Hi Team, I am using the honda driving dataset which consists of images of the videos (3fps). I am using repo (online action detection with transformer and they have used RGB and optical flow feature for the input to the transformer. I have some questions related to honda driving data How could...

Start reporting TypeError: forward_ Dummy() got an unexpected keyword argument 'softmax' error Then I remove the softmax from the code Then run to report ValueError: scale_ factor shape must match input shape. Input is 4D, scale_ Factor is 3 I will reduce the input by one dimension RuntimeError: Inp...

开始报TypeError: forward_dummy() got an unexpected keyword argument 'softmax'错误 然后我将代码里的softmax去除掉 然后运行分别报ValueError: scale_factor shape must match input shape. Input is 4D, scale_factor is 3错 我将输入减少一个维度 运行报RuntimeError: Input and output sizes should be greater than...

Hello, Is there a way to use TimeSformer instead of SlowOnlyResnet backbone in PoseC3D. I have made some changes to make the network work but the network does not seem to be learning even if I tried different learning rates. Can you help for this? I am also sending my config file. Thanks! model = di...

What I try to do is given below. from mmaction.apis import inference_recognizer, init_recognizer config = 'configs/detection/acrn/slowfast_acrn_kinetics_pretrained_r50_8x8x1_cosine_10e_ava22_rgb.py' checkpoint = 'checkpoints/slowfast_acrn_kinetics_pretrained_r50_8x8x1_cosine_10e_ava22_rgb-2be32625....

I am currently trying to train action detection model based on transformer network. However, mmaction2 does not provide a config for training AVA dataset with TimeSformer. Performance of tranformer-based models on the AVA dataset are better than methods based on conv nets, like SlowFast; it would be...

I'm using mmAction2 do some audio recognition tasks. I want to reuse the ResNetAudio backbone in my own project. In this backbone we need to build a basic ConvModule named ConvAudio: However i cannot find this module in mmcv.cnn: which means i cannot build this backbone for now. There is another i...

This question is related to this document. More specifically, the PoseC3D model on the data set NTU120_XSub. In that document, the Top1 Acc of PoseC3D (keypoint heatmat) is 86.3 (in paper is 86.0) However, in the corresponding log file, the final val Top-1 Acc is 0.8596, which is lower the value...

在模型训练完成之后，我们想画ROC曲线和混淆矩阵来补充实验数据，其中模型输出的label和predict probability 需要我们手动保存到一个文件中。在之前看到过的代码中，训练过程中有train_one_epoch这个函数，我可以直接在这个函数中添加我需要的代码。但是在这个工具箱中，由于我的粗心暂时未能找到类似于train_one_epoch函数...

Hello, I am currently running slowonly_r50_u48_240e_ntu120_xsub_keypoint, the question I wanted to ask is can I make changes in the backbone (resnet50) of the training (such as using channel normalization instead of batch normalization)? If i can, is there any guide that I can use of (or any names f...