How did you calculate the normalization constants?

[ultimatedigiman]

Hi,

I found some normalization constants in the repo, such as

• normalizing the future coordinates:
  

  waymo-motion-prediction-challenge-2022-multipath-plus-plus/code/train.py

  Line 85 in cf020ff

  xy_future_gt = (data["target/future/xy"] - torch.Tensor([1.4715e+01, 4.3008e-03]).cuda()) / 10.

• normalizing inputs:
  

  waymo-motion-prediction-challenge-2022-multipath-plus-plus/code/prerender/utils/utils.py

  Lines 37 to 61 in cf020ff

  	def get_normalize_data():
  	return {
  	"target": {
  	"xy": {
  	"mean": np.array([[[-3.0173979, 0.00575967]]]),
  	"std": np.array([[[3.7542882, 0.11941358]]])},
  	"yaw": {
  	"mean": np.array([[[0.00815599]]]),
  	"std": np.array([[[1.0245908]]])},
  	"speed": {
  	"mean": np.array([[[6.1731253]]]),
  	"std": np.array([[[5.53667]]])}},
  	"other": {
  	"xy": {
  	"mean": np.array([[[9.855061 , 2.6597235]]]),
  	"std": np.array([[[44.58452 , 34.069477]]])},
  	"yaw": {
  	"mean": np.array([[[1.6482836]]]),
  	"std": np.array([[[3.7098966]]])},
  	"speed": {
  	"mean": np.array([[[2.5248919]]]),
  	"std": np.array([[[4.806048]]])}},
  	"road_network_segments": {
  	"mean": np.array([[[11.440233, 3.4300654]]]),
  	"std": np.array([[[66.125916, 53.79835]]])}}

I have a few questions about normalization, please let me know if you have any suggestions:

1. Is there any script to get these constants?
2. Have you tried training without any normalization? What's the result?
3. How to modify the training code to non-normalization version? Just set these 2 lines below to False?
   

   waymo-motion-prediction-challenge-2022-multipath-plus-plus/code/configs/final_RoP_Cov_Single.yaml

   Lines 20 to 21 in cf020ff

   	normalize: True
   	normalize_output: True

[stepankonev]

Hi!
Thank you for your question. First, it seems like get_normalize_data() is outdated. It used to provide normalization during data preparation, however now it is done in dataloader. I will remove it in the following commits with the normalization option in prerender config.
As for your questions:

1. No script is provided for computing these constants yet.
2. Normalization seemed to provide numerical stability improvements. I did not train the network for long without normalization.
3. Exactly

Hope it was helpful

[ultimatedigiman]

Thanks! Steven