Skip to content

Releases: rapidsai/cucim

v22.04.00

06 Apr 14:51
@GPUtester GPUtester
v22.04.00
8dfed80
Compare
Choose a tag to compare
v22.04.00

This version would be available through both Conda (https://anaconda.org/rapidsai/cucim) and PyPI package (https://pypi.org/project/cucim/22.04.00/).

cuCIM's GPUDirectStorage (GDS) API was introduced at GTC 2022 Spring "Accelerating Storage IO to GPUs with Magnum IO [S41347]" session on March 24.

cuCIM's GDS API examples are available at https://github.com/NVIDIA/MagnumIO/tree/main/gds/readers/cucim-gds.

In the below sections, ✨ means that the item is a new change in cuCIM's PyPI package, compared with the previous release (v22.02.01).

🚨 Breaking Changes

  • Apply fixes to skimage.transform scheduled for scikit-image 0.19.2 (#208) @grlee77

🐛 Bug Fixes

  • ✨ Fix ImportError from vendored code (#252) @grlee77
  • ✨ Fix wrong dimension in metadata (#248) @gigony
  • Handle file descriptor ownership and update documents for GDS (#234) @gigony
  • Check nullptr of handler in CuFileDriver::close() (#229) @gigony
  • Fix docs builds (#218) @ajschmidt8
  • Apply fixes to skimage.transform scheduled for scikit-image 0.19.2 (#208) @grlee77

🚀 New Features

  1. Randomization of transforms per image per batch (#231) @shekhardw

  2. Expose data type of CuImage object for interoperability with NumPy (#246) @gigony

1. Randomization of transforms per image per batch

Random Color Jitter transform implemented. Random Image Flip, Random Image Rotate90 and Random Zoom transforms are updated to apply transforms per image per batch. Execution of these transforms on per image per batch basis may result in increase in runtime.

2. Expose data type of CuImage object for interoperability with NumPy

  • CuImage object exposes typestr property.
  • DLDataType and DLDataTypeCode type is available under cucim.clara.

Prior to this change, it was not easy to convert CuImage's dtype (DLDataType) to NumPy's dtype and had to use the below workaround.

>>> from cucim import CuImage
>>> a = CuImage("notebooks/input/image.tif")
>>> b = a.read_region((0,0), (10,10))
>>> import numpy as np
>>> np.dtype(b.__array_interface__["typestr"]) # b would expose `__cuda_array_interface__` if memory is in GPU.
dtype('uint8')

With this change, we can convert the data type of CuImage to NumPy's dtype easily, and also can access cuCIM's DLDataType.

>>> from cucim import CuImage
>>> a = CuImage("notebooks/input/image.tif")
>>> b = a.read_region((0,0), (10,10))
>>> import numpy as np
>>> b.typestr
'|u1'
>>> np.dtype(b.typestr) == np.uint8
True
>>> from cucim.clara import DLDataType, DLDataTypeCode
>>> b.dtype == DLDataType(DLDataTypeCode.DLUInt, 8, 1)
True

🛠️ Improvements

  • Remove verbose plugin messages temporarily (only available in the PyPI package)
    • Address #109 ([BUG] - Info messages appearing as warnings in Jupyter notebooks)
    • Will support log level configuration later with the official PR.
  • Temporarily disable new ops-bot functionality (#239) @ajschmidt8
  • Add .github/ops-bot.yaml config file (#236) @ajschmidt8

📖 Documentation

Assets 2
Loading

v22.02.00

02 Feb 15:58
@GPUtester GPUtester
v22.02.00
29bfe6f
Compare
Choose a tag to compare
v22.02.00

Version 22.02.00 (Feb 02, 2022)

This version would be available through both Conda (https://anaconda.org/rapidsai/cucim) and PyPI package (https://pypi.org/project/cucim/22.02.00/).

🚨 Breaking Changes

  • Update cucim.skimage API to match scikit-image 0.19 (#190) @glee77

🐛 Bug Fixes

  • Fix a bug in v21.12.01 (#191) @gigony
    • Fix GPU memory leak when using nvJPEG API (when device='cuda' parameter is used in read_region method).
  • Fix segfault for preferred_memory_capacity in Python 3.9+ (#214) @gigony

📖 Documentation

🚀 New Features

  1. Update cucim.skimage API to match scikit-image 0.19 (#190) @glee77
  2. Support multi-threads and batch, and support nvJPEG for JPEG-compressed images (#191) @gigony
  3. Allow CuPy 10 (#195) @jakikham

1. Update cucim.skimage API to match scikit-image 0.19 (🚨 Breaking Changes)

channel_axis support

scikit-image 0.19 adds a channel_axis argument that should now be used instead of the multichannel boolean.

In scikit-image 1.0, the multichannel argument will likely be removed so we start supporting channel_axis in cuCIM.

This pulls changes from many scikit-image 0.19.0 PRs related to deprecating multichannel in favor of channel_axis. A few other minor PRs related to deprecations and updates to color.label2rgb are incorporated here as well.

The changes are mostly non-breaking, although a couple of deprecated functions have been removed (rgb2grey, grey2rgb) and a change in the default value of label2rgb's bg_label argument. The deprecated alpha argument was removed from gray2rgb.

Implements:

Update float32 dtype support to match scikit-image 0.19 behavior

Makes float32 and float16 handling consistent with scikit-image 0.19. (must functions support float32, float16 gets promoted to float32)

Deprecate APIs

Introduces new deprecations as in scikit-image 0.19.

Specifically:

  • selem -> footprint
  • grey -> gray
  • iterations -> num_iter
  • max_iter -> max_num_iter
  • min_iter -> min_num_iter

2. Supporting Multithreading and Batch Processing

cuCIM now supports loading the entire image with multi-threads. It also supports batch loading of images.

If device parameter of read_region() method is "cuda", it loads a relevant portion of the image file (compressed tile data) into GPU memory using cuFile(GDS, GPUDirect Storage), then decompress those data using nvJPEG's Batched Image Decoding API.

Current implementations are not efficient and performance is poor compared to CPU implementations. However, we plan to improve it over the next versions.

Example API Usages

The following parameters would be added in the read_region method:

  • num_workers: number of workers(threads) to use for loading the image. (default: 1)
  • batch_size: number of images to load at once. (default: 1)
  • drop_last: whether to drop the last batch if the batch size is not divisible by the number of images. (default: False)
  • preferch_factor: number of samples loaded in advance by each worker. (default: 2)
  • shuffle: whether to shuffle the input locations (default: False)
  • seed: seed value for random value generation (default: 0)

Loading entire image by using multithreads

from cucim import CuImage

img = CuImage("input.tif")

region = img.read_region(level=1, num_workers=8)  # read whole image at level 1 using 8 workers

Loading batched image using multithreads

You can feed locations of the region through the list/tuple of locations or through the NumPy array of locations.
(e.g., ((<x for loc 1>, <y for loc 1>), (<x for loc 2>, <y for loc 2>)])).
Each element in the location should be int type (int64) and the dimension of the location should be
equal to the dimension of the size.
You can feed any iterator of locations (dimensions of the input don't matter, flattening the item in the iterator once if the item is also an iterator).

For example, you can feed the following iterator:

  • [0, 0, 100, 0] or (0, 0, 100, 0) would be interpreted as a list of (0, 0) and (100, 0).
  • ((sx, sy) for sy in range(0, height, patch_size) for sx in range(0, width, patch_size)) would iterate over the locations of the patches.
  • [(0, 100), (0, 200)] would be interpreted as a list of (0, 0) and (100, 0).
  • Numpy array such as np.array(((0, 100), (0, 200))) or np.array((0, 100, 0, 200)) would be also available and using Numpy array object would be faster than using python list/tuple.
import numpy as np
from cucim import CuImage

cache = CuImage.cache("per_process", memory_capacity=1024)

img = CuImage("image.tif")

locations = [[0,   0], [100,   0], [200,   0], [300,   0],
             [0, 200], [100, 200], [200, 200], [300, 200]]
# locations = np.array(locations)

region = img.read_region(locations, (224, 224), batch_size=4, num_workers=8)

for batch in region:
    img = np.asarray(batch)
    print(img.shape)
    for item in img:
        print(item.shape)

# (4, 224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (4, 224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)

Loading image using nvJPEG and cuFile (GDS, GPUDirect Storage)

If cuda argument is specified in device parameter of read_region() method, it uses nvJPEG with GPUDirect Storage to load images.

Use CuPy instead of Numpy, and Image Cache (CuImage.cache) wouldn't be used in the case.

import cupy as cp
from cucim import CuImage

img = CuImage("image.tif")

locations = [[0,   0], [100,   0], [200,   0], [300,   0],
             [0, 200], [100, 200], [200, 200], [300, 200]]
# locations = np.array(locations)

region = img.read_region(locations, (224, 224), batch_size=4, device="cuda")

for batch in region:
    img = cp.asarray(batch)
    print(img.shape)
    for item in img:
        print(item.shape)

# (4, 224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (4, 224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)
# (224, 224, 3)

Experimental Results

We have compared performance against Tifffile for loading the entire image.

System Information
Experiment Setup

Benchmarked loading several images with Tifffile.

  • Use read_region() APIs to read the entire image (.svs/.tiff) at the largest resolution level.
    • Performed on the following images that use a different compression method
      • JPEG2000 YCbCr: TUPAC-TR-467.svs, 55MB, 19920x26420, tile size 240x240
      • JPEG: image.tif (256x256 multi-resolution/tiled TIF conversion of TUPAC-TR-467.svs), 238MB, 19920x26420, tile size 256x256
      • JPEG2000 RGB: [CMU-1-JP2K-33005.svs](https://opens...
Read more
Assets 2
Loading

v21.12.00

08 Dec 17:05
@GPUtester GPUtester
v21.12.00
6d1f082
Compare
Choose a tag to compare
v21.12.00

Version 21.12.00 (Dec 08, 2021)

This version would be available through both Conda (https://anaconda.org/rapidsai/cucim) and PyPI package (https://pypi.org/project/cucim/21.12.0/).

🚀 New Features

  1. Support Aperio SVS with CPU LZW and jpeg2k decoder (#141) @gigony
  2. Add NVTX support for performance analysis (#144) @gigony
  3. Normalize operation (#150) @shekhardw

1. Support Aperio SVS (.svs)

cuCIM now supports Aperio SVS format with help of OpenJpeg for decoding jpeg2k-compressed data.

Please check this notebook to see how to use the feature.

Unaligned Case (per_process, JPEG-compressed SVS file)

image

Unaligned Case (per_process, JPEG2000 RGB-compressed SVS file)

image

Unaligned Case (per_process, JPEG2000 YCbCr-compressed SVS file)

image

The detailed data is available here.

2. Add NVTX support for performance analysis

Important methods in cuCIM are instrumented with NVTX so can see performance bottlenecks easily with NSight systems.

Tracing can be enabled through config file or environment variable or through API and less than 1% performance overheads in normal execution.

Enabling Tracing

Through .cucim.json file
{
        "profiler" : { "trace": true }
}
Through Environment variable
CUCIM_TRACE=1 python
Through API
from cucim import CuImage

CuImage.profiler(trace=True)
#or
CuImage.profiler().trace(True)

CuImage.profiler().config
# {'trace': True}
CuImage.profiler().trace()
# True
CuImage.is_trace_enabled # this is simpler method.
# True

Profiling with NVIDIA Nsight Systems

nsys profile -f true -t cuda,nvtx,osrt -s cpu -x true --trace-fork-before-exec true -o my_profile `which python` benchmark.py
# can add `--stats true`

Then, execute nsight-sys to open the profile results (my_profile.qdrep).

image

With this feature, a bug in cuCIM v21.10.01 (thread contention in Cache) was found and fixed (#145).

3. Normalize operation

CUDA-based normalization operation is added. Normalization supports the following types.

  1. Simple range based normalization
  2. Arctangent based normalization

Arctangent-based normalization helps to stretch lower intensity pixels in the image slightly better than range-based normalization. If you look at its graph, there is a huge variation at certain lower intensities, but as intensities become higher, the curve becomes flatter. This helps in isolating regions like lungs (and regions within lungs) more efficiently. There can be separate use cases depending on the modality and the application.

Please check the test cases to see how you can use the operation.

🐛 Bug Fixes

  • Load libcufile.so with RTLD_NODELETE flag (#177) @gigony
  • Remove rmm/nvcc dependencies to fix cudaErrorUnsupportedPtxVersion error (#175) @gigony
  • Do not compile code with nvcc if no CUDA kernel exists (#171) @gigony
  • Fix a segmentation fault due to unloaded libcufile (#158) @gigony
  • Fix thread contention in Cache (#145) @gigony
  • Build with NumPy 1.17 (#148) @jakirkham

📖 Documentation

🛠️ Improvements

Assets 2
Loading

v21.10.00

07 Oct 00:26
@GPUtester GPUtester
v21.10.00
d7ac21f
Compare
Choose a tag to compare
v21.10.00

Version 21.10.00 (Oct 06, 2021)

This version would be available through both Conda (https://anaconda.org/rapidsai/cucim) and PyPI package (https://pypi.org/project/cucim/21.10.0/).

🚀 New Features

  1. Add transforms for Digital Pathology (#100) @shekhardw @chirayuG-nvidia
  2. Enable GDS and Support Runtime Context (__enter__, __exit__) for CuFileDriver and CuImage (#106) @gigony
  3. Add a mechanism for user to know the availability of cucim.CuImage (#107) @gigony
  4. Support raw RGB tiled TIFF (#108) @gigony

1. Add high-performance transforms for Digital Pathology (#100)

We have implemented the following methods in this release.

  • cucim.core.operations.color module
    • color_jitter
  • cucim.core.operations.intensity module
    • rand_zoom
    • scale_intensity_range
    • zoom
  • cucim.core.operations.spatial module
    • image_flip
    • image_rotate_90
    • rand_image_flip
    • rand_image_rotate_90

The above methods are exposed to MONAI (CuCIM Transforms #2932) through cucim.core.operations.expose.transform module

  • color_jitter
  • image_flip
  • image_rotate_90
  • scale_intensity_range
  • zoom
  • rand_zoom
  • rand_image_flip
  • rand_image_rotate_90

Special thanks to Chirayu Garg(@chirayuG-nvidia) for implementing C++ CUDA Kernels of the above transform methods!

2. Enable GDS and Support Runtime Context (__enter__, __exit__) for CuFileDriver and CuImage (#106)

Enabling GDS

GPU-Direct-Storage feature was disabled before (#3) because there was no public package available.

Now we re-enabled the feature. It's up to the user to install GDS libraries (GDS libraries are available through CUDA Toolkit >=11.x), or can install manually by following the instruction https://docs.nvidia.com/gpudirect-storage/troubleshooting-guide/index.html#gds-installing

If the GDS library is not available, the filesystem API still works with compatibility mode (our own implementation to do the same behavior).

Please see the Tutorial (Accessing File with GDS) for detail.

Support Runtime Context for CuFileDriver

CuFile Driver now can use Runtime Context so with statement is available.

with for cucim.clara.filesystem.open() would be available in the next release (v21.12.00)

from cucim.clara.filesystem import CuFileDriver
import cucim.clara.filesystem as fs
import os, cupy as cp, torch


# Assume a file ('nvme/input.raw') with size 10 in bytes

# Create a CuPy array with size 10 (in bytes)
cp_arr = cp.ones(10, dtype=cp.uint8)
# Create a PyTorch array with size 10 (in bytes)
cuda0 = torch.device('cuda:0')
torch_arr = torch.ones(10, dtype=torch.uint8, device=cuda0)

print(f"fs.is_gds_available(): {fs.is_gds_available()}")

# Using CuFileDriver
# (Opening a file with O_DIRECT flag is required for GDS)
fno = os.open("nvme/input.raw", os.O_RDONLY | os.O_DIRECT)
with CuFileDriver(fno) as fd:
    # Read 8 bytes starting from file offset 0 into buffer offset 2
    read_count = fd.pread(cp_arr, 8, 0, 2)
    # Read 10 bytes starting from file offset 3
    read_count = fd.pread(torch_arr, 10, 3)

# Another way of opening file with cuFile (>= v21.12.00)
# with fs.open("nvme/output.raw", "w") as fd:
#     # Write 10 bytes from cp_array to file starting from offset 5
#     write_count = fd.pwrite(cp_arr, 10, 5)
fd = fs.open("nvme/output.raw", "w")
# Write 10 bytes from cp_array to file starting from offset 5
write_count = fd.pwrite(cp_arr, 10, 5)
fd.close()

Support Runtime Context for CuImage

CuImage class didn't have close() method (file handle is close when the object is destructed) but now close() method is available.

Once the file handle for the CuImage object is closed, it cannot call operations that require file access.

from cucim import CuImage

with CuImage("image.tif") as cuimg:
  region = cuimg.read_region((0, 0), (100, 100))

print(region.shape)
region2 = cuimg.read_region((100, 100), (200, 200))
[Plugin: cucim.kit.cuslide] Loading...
[Plugin: cucim.kit.cuslide] Loading the dynamic library from: /ssd/repo/cucim/python/cucim/src/cucim/clara/cucim.kit.cuslide@21.10.00.so
[Plugin: cucim.kit.cuslide] loaded successfully. Version: 0
Initializing plugin: cucim.kit.cuslide (interfaces: [cucim::io::IImageFormat v0.1]) (impl: cucim.kit.cuslide)
[100, 100, 3]
Traceback (most recent call last):
  File "test_cucim.py", line 7, in <module>
    region2 = cuimg.read_region((100, 100), (200, 200))
RuntimeError: [Error] The image file is closed!
[Plugin: cucim.kit.cuslide] Unloaded.

3. Add a mechanism for users to know the availability of cucim.CuImage (#107)

Added cucim.is_available(module_name) method.
(module_name (str): Name of the module to check. (e.g. "skimage", "core", and "clara"))

Users can query if a specific module is available or not.

For example, if cucim.clara module which requires C shared library is not available:

>>> import cucim
>>> cucim.is_available()
False
>>> cucim.is_available("skimage")
True
>>> cucim.is_available("core")
True
>>> cucim.is_available("clara")
False

4. Support raw RGB tiled TIFF (#108)

Due to #17 ([BUG] Compression scheme 33003 tile decoding is not implemented), #19 (Check compression method used in the image) was merged so cuCIM has been handling only particular formats (only jpeg/deflate-compressed image).

If the input TIFF image has no-compressed RAW tile image, cuCIM showed the following error message:

RuntimeError: This format (compression: 1, sample_per_pixel: 1, planar_config: 1, photometric: 1) is not supported yet!.

https://github.com/Project-MONAI/MONAI/pull/2987/checks?check_run_id=3667530814#step:7:15641

Now cuCIM supports a raw RGB tiled TIFF with fast-path, re-allowing non-compressed image with slow-path (in case the input is not tiled RGB image).

This supports Project-MONAI/MONAI#2987.

🐛 Bug Fixes

📖 Documentation

  • Forward-merge branch-21.08 to branch-21.10 (#88) @jakirkham
  • Update PyPI cuCIM v21.08.01 README.md and CHANGELOG.md (#87) @gigony

🛠️ Improvements

Assets 2
Loading

v21.08.01

05 Aug 15:31
@GPUtester GPUtester
v21.08.01
a89f250
Compare
Choose a tag to compare
v21.08.01

Version 21.08.01 (August 10, 2021)

There are no new critical features in this release, but there are several enhancements/bug fixes.

We announced the cuCIM at the SciPy 2021 Conference (July 14). You can view the recorded video on YouTube (link).

🐛 Bug Fixes

  • Remove int-type bug on Windows in skimage.measure.label (#72) @grlee77
  • Fix missing array interface for associated_image() (#65) @gigony
  • Handle zero-padding version string (#59) @gigony
  • Remove invalid conda environment activation (#58) @ajschmidt8

📖 Documentation

🚀 New Features

🛠️ Improvements

  • Add SciPy 2021 to README (#79) @jakirkham
  • Use more descriptive ElementwiseKernel names in cucim.skimage (#75) @grlee77
  • Add initial Python unit/performance tests for TIFF loader module (#62) @gigony
  • Fix 21.08 forward-merge conflicts (#57) @ajschmidt8
Assets 2
Loading

v21.08.00

04 Aug 14:39
@GPUtester GPUtester
v21.08.00
2e2e1fd
Compare
Choose a tag to compare
v21.08.00

Version 21.08.00

There are no new critical features in this release, but there are several enhancements/bug fixes.

We announced the cuCIM at the SciPy 2021 Conference (July 14). You can view the recorded video on YouTube (link).

🐛 Bug Fixes

  • Remove int-type bug on Windows in skimage.measure.label (#72) @grlee77
  • Fix missing array interface for associated_image() (#65) @gigony
  • Handle zero-padding version string (#59) @gigony
  • Remove invalid conda environment activation (#58) @ajschmidt8

📖 Documentation

🚀 New Features

🛠️ Improvements

  • Add SciPy 2021 to README (#79) @jakirkham
  • Use more descriptive ElementwiseKernel names in cucim.skimage (#75) @grlee77
  • Add initial Python unit/performance tests for TIFF loader module (#62) @gigony
  • Fix 21.08 forward-merge conflicts (#57) @ajschmidt8
Assets 2
Loading

v21.06.00

09 Jun 16:23
@GPUtester GPUtester
v21.06.00
0886c13
Compare
Choose a tag to compare
v21.06.00

Version 21.06.00 (June 8, 2021)

🚀 What's new?

  1. Change version scheme from SemVer to CalVer (#32)
  2. Implement cache mechanism (#48)
  3. Support __cuda_array_interface__ (#48)

Cache Mechanism

cuCIM now supports the cache mechanism for the unaligned use case.

image
image

Please check this notebook to see how to use the feature.

Support __cuda_array_interface__

In addition to the __array_interface__, cuCIM now supports __cuda_array_interface__ protocol.

Please check this notebook to see how to use the feature.

🐛 Fixes

  • Check compression method used in the image for SVS format (#19)
  • Avoid use of deprecated cp.float, cp.int, cp.bool (#29)
  • Fix a memory leak in Deflate decoder (#48)
  • Update update-version.sh (#42) @ajschmidt8

🚀 Improvements

  • Use CuPy 9 (#20)
    • CuPy 9 is officially released so you don't need to install the release candidate (e.g., 9.0.0b3) anymore.
  • Update environment variable used to determine cuda_version (#43) @ajschmidt8
  • Update version script to remove bump2version dependency (#41) @gigony
  • Update changelog (#40) @ajschmidt8
  • Update docs build script (#39) @ajschmidt8

Contributors

gigony and ajschmidt8
Assets 2
Loading

v0.19.0

21 Apr 20:56
@GPUtester GPUtester
v0.19.0
fc9803f
This commit was signed with the committer’s verified signature.
raydouglass Ray Douglass
GPG key ID: 9931FB82CEDBB451
Learn about vigilant mode.
Compare
Choose a tag to compare
v0.19.0 
REL v0.19.0
Assets 2
Loading