Add cvpr workshop

vista3d/cvpr_workshop/README.md

@@ -15,20 +15,27 @@ limitations under the License.
 This repository is written for the "CVPR 2025: Foundation Models for Interactive 3D Biomedical Image Segmentation"([link](https://www.codabench.org/competitions/5263/)) challenge. It 
 is based on MONAI 1.4. Many of the functions in the main VISTA3D repository are moved to MONAI 1.4 and this simplified folder will directly use components from MONAI.
 
-It is overly simplied to train interactive segmentation models across different modalities. The sophisticated transforms and recipes used for VISTA3D are removed. 
+
+It is simplified to train interactive segmentation models across different modalities. The sophisticated transforms and recipes used for VISTA3D are removed. The finetuned VISTA3D checkpoint on the challenge subsets is available [here](https://drive.google.com/file/d/1r2KvHP_30nHR3LU7NJEdscVnlZ2hTtcd/view?usp=sharing)
 
 # Setup
 ```
 pip install -r requirements.txt
 ```
 
 # Training
-Download VISTA3D pretrained checkpoint or from scratch. Generate a json list that contains your traning data.
+Download the challenge subsets finetuned [checkpoint](https://drive.google.com/file/d/1r2KvHP_30nHR3LU7NJEdscVnlZ2hTtcd/view?usp=sharing) or VISTA3D original [checkpoint]((https://drive.google.com/file/d/1DRYA2-AI-UJ23W1VbjqHsnHENGi0ShUl/view?usp=sharing)). Generate a json list that contains your traning data and update the json file path in the script.
 ```
 torchrun --nnodes=1 --nproc_per_node=8 train_cvpr.py
 ```
 
 # Inference
-We provide a Dockerfile to satisfy the challenge format. For more details, refer to the [challenge website]((https://www.codabench.org/competitions/5263/))
+You can directly download the [docker file](https://drive.google.com/file/d/1r2KvHP_30nHR3LU7NJEdscVnlZ2hTtcd/view?usp=sharing) for the challenge baseline.
+We provide a Dockerfile to satisfy the challenge format. For more details, refer to the [challenge website]((https://www.codabench.org/competitions/5263/)).
+```
+docker build -t vista3d:latest .
+docker save -o vista3d.tar.gz vista3d:latest
+```
+
 
 

vista3d/cvpr_workshop/train_cvpr.py

@@ -104,23 +104,24 @@ def __getitem__(self, idx):
         return data
 # Training function
 def train():
+    json_file = "subset.json"  # Update with your JSON file
     epoch_number = 100
-    start_epoch = 30
+    start_epoch = 0
     lr = 2e-5
     checkpoint_dir = "checkpoints"
+    start_checkpoint = '/workspace/CPRR25_vista3D_model_final_10percent_data.pth'
     os.makedirs(checkpoint_dir, exist_ok=True)
     dist.init_process_group(backend="nccl")
     world_size = int(os.environ["WORLD_SIZE"])
     local_rank = int(os.environ["LOCAL_RANK"])
     torch.cuda.set_device(local_rank)
     device = torch.device(f"cuda:{local_rank}")
-    json_file = "subset.json"  # Update with your JSON file
     dataset = NPZDataset(json_file)
     sampler = torch.utils.data.distributed.DistributedSampler(dataset, num_replicas=world_size, rank=local_rank)
     dataloader = DataLoader(dataset, batch_size=1, sampler=sampler, num_workers=32)
     model = vista3d132(in_channels=1).to(device)
-    # pretrained_ckpt = torch.load('/workspace/VISTA/vista3d/bundles/vista3d/models/model.pt', map_location=device)
-    pretrained_ckpt = torch.load(os.path.join(checkpoint_dir, f"model_epoch{start_epoch}.pth"))
+    pretrained_ckpt = torch.load(start_checkpoint, map_location=device)
+    # pretrained_ckpt = torch.load(os.path.join(checkpoint_dir, f"model_epoch{start_epoch}.pth"))
     model = DDP(model, device_ids=[local_rank], find_unused_parameters=True)
     model.load_state_dict(pretrained_ckpt['model'], strict=True)
     optimizer = optim.AdamW(model.parameters(), lr=lr, weight_decay=1.0e-05)