talsperre / infer Goto Github PK

Hi Shashank,

I run the infer-main.ipynb. By default, it is using INFER-Skip (Top 5) in the KITTI dataset, but the result that I get is

1s: 2.209328362278883, 2s: 2.9157824738774303, 3s: 3.683025655408532, 4s: 5.294841714296214

I believe this is the same ADE metric that you used in Table 1 of your paper, but why the ADE here is significantly worser than what your paper recorded? How should I replay the result in the paper?

Regards,
DK

Hi, @talsperre
thanks for your work. Could you share your official code for data preprocessing? Or official guidance is ok.

The requirements.txt file is broken. Do not use that one to set up your virtual environment.

Do this

# Create a conda environment first
conda create --name pytorch04

# Enter the environment
conda activate pytorch04

# Install pytorch 
conda install pytorch=0.4.1 cuda90 -c pytorch

Replace everything in the requirements.txt file to the following

matplotlib
pandas
jupyterlab
torchvision==0.2.1

then run pip install -r requirements.txt. Now you should be able to run the infer-main.ipynb and train.py

In the paper you say that you add a safety loss term, that penalizes all predicted
states of vehicles that lie in an obstacle cell. In the args.txt file supplied with the pretrained KITTI model, --lossOT = True, suggesting you have used this loss function. However, in the training code you define the obstacleLossFun then never use it.

This is just one example of a load of seemingly unused args, making it very difficult to recreate your results given that it is not clear exactly which args you used in your training.

I have re-train many times, but I can't get as good as your release. My result is followed. Can you give me some advice?

Hello, I would like to know how you generate a picture like Figure 5 in the paper. I think such a visual picture is very intuitive, and I want to generate it myself.I would be more grateful if you could provide the source code。Thanks！

In order to predict multiple time steps ahead for the training/validation/testing where you already have access to the future intermediate representations you simply combine the prediction at the previous time step with the next incoming intermediate representation and use this as the next time step's input. But how do you predict multiple time steps ahead live at inference time when you don't have access to the future intermediate representations? I couldn't see any code for this in the repo.