Skip to content

CHANGELOG.md

Latest commit

 

History

History
54 lines (36 loc) · 2.41 KB

CHANGELOG.md

File metadata and controls

54 lines (36 loc) · 2.41 KB

CHANGE LOG

v1.2:

  • add standard deviation of DSC in coarse2fine_testing.py
  • Our codebase is also compatible with PyTorch 0.4.1.

v1.1:

  • Thank Qihang Yu for finding the bug which affects performance when batch > 1 in model.py and having fixed it elegantly.
  • remove the redundant clone() in model.py

v1.0:

  • make get_parameters in model.py more robust

v0.5:

  • add logs/ which contains training logs and testing results in FOLD #0. please see section 5
  • add RSTN pre-trained models in section 5
  • add oracle_testing.py & oracle_fusion.py to evaluate fine models. please see 4.6 & 4.7

v0.4:

  • we introduce epoch hyperparameter to replace max_iterations because the size of datasets varies.
    • Epoch dict {2, 6, 8} for (S, I, J) is intended for NIH dataset. You may modify it according to your dataset.
  • Add training_parallel.py to support multi-GPU training:
    • please see 4.3.4 section for details.
  • Simplify the bilinear weight initialization in ConvTranspose layer (issue #1)
  • Add coarse_fusion.py
  • training.py & training_parallel.py: print coarse/fine/average loss, giving more information of training loss
    • Thank Angtian Wang and Yingwei Li for finding bugs on multi-GPU training.

v0.3: no big improvements.

v0.2:

  • utils.py: two faster functions post_processing and DSC_computation are re-implemented in C for python3.6
    • give instructions in section 4.8.3 on how to compile fast_functions.i to get _fast_functions.so for different version of python like 3.5.
  • training.py : now trains by iterations instead of epoches, and learning rate will decay in J mode every 10k iterations.
    • performance of current version is 84.3% in NIH dataset, which is slightly lower than 84.4-84.6% in CAFFE implementation.

v0.1: init version.

Differences from OrganSegRSTN

Improvements:

  • We merge indiv_training.py, joint_training.py into training.py
  • We merge all *.prototxt to model.py
  • Our code runs almost twice faster than original one in CAFFE.
  • The minimum of DSC in test cases is a little higher (63.4%) than original minimum (62.8%).

Performance: in NIH Pancreas Dataset, average DSC is a little poorer (84.25% - 84.45%) than original one in CAFFE (84.4% - 84.6%).