AssertionError: Database version not found: v1.0-trainval

[lyzxcb]

Hello, I installed conda ,python environment and downloaded the dataset, but it is running/ gene_ woDTR. SH times wrong. AssertionError: Database version not found: v1.0-trainval

[ramdrop]

Hi, you may need to define your downloaded nuScenes dataset directory in generic.py: line 29.

[lyzxcb]

Thanks!This problem has been solved, but it is running "/gene_woDTR.sh" There are still some problems, which shows that “cannot be found n008-2018-08-01-16-03-27-0400RADAR BACK_ LEFT__1533153857956096.pcd”. I downloaded the "radar blobs" part of these datasets in the figure below and put it under the path related to date/nuscenes. I don't know which step is wrong

[ramdrop]

You need to download the Test split, which includes :

[lyzxcb]

thank! I was confused and forgot it. However, at present, I put the.pcd file in test into the "/date" path, and there are still problems when running "/ gene_wodtr. Sh", prompting "valueerror: not enough values to unpack (expected 3, got 2)". I haven't modified the source code. Do I need to modify some parameters?

[ramdrop]

I double-checked and found this is due to a recent change of nuscenes-devkit API . I have updated the requirements.txt file, now you can fix this issue by updating your environment by pip install -r requirements.txt or pip install nuscenes-devkit==1.1.1.

[lyzxcb]

Hello, I re established a CONDA environment and re downloaded and ran your newly released "autoplace" and "environment setup". The same problem still occurred when running ". / gene_wodtr. Sh". “ValueError: not enough values to unpack (expected 3, got 2)” 。

[ramdrop]

Could you please show me your nuscene-devkit version by running pip show nuscenes-devkit?

[lyzxcb]

Hello, I deleted the previously downloaded "nuscenes devkit" and downloaded "nuscenes devkit = = 1.1.1" by running "pip install - r requirements. txt"

[ramdrop]

Hi, sorry I realized this API is indeed my customized function. To solve this issue, please run git pull and then modify the specific nuscenes-devkit API as I did: go to the nuscenes-devkit package directory (depends on which python you are using), and override the function from_file_multisweep(...) in site-packages/nuscenes/utils/data_classes.py with the function provided in autoplace/nuscenes-devkit_override.py. BTW, it is still okay to use nuscenes-devkit==1.1.1.

[lyzxcb]

Thank you for your reply. According to your prompt, the data preprocessing has been completed. There are some problems during training. When I run "3: train spatial encoder + temporary encoder (se_te)", the error appears to be that the graphics memory of the graphics card is not enough. For this problem, I have two questions: 1 Do I need to change my computer with higher configuration for training?

2. Can I continue to use this machine for training without changing the computer? Is there any difference in effect?

[ramdrop]

1. To reproduce the same paper results I think you may need to resort to a graphics card with larger memory (~34GB).
2. To reduce the memory requirements, you may need to reduce either spatial feature dimensions or temporal feature dimensions. (1) For reducing spatial feature dimensions, you can override the Maxpooling layer with a larger kernel size or stride. (2) For reducing temporal feature dimensions, you can reduce the hidden size of the LSTM layer. But neither cases will promise a similar/comparable performance as the original network.

[lyzxcb]

Hello, I ran four "train" commands，in the evaluation report, "modulenotfounderror: no module named 'autoplace. Logs_autoplace. Networks'", I'm sorry to bother you. I'm really not familiar with Python. Is the path I added wrong?

[ramdrop]

Your resume path should narrow down to the checkpoint file directory, as shown in resume_path.json.

BTW, it is a better practice to open a new issue for a new error.