0.4.0

版本发布时间: 2020-12-15 23:21:13

Project-MONAI/MONAI最新发布版本:1.4.0(2024-10-17 08:54:28)

Overview document for feature highlights in v0.4.0
Torchscript support for the net modules
New networks and layers:
- Discrete Gaussian kernels
- Hilbert transform and envelope detection
- Swish and Mish activation
- Acti-norm-dropout block
- Upsampling layer
- Autoencoder, Variational autoencoder
- FCNet
Support of initialisation from pre-trained weights for densenet, SENet, multichannel AHNet
Layer-wise learning rate API
New model metrics and event handlers based on occlusion sensitivity, confusion matrix, surface distance
CAM/GradCAM/GradCAM++
File format-agnostic image loader APIs with Nibabel, ITK readers
Enhancements for dataset partition, cross-validation APIs
New data APIs:
- LMDB-based caching dataset
- Cache-N-transforms dataset
- Iterable dataset
- Patch dataset
Weekly PyPI release
Fully compatible with PyTorch 1.7
CI/CD enhancements:
- Skip, speed up, fail fast, timed, quick tests
- Distributed training tests
- Performance profiling utilities
New tutorials and demos:
- Autoencoder, VAE tutorial
- Cross-validation demo
- Model interpretability tutorial
- COVID-19 Lung CT segmentation challenge open-source baseline
- Threadbuffer demo
- Dataset partitioning tutorial
- Layer-wise learning rate demo
- MONAI Bootcamp 2020

Base Docker image upgraded to nvcr.io/nvidia/pytorch:20.10-py3 from nvcr.io/nvidia/pytorch:20.08-py3

monai.apps.CVDecathlonDataset is extended to a generic monai.apps.CrossValidation with an dataset_cls option
Cache dataset now requires a monai.transforms.Compose instance as the transform argument
Model checkpoint file name extensions changed from .pth to .pt
Readers' get_spatial_shape returns a numpy array instead of list
Decoupled postprocessing steps such as sigmoid, to_onehot_y, mutually_exclusive, logit_thresh from metrics and event handlers, the postprocessing steps should be used accordingly before computing the metrics
ConfusionMatrixMetric and DiceMetric computation now returns an additional not_nans flag to indicate valid results
UpSample optional mode supports "deconv", "nontrainable", "pixelshuffle"; interp_mode is only used when mode is "nontrainable"
SegResNet optional upsample_mode now supports "deconv", "nontrainable", "pixelshuffle"
monai.transforms.Compose class inherits monai.transforms.Transform
In Rotate, Rotated, RandRotate, RandRotated transforms, the angle related parameters are interpreted as angles in radians instead of degrees.
SplitChannel and SplitChanneld moved from transforms.post to transforms.utility