Action Recognition Code

[jim4399266]

Dear author, regarding the Action Recognition, how can I modify the "Frames x Views" parameter in the code? I noticed the "num_segments" parameter in the config yaml file, does "num_segments" correspond to the number of frames?

[HJYao00]

Frames x Views in evaluation: https://github.com/HJYao00/Side4Video/blob/main/Recognition/test_vision.py#L58 and L59;
In training, num_segments means the number of frames

[jim4399266]

Thanks a lot :D

[jim4399266]

Could you please explain how the FLOPS is calculated? I used the thop and inserted the code between lines 230 and 233, but the calculated TFLOPs do not match the data shown in Table 4 of the paper. Do I need to use a specific batch size or other parameters, or is the code placed in the wrong location?

Thank you very much for your help!

image_input = image.to(device).view(-1, c, h, w)
statistic(model, image_input)
logits = model(image_input)  # bt n_class

def statistic(model, *args):
    # 1.  torchinfo 
    model_stats = summary(model, input_data=(*args,), verbose=1)
    # 1 MAC ≈ 2 FLOPs
    flops = model_stats.total_mult_adds * 2

    # 2.  thop 
    macs, params = profile(model, inputs=(*args,), verbose=False)
    flops = macs * 2