Skip to content

Version 3.5.11

Choose a tag to compare

View all tags
@danifranco danifranco released this 17 Mar 08:34
· 961 commits to master since this release
v3.5.11
5d4f7bf

Major:

  • Adapt DFCAN to 3D and add SSIM-based loss
  • Adapt RCAN to 3D
  • Ensure appropiate data range when calculating metrics in SR, SSL and I2I workflows.
  • Implement more data/sample filtering methods: 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio'
  • Distribute better loading data code into classes
  • Add in all most code Typing checks for Pydantic/Pylance
  • Rename 'custom' normalization into 'zero_mean_unit_variance'
  • Separate into a class the normalization module to reduce repeated code
  • Avoid doing normalization for test GT data in I2I, SR, SSL and Denoising
  • Organize metrics for I2I, SR and SSL depending on train and test
  • Avoid creating X data during Instance segmentation saving disk space
  • Improve Detection workflow when multihead output is created
  • Add SSIM, W_MAE_SSIM and W_MSE_SSIM losses for I2I, SR and SSL workflows
  • Move some variables from TEST.BY_CHUNKS to DATA.TEST

Minor:

  • Change 2D image load to be more robust
  • BMZ connection:
    • Add task description option
    • Add model version
    • Change env.yaml created
    • Change BMZ model import message error
    • Add more Instance segmentation model support when consuming BMZ models
  • Allow multiple ddp runs without closing the process group initialization
  • Add script to change a parameter in the RDF file of a BMZ model
  • Update DDP messages and wait points
  • Remove mixed precision in evaluation
  • Add find_unused_parameters when using resunet_se with DDP
  • Add more information while gathering training/validation data
  • Add more options to tune RCAN model
  • Add script for blur estimation
  • Add script to measure similarity metrics common in I2I and SR workflows
  • Add tensor conversion in metric calculation
  • Add changes to convert_old_model_cfg_to_current_version in order to convert old configurations into new
  • Upgrade filtering saving examples of the patches/images filtered so the user can check them
  • Change DATA.FILTER_BY_IMAGE default value to False
  • Wrap rotate function to allow float16

Bugs fixed:

  • Fix minor bug in instance segmentation when test is not enabled and test path does not exist
  • Instance masks folder name update to not set always contour info
  • Solve bugs in U-Next V1 and U-Next V2 models when using more than one channel
  • Fix 3D U-NeXt models
  • BMZ connection:
    • Resize cover to ensure the shape
    • Avoid adding duplicate tags
    • Ensure only a patch is taken for BMZ input when working with H5/Zarr files
    • Ensure only pytorch_state_dict models are consumed
  • Correct C channel activation during Instance segmentation when it is used alone.
  • Fix minor error during MODEL.BMZ.EXPORT.DATASET_INFO check
  • Solve minor bug in convert_old_model_cfg_to_current_version function
  • Fix conversion to RGB in generators
  • Fix problem with TEST.REDUCE_MEMORY
  • Calculate metrics when reusing predictions (TEST.REUSE_PREDICTIONS). They are calculated as "merge_patches"
  • Allow SSL pretrainings during model check
  • Add restrictions in SR, SSL and I2I workflows to not use wdsr in 3D

Full Changelog: v3.5.10...v3.5.11

Assets 2
Loading