MIP Candy brings ready-to-use training, inference, and evaluation pipelines together with aesthetics, so you can focus on your experiments, not boilerplate.
MIP Candy powers research across top institutions.
Designed for modern medical image research pipelines.
We carefully designed the console outputs to be informative and actionable, providing insights into model performance and training progress. Not only scalar metrics are printed, we also track their trends over epochs and show per-case validation metrics.
Sometimes even when all metrics seem perfect, the model can still be misleading. MIP Candy does its best to help you "see" the model's performance by providing powerful built-in visualization utilities.
Visualizing 3D volumes, which are widespread in medical imaging and other 3D data analysis tasks, can be challenging. MIP Candy provides simple and intuitive 3D visualization interfaces that allow you to visualize 3D volumes in a user-friendly manner. These methods are well sculpted to ensure negligible overhead added to the training process.
Everything saved into metrics will be automatically plotted and saved into the trainer folder.
Connect to Notion, Weights & Biases, and TensorBoard for rich experiment tracking and sharing with your team.
As pipelines grow bigger and bigger, memory leakage and computational overhead become invisible problems. Our built-in profiler helps you investigate into sub-epoch level loggings to find the culprit.
Start from SegmentationTrainer and only implement the network construction. MIP Candy handles data flow, loss computation, augmentation, checkpointing, and evaluation out of the box. Learn all you can do in MIP Candy Docs.
from typing import override
from torch import nn
from mipcandy import SegmentationTrainer
class MyTrainer(SegmentationTrainer):
@override
def build_network(self, example_shape: tuple[int, ...]) -> nn.Module:
...
Explore an interactive MIP Candy frontend demo directly in Notion.
Download a dataset, create a dataset wrapper, and hand it to a trainer. Below is an example using the ACDC dataset. The example code to the right replicates most of nnU-Net's features but without augmentations.
from typing import override
from os.path import exists
from monai.networks.nets import DynUNet
from torch import nn
from torch.utils.data import DataLoader
from mipcandy import SegmentationTrainer, AmbiguousShape, auto_device, download_dataset, NNUNetDataset, inspect, \
load_inspection_annotations, RandomROIDataset
class UNetTrainer(SegmentationTrainer):
@override
def build_network(self, example_shape: AmbiguousShape) -> nn.Module:
kernel_size = [[3, 3, 3]] * 5
strides = [[1, 1, 1]] + [[2, 2, 2]] * 4
return DynUNet(spatial_dims=3, in_channels=example_shape[0], out_channels=self.num_classes,
kernel_size=kernel_size, strides=strides, upsample_kernel_size=strides,
deep_supervision=self.deep_supervision, deep_supr_num=2, res_block=True)
if __name__ == "__main__":
device = auto_device()
download_dataset("nnunet_datasets/ACDC", "tutorial/datasets/ACDC")
dataset = NNUNetDataset("tutorial/datasets/ACDC", align_spacing=True)
if exists("tutorial/datasets/ACDC/annotations.json"):
annotations = load_inspection_annotations("tutorial/datasets/ACDC/annotations.json", dataset)
else:
dataset.device(device=device)
annotations = inspect(dataset)
dataset.device(device="cpu")
annotations.save("tutorial/datasets/ACDC/annotations.json")
dataset = RandomROIDataset(annotations, 2)
train, val = dataset.fold()
train_loader = DataLoader(train, 2, True, num_workers=2, prefetch_factor=2, persistent_workers=True)
val_loader = DataLoader(val, 1, False)
trainer = UNetTrainer("tutorial", train_loader, val_loader, device=device)
trainer.train(1000, note="example with the ACDC dataset")
MIP Candy requires Python ≥ 3.12. Install the standard bundle from PyPI:
MIP Candy Bundles provide verified model architectures with corresponding trainers and predictors. You can install it along with MIP Candy.
Jump into the ecosystem around MIP Candy.