naver-ai / egtr

[CVPR 2024 Best paper award candidate] EGTR: Extracting Graph from Transformer for Scene Graph Generation
https://arxiv.org/abs/2404.02072
Apache License 2.0
78 stars 2 forks source link

Trained EGTR model #8

Closed thomasfoltz closed 2 months ago

thomasfoltz commented 2 months ago

@jinbae I appreciate the documentation for this work! However, I currently don't have enough compute to train my own custom EGTR model on the Visual Genome Dataset. Would you make those weights publicly available? Thank you!

For others to reference, this is the memory error I run into when there isn't sufficient GPU memory to train. I had to set the --gpus arg to 1 since I don't have 8 GPUS available.

image

jinbae commented 2 months ago

Our model weights were already made public. Please refer to the readme.

image
trongminh03 commented 2 months ago

The error is due to insufficient shared memory. If you're using Docker, try allocating more shared memory by using the "--shm-size" option when running the Docker image. Hope this helps!

thomasfoltz commented 2 months ago

Thanks for the advice! It turns out it wasn't an issue with the weights I was using, but rather the insufficient shared memory as @trongminh03 mentioned. Now it is running, but unfortunately, since I am only using one RTX 4070, I can only train with a batch size of 1.