Issues
- 0
Add serialization/deserialization unit tests - checks whether writing model snapshots to disc actually works properly
#303 opened by yalaudah - 0
prepare_dutchf3.py vertical sample locations drops patches at the bottom of the volume - leads to worse results in that region
#259 opened by yalaudah - 0
make output transform unpad and scale down mask in penobscot metrics - facilitates correctly reported inline mIoU
#276 opened by maxkazmsft - 0
add code coverage badge to github front page - facilitates visual certainty in unit test code coverage
#316 opened by maxkazmsft - 0
merge staging into master and cut a 0.1.3 release - keeps master as latest stable branch
#314 opened by maxkazmsft - 0
Generated data splits should be tracked along with model outputs, and not stored with data.
#286 opened by yalaudah - 1
debug test.py returns zero output - determine why deconvnet and unet model snapshots produce zero output (same class predictions)
#304 opened by yalaudah - 0
- 1
run master and provide feedback on state of the repo - makes repo friendly to new users
#315 opened by maxkazmsft - 0
- 0
create dev/user Docker version - enables repo to work in platform-agnostic way
#255 opened by maxkazmsft - 0
create script to read build yaml file and trigger builds on the dev machine - enables developer to test builds locally without build machine
#305 opened by maxkazmsft - 0
correctness enhancements (data dumps and data generators) - facilitates a better debugging experience for the user
#297 opened by maxkazmsft - 0
Add unit tests for _patch_label_2d()
#301 opened by yalaudah - 0
debug correctness dataset metrics and experiment performance for all models - facilitates correct metrics which are bug-free
#289 opened by maxkazmsft - 0
replace decode_segmap custom code with qualitative matplotlib colormap - fixes max number of classes to be 6 and provides a consistent colormap across the repo
#285 opened by maxkazmsft - 0
Host DutchF3 on Azure open datasets - overcomes reliability problem of data availability
#279 opened by sharatsc - 0
add Azure ML test build pipeline status badges to main readme - users are able to see that status of AML tests
#241 opened by maxkazmsft - 1
make sure train.py and test.py data augmentation is consistent - addressed correctness of model performance during scoring
#270 opened by maxkazmsft - 1
test HRNet section depth patch-based model on centos - facilitates correct model behavior on redhat systems
#254 opened by maxkazmsft - 0
read model parameters from configuration file for texture_net - facilitates easy model architecture changes
#277 opened by maxkazmsft - 0
add distributed model scoring for test.py methods - facilitates faster scoring if multiple GPUs are available
#268 opened by maxkazmsft - 0
make sure we can download all models provided in the repo in pre-trained form - facilitates pre-trained model consistency
#267 opened by maxkazmsft - 0
check what type of data augmentation pre-trained models have - prevents making a mistake of using a model with wrong augmentation
#266 opened by maxkazmsft - 0
remove hardcoded model selection logic for pre-trained modles - facilitates more intuitive selection of model with no code (no effect on functionality)
#265 opened by maxkazmsft - 0
remove unused configuration parameters - makes it intuitive for the user to modify configuration files
#248 opened by maxkazmsft - 0
experiment builds take almost an hour to run - facilitates faster turn-around time for PRs with parallel builds
#245 opened by maxkazmsft - 0
take in notebook changes batch size - user sees batch size of 10 when take is used
#238 opened by maxkazmsft - 1
hotfix master branch with accurate documentation regarging penobscot dataset - allows user to more carefully gauge the performance of the repo
#235 opened by maxkazmsft - 0
dutch F3 patch test.py and notebook: refactor helper functions into common experiment utilities - helps the user navigate the codebase and removes code duplication
#234 opened by maxkazmsft - 0
add test.py visualizations for dutchf3_patch experiment - allows the user to see how the patches are scored on the test set
#229 opened by maxkazmsft - 1
extend correctness work to current Dutch F3 notebook - shows correctness test and dataset running properly
#288 opened by maxkazmsft - 0
- 0
detect and disable rotation patch augmentations - facilitates correct physics of the learned model
#275 opened by maxkazmsft - 0
remove hardcoding when picking splits in TestPatchLoader - facilitates correctness of reported test set results
#274 opened by maxkazmsft - 0
explore whether we should shift the patches or pad them - facilitates better training data for patch-based models
#273 opened by maxkazmsft - 0
explore correctness of constants for mean and std - effects all models pre-trained on imagenet
#269 opened by maxkazmsft - 0
correctness train tests - adds confidence to the current set of experiments with ignite
#249 opened by maxkazmsft - 0
enable switch of crossline or inline training and scoring - easy fix for training set leakage
#253 opened by maxkazmsft - 2
Dutch F3 validation data partially overlaps with training data, leads to overly inflated validation results in Tensorboard
#233 opened by yalaudah - 1
snapshot ignite dev version - enables to continue running on dev version of ignite if goes obsolete
#252 opened by maxkazmsft - 0
Docker file prepare Dutch F3 error on staging - user cannot use Docker file
#237 opened by maxkazmsft - 0
determine why RedHat trained models are scoring only a single class - provides insight into DNN models run on other OSs
#284 opened by maxkazmsft - 0
Go through TODOs reformulate into issues - facilitates correctness tracking of any outstanding items in the code
#280 opened by maxkazmsft - 2
Notebooks not using same seed as experiments, affects reproducibibilty across notebooks and experiments
#244 opened by yalaudah - 1
interpret the padding and validation set mIoU vs predictoins - currently the quality of output predictions doesn't match the mIoU metric
#230 opened by maxkazmsft - 1
Refactoring train.py to enhance readability, maintainability, and add training results to Tensborboard.
#236 opened by yalaudah - 1
explore whether a generator can be used with voxelloader - facilitates lower memory footprint when loading voxel data
#272 opened by maxkazmsft - 0
- 0
Dutch F3 patch HRNet notebook and train cleanup uber item - reduces user friction on this repo and facilitates consistency between Dutch F3 patch experiment and notebook
#226 opened by maxkazmsft