Discussion on SiamMOT baseline

Hi,

Thanks for providing starterr kit.

I am curious about which dataset are the baseline (siammot: 0.66990,0.45570) trained on, full dataset or partial dataset (500G) ?
And how long to train the model and how many about the gpu devices?

By the way, would the competition be postponed? It may be cost much time for us to train the model due to the large dataset.

Thanks!

2 Likes

@shivam Hi, could you provide some idea? Thanks

Hi @octo,

Apologies for missing your question.
We have recently provided the baseline code in the starter kit and the model.
It does not use the full dataset, it only relies on frames with airborne objects in range < 1200 m
It takes ~ 5 hours to train the network with 4 GPUs.
We are checking the option to postpone the end date of the challenge.

Hope this helps.

1 Like

@zontakm9

Hi, It too large for us to download full dataset(13T). Could you provide compressed full data or compressed frames with airborne objects in range < 1200 m.

Thanks

Hi @octo,

You can get the partial dataset using the partial=True flag in the helper library for downloading dataset.

Details here:
:point_right: https://discourse.aicrowd.com/t/faqs-and-common-mistakes-while-making-a-submission/5781#8-tip-dataset-too-big-to-play-around-with

Dataset library (part of starter kit, in case you missed it):
:point_right: DATASET.md

In case you are downloading the dataset manually, you can use range_distance_m in groundtruth.json for selecting frames you are interested in and download only those images.

1 Like

Hi,
Got it!, Thanks!

2 Likes

@shivam @zontakm9
Could you provide some suggestion about the issue?

Thanks!

Getting error in siam-mot baseline submission “pack exceeded maximum allowed size”

1 Like

Hello @octo

Replied on the thread. Let us know if you need any further help on this :slight_smile:

1 Like

Have you used the same trainer.py as in the repo and the same config files? I have recently setup the training pipeline and loss stagnants for me very quickly in 4-5 mins

1 Like

We did use the provided config file and the trainer.py
One suggestion might be to check that the hyperparameters are adjusted based on the number of GPUs you are working with

Can you provide and clarify the specific details? WIll be of great help to reproduce the results: a)the number of GPUs used along with their memory b)Number of images per batch and hence per GPU c)The loss aggregation scheme across GPUs is “sum” and per GPU is “mean”,right?