[ERROR] The example script for agent_motion_prediction possibly has issues

[giaminhhoang]

Hi,

I tried to run the example scripts. While visualise_data worked well, the agent_motion_prediction first has strange eval_dataloader as follows:
`dm = LocalDataManager(None)

===== INIT DATASETS

rasterizer = build_rasterizer(cfg, dm)
train_dataloader = build_dataloader(cfg, "train", dm, AgentDataset, rasterizer)
eval_dataloader = build_dataloader(cfg, "val", dm, AgentDataset, rasterizer)`

This yielded the size of eval_dataloader is zero which led to error at the evaluation (training was fine).

I tried change the scene_indices for val_data_loader to 1 in the file agent_motion_config and got non-zero eval_dataloader and the result below

+-------------+--------+
| future step |  MSE   |
+-------------+--------+
|      1      |  2.30  |
|      2      |  7.10  |
|      3      | 54.06  |
|      4      | 106.99 |
|      5      | 134.82 |
|      6      | 205.89 |
|      7      | 181.60 |
|      8      | 192.88 |
|      9      | 198.58 |
|      10     | 225.66 |
...

However, the Visualise Results doesn't work.
`---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
in
28
29 # convert coordinates to AV point-of-view so we can draw them
---> 30 predicted_positions = transform_points(np.concatenate(predicted_positions), data_ego["world_to_image"])
31 target_positions = transform_points(np.concatenate(target_positions), data_ego["world_to_image"])
32

<array_function internals> in concatenate(*args, **kwargs)

ValueError: need at least one array to concatenate`

Best.

[lucabergamini]

whoops, we messed up with the config file during refactoring. It will be solved by noon :)

no eval data:

• scene indices set to -1 will solve it by picking the entire dataset (note you're training and testing on the same dataset, to try a full proper train change the val_data_loader/datasets/key to something else than sample.zarr)

error in visualisation:

• that happens when a frame doesn't have any agent in it. I'm changing the code to pick the first frame with at least one agent

you can track the development of this in #40

[lucabergamini]

fixed in #40