SAMed

SAMed copied to clipboard

Technical issues

1

# Select the correct mask or masks for output # if multimask_output: # mask_slice = slice(1, None) # else: # mask_slice = slice(0, 1) # masks = masks[:, mask_slice, :,...

yuanchaoX

'if args.skip_hard and iter_num > 3 * args.warmup_period and loss.item() > 0.4:' appears in '/data/gauss/wr/SAMed/SAMed_h/trainer.py', but does not define skip_hard This parameter, what does it mean and what is the...

AdamRii

Hi, SAM2 has released ! Have anyone has idea how to use Lora to fine-tuned on specific data Thanks

anchichuang

请问有人知道这个文件对应的是怎么样的结构吗？论文中似乎没有给出

Saillxl

结果复现

1

parser = argparse.ArgumentParser() parser.add_argument('--root_path', type=str, default='/data/LarryXu/Synapse/preprocessed_data/train_npz', help='root dir for data') parser.add_argument('--output', type=str, default='/output/sam/results') parser.add_argument('--dataset', type=str, default='Synapse', help='experiment_name') parser.add_argument('--list_dir', type=str, default='./lists/lists_Synapse', help='list dir') parser.add_argument('--num_classes', type=int, default=8, help='output channel of network') parser.add_argument('--max_iterations',...

lpc-97

类别

4

您好，请问一下如果我做10类别微调的话，num_classes设置成10还是11，因为我设置成10，黑背景也会产生类别预测，请问有什么方法解决嘛

514360393

你好，我这边在网上了解到的关于cudnn.benchmark设置为True一般是可以使得训练变快的，但是实际训练的时间反而变慢了很多，想知道这是什么原因导致的。 [09:40:27.459] Namespace(root_path='../MedDataset/H-sam/train_npz', output='test_output_deterministic0', dataset='Synapse', list_dir='./lists/lists_Synapse', num_classes=8, max_iterations=30000, max_epochs=200, stop_epoch=160, batch_size=12, n_gpu=1, deterministic=0, base_lr=0.005, img_size=512, seed=1234, vit_name='vit_b', ckpt='checkpoint/sam_vit_b_01ec64.pth', lora_ckpt=None, rank=4, warmup=True, warmup_period=250, AdamW=True, module='sam_lora_image_encoder', dice_param=0.8, is_pretrain=True, exp='Synapse_512') [09:40:27.465] 185 iterations...

bookpen

If I want to train on my own dataset, do I need to download and load the pretrained file called: sam_vit_b_01ec64.pth, instead of loading epoch_159.pth?

HANJIAC

请问为什么用自己数据集训练过程正常，但是加载lora_checkpoint后推理各项指标全为0？

kioic