attention/local attention而不是全局attention;
- 一开始从I3D得到的特征是否需要通过MLP后作动作识别;
- 逆视频输入,加个可学习的正逆特征,像PE特征一样;
- 初始特征增强,加噪声
- 我们采用8frames通过I3D得到每clip的特征,但这些特征会不会由于时序太短而无法捕获到时序信息,更多是空间信息;
- For simplicity, we sample the training data in each pure action segment, while we construct the evaluation datas with continuous frames.
- lr = 0.0001时,训练会明显震荡,检查数据好像没啥问题,改成0.00001会好点;
- 所有层使用default初始化比xavier_uniform好;
- input feature先试用L2norm归一化或L2norm,好像也没啥作用;
- batchsize设小点好像效果更好;
- 没有Positional embedding效果好差;
- I3D features没有经过FC效果会差点,加两层MLP会好点;
- dropout in PE效果不好;
- decoder中PE加offset而不是从0位置计起,效果没提升;
- 4层效果比2层好,6层与4层相当;
- 原始特征加noise会好点
CUDA_VISIBLE_DEVICES=1 python train.py --nw=4 --lr=0.00001 --bs=4 --e_v='L2_dp0.3_lr0.00001_bs64_dfinit_alldata_inputl2norm'