myron

PersistentDataset, CacheDataset improvements

5

It would nice to have several improvements to caching of data on disk and in memory. For caching on disk with **PersistentDataset**, - the first epoch takes a long time,...

import cv2 and DistributedDataParallel. (FIND was unable to find an engine)

13

EDIT: the bug is reproducable in the newest nvidia/pytorch:22.09-py3 docker container, but is not reproducible in older container (older pytorch/cudnn) Something in MetaTensor makes DistributedDataParallel fail (this is in addition...

fixes DiceCELoss for multichannel targets

1

Fixes DiceCELoss for multichannel targets. Currently if "target" (ground truth label) is provided as a multichannel data (each channel is binary or float), then current DiceCELoss attempts to convert it...

updates to Gaussian map for sliding window inference

4

This updates the calculation of Gaussian map (weights) during sliding window inference with "gaussian" Current version had multiple small issues - it computed Guassian weight map (image) via nn.Conv1d sequence...

This introduces a new SharedCacheDataset(Dataset) class to address https://github.com/Project-MONAI/MONAI/issues/3843 It has several improvements / changes compared to the CacheDataset https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/dataset.py#L690 - It caches into a shared memory in RAM, so...

This adds improvements to SegResNet semantic segmentation to include deep supervision feature (among other things). I've decided to add a new network named SegResnetDS instead of replacing the existing SegResNet,...

SlidingWindowInferer: option to adaptively stitch in cpu memory for large images. This adds an option to provide maximum input image volume (number of elements) to dynamically change stitching to cpu...

MetaTensor and DistributedDataParallel. bug (SyncBatchNormBackward is a view and is being modified inplace)

3

When upgrading from MONAI 0.9.0 to 1.0.0, my 3D segmentation code fails due to (most likely) new MetaTensor in transforms, when using DistributedDataParallel (multi-gpu) the error is `RuntimeError: Output 0...

get_norm_layer default to affine=True

2

https://github.com/Project-MONAI/MONAI/issues/5222 Sets the default normalization parameter (affine) to True. Most normalizations in pytorch by default estimates the affine parameters (e.g. Batchnorm, Groupnorm). One notable exception is Instancenorm (for which default...

For many monai networks (including Unet) we use the default normalization - Instancenorm , using a short string 'instance' or Norm.INSTANCE https://github.com/Project-MONAI/MONAI/blob/fb6d95058c26bcbaaf44e5e8cb75625ebf9f67ac/monai/networks/nets/unet.py#L124 But the way it's implemented via Monai 'factory',...