-- Bug fixes -- Adding Code-based Tutorial (https://matjesg.github.io/deepflash2/tutorial_monuseg.html)
-
new features (#42), thanks to @matjesg
- Multiclass GT Estimation, closes #34
- Torchscript ensemble class for inference / tta adjusted
- ONNX export possible
-
Major changes
- Different classes for training (EnsembleLearner) and Inference (EnsemblePredictor)
- Normalization based on uin8 images (0...255)
-
Instance segmentation metrics (#32), thanks to @matjesg
- Instance segmentation metrics for GT estimation and prediction
- Dependency fix (opencv-python-headless>=4.1.1,<4.5.5)
- Minor improvements and fixes
- Fixes bug when predicting on new data.
- Removing git dependency for cellpose installation.
- Bug fixes and minor improvements.
- Instance segmentation metrics (#32), thanks to @matjesg
-
- Instance segmentation metrics for GT estimation and prediction
-
- Tutorials
-
Adding tifffile imread (#13), thanks to @matjesg
- Closes #12
-
Multichannel tiffs are imported as single channel greyscale when imported via ui (#12)
- Problem: .tiff images are with shape z,c,y,x = 1,4,1024,2048 are selected in the user deepflash google colab user interface. Image preview only shows first channel. After inspecting the image import, I saw, that the tiffile has dimensions 1024x2048 after import.
Proposed solution: Images may be imported with tifffile.imread instead of imageio.imread. Here, the channel dimensions are preserved.
- Real-Time loss weight computation and large file support (#11), thanks to @matjesg
-
- Real-Time loss weight computation via fast convolutional distance transform on GPU
-
- temp storage using zarr instead of RAM
- zarr dependency
- Prediction not possible on Windows machines due to cuda error (#10)
- During prediction on a windows machine a a OS Error occurs due to:
"RuntimeError: cuda runtime error (801) : operation not supported at ..\torch/csrc/generic/StorageSharing.cpp:247"
Problem: Storage sharing currently not supported on windows.
Proposed solution: Ensemble learner takes "num_workers" argument and passes it to subsequent functions. If num_workers == 0, prediction works for me.
- Bug fixes and minor improvements.
- Type Error when starting training (#5)
- When I try to start the training process in google colab, this error occurs:
TypeError: no implementation found for 'torch.nn.functional.cross_entropy' on types that implement torch_function: [<class 'fastai.torch_core.TensorImage'>, <class 'fastai.torch_core.TensorMask'>]
as well as
FileNotFoundError: [Errno 2] No such file or directory: 'models/model.pth'
in the end.
Hope you know whats the problem.
- Type Error when starting training (#5)
- When I try to start the training process in google colab, this error occurs:
TypeError: no implementation found for 'torch.nn.functional.cross_entropy' on types that implement torch_function: [<class 'fastai.torch_core.TensorImage'>, <class 'fastai.torch_core.TensorMask'>]
as well as
FileNotFoundError: [Errno 2] No such file or directory: 'models/model.pth'
in the end.
Hope you know whats the problem.