search

aswahd / SamRadiology

1 stars 0 forks source link

reproduce the results #1

Open kaigelee opened 3 days ago

kaigelee commented 3 days ago

What kind of hardware can reproduce the results? For example, 1*3090?

aswahd commented 3 days ago

Hi @kaigelee, I’m using an A100 GPU, but many of the models are <40M parameters and I believe a single 3090 should also work. I’m currently writing a README for a quick start guide, so keep an eye out for that!

kaigelee commented 21 hours ago

Thank you for your reply. I would like to know whether your method is applicable to the detection of multiple objects in one image? If not, how can I modify the code?

aswahd commented 7 hours ago

@kaigelee Yes.

If it is semantic segmentation, you simply need to specify the number of classes in your config.

Example: Semantic Segmentation: ACDC dataset.

image_size: 1024
image_encoder: "sam2_tiny_hiera_adapter"
mask_decoder: "sam2_lora_mask_decoder"
sam_checkpoint: "weights/sam2_hiera_tiny.pt"
wandb_project_name: "ACDC"

dataset:
  name: acdc
  root: /path/to/your/dataset
  image_size: 1024
  num_classes: 3
  num_tokens: 10

training:
  max_epochs: 200
  save_path: checkpoints/ACDC

If you’re doing instance segmentation, while prompt learning works well, you’ll need a bipartite matching loss similar to DETR.

Hope it helps!