Train configuration problem

[Thang1703hrsh]

Let me ask if train model with full nuscenes data, what is your configuration need, I have problem with too much RAM when training, my RAM is 13GB, I use colab or kaggle

Thank you very much

[basbaba]

I'm training mono-semantic-maps on a full nuscene dataset, with 16GB nvidia T4, 200 epoch, it's about 460 hours(each epoch 2.3 hours). I do think translating-imgs-into-maps will cost more time. On mini dataset, translating-imgs-into-maps took about 7 hours for 600 epoch under the default configuration.

[Thang1703hrsh]

Thanks you very much, with the mini dataset, is the prediction result good or not?

[Thang1703hrsh]

I am working on a thesis related to this topic, can I ask for your code for reference, thank you very much

[basbaba]

mono-semantic-maps url: https://github.com/tom-roddick/mono-semantic-maps

you'd better train with the full dataset, prediction on mini-trained model is terrible.

[Thang1703hrsh]

Do you use a personal computer or rent a platform, I was able to run it but colab pro is not enough for 1 month, because there are only 100 computing units

[basbaba]

I'm working on a server in office when there is no job running on it. In fact you don't need to do 200 epochs in one time, I think after training 10 epochs, a testable model would be generated. By saving and reloading you can get a procedural checkpoint for tests.

[Thang1703hrsh]

Hi, I bought a google colab pro account but it still doesn't meet the requirements, can you share me some checkpoints, thank you very much

[basbaba]

Tomorrow when arriving office, I'll try to post a checkpoint to huggingface. It's trained on the full nuscene database.

[basbaba]

Here: https://huggingface.co/roam/TIIM/tree/main Only trained 4 epoch, don't expect too much on the result.

[Thang1703hrsh]

Thank you very much, nice to meet you

[Yutong-gannis]

@basbaba can you share your nuscenes convert code please

[basbaba]

1. There is a pull request(New DataLoader) to this project, you can find here: https://github.com/avishkarsaha/translating-images-into-maps/pulls
2. I modified few things from the New DataLoader
3. the modified code is zipped and pushed to https://huggingface.co/roam/TIIM/tree/main
4. After downloading nuscenes full dataset, remember to pre-process it using https://github.com/tom-roddick/mono-semantic-maps
5. import dataloader_new.py to load data from the processed dataset

[Yutong-gannis]

@basbaba How can I produce lane data?

[basbaba]

If you mean how to get lane from datasets, I have no ideal. In Nuscenes, there is a definition: NUSCENES_CLASS_NAMES = [ 'drivable_area', 'ped_crossing', 'walkway', 'carpark', 'car', 'truck', 'bus', 'trailer', 'construction_vehicle', 'pedestrian', 'motorcycle', 'bicycle', 'traffic_cone', 'barrier' ] 'drivable_area', 'ped_crossing', 'walkway' may be what you want.

[Thang1703hrsh]

how did you go about getting the ground truth image, i did the same in https://github.com/tom-roddick/mono-semantic-maps however it is not the same as the ground truth in the mini episode, can you give me an example image of your ground truth. Thank you very much

[Thang1703hrsh]

[basbaba]

Almost the same

[Thang1703hrsh]

@basbaba do you have facebook, skype or any social network, i want to exchange something about datasets and how to train models, there are some places i'm a bit confused, tks you very much

[Thang1703hrsh]

tks u, i am ducthang170301@gmail.com

[Thang1703hrsh]

you can add contact me, https://t.me/ducthang1703

[basbaba]

@basbaba How can I produce lane data?

Recently I'm working on nuScenes map datasets and found the Lane data in it's NuScenesMap:

def _load_layers(): self.lane = self._load_layer('lane')

you can get it by NuScenesMap or directly from json file.