Commits

Breadcrumbs

History for

onmain

User selector

All users

All time

Commit History

Commits on Apr 10, 2021

• Small tweaks in create_data/train_seg/eval_seg to produce the best segmentation model. Also added a jupyter notebook example for training deep segmentation models in Google Colab.

  andreped

  committed
  b8307cb

  Copy full SHA for b8307cb

• small refactoring

  andreped

  committed
  ddc4c8b

  Copy full SHA for ddc4c8b

• Improved segmentation result print using PrettyTable, and added example figure showing segmentation results.

  andreped

  committed
  dd7939d

  Copy full SHA for dd7939d

• Major revision. Added support for training and evaluating segmentation models. Current supported architectures are U-Net and ResU-Net++. The networks and training sets are somewhat tuned, focusing …

  Show description for 957dae2
  andreped

  committed
  957dae2

  Copy full SHA for 957dae2

Commits on Apr 3, 2021

• Fixed bug in U-Net related to the usage of spatial_dropout when performing multi-class semantic segmentation. Augmentation wasn't being applied during training, added it now. Added improved trainin…

  Show description for 8394f61
  andreped

  committed
  8394f61

  Copy full SHA for 8394f61

Commits on Apr 2, 2021

• Fixed allow_growth-usage (TF-related)

  andreped

  committed
  75a90bc

  Copy full SHA for 75a90bc

• Added pipeline for training multi-class segmentation models (using U-Net currently), with preprocessing script for open data set, as well as data augmentation methods, batch generator, evaluation s…

  Show description for ca479e2
  andreped

  committed
  ca479e2

  Copy full SHA for ca479e2

Commits on Mar 30, 2021

• Updated notebook and lowered number of epochs for convenience in the notebook verbose.

  andreped

  committed
  76a4d73

  Copy full SHA for 76a4d73

• Changed paths to use respective paths, instead of hardcoded ones, assuming that the code is ran from the python/ folder, this should run fine. Also added a jupyter notebook project, relevant for de…

  Show description for 881b6de
  andreped

  committed
  881b6de

  Copy full SHA for 881b6de

Commits on Mar 27, 2021

• minor edits

  andreped

  committed
  90bfb9f

  Copy full SHA for 90bfb9f

• observed memory leak in data augmentation. Disabled for now.

  andreped

  committed
  8c6d696

  Copy full SHA for 8c6d696

• data augmentation now works. However, haven't spent a lot of time finding the optimal augmentation design.

  andreped

  committed
  abdcedf

  Copy full SHA for abdcedf

• Now both basic CNN and multi-output classifiers work as intended, both from training and evaluation. Class imbalance is still a problem on the multi-class task.

  andreped

  committed
  478cbbe

  Copy full SHA for 478cbbe

• Added support and interface for multi-class and multi-task (output) models. I have tested both and they work. However, need to find the distribution of the 5 classes to balance them properly.

  andreped

  committed
  4a129f5

  Copy full SHA for 4a129f5

• Added convenience function to utils, and removed redundant script

  andreped

  committed
  200f863

  Copy full SHA for 200f863

• Added some figures illustrating some preliminary results. Also introduced an MTL model which will be used in the future on the kaggle dataset.

  andreped

  committed
  e6d5c83

  Copy full SHA for e6d5c83

• Added option to train multi-class model using the kaggle dataset if of interest.

  andreped

  committed
  699dc3a

  Copy full SHA for 699dc3a

Commits on Mar 26, 2021

• Introduced pretrained model and finetuning for improving convergence, added augmentation, tuned the approach and architecture futher. I have reached close to 90% macro average F1 score now, without…

  Show description for 0d92ca2
  andreped

  committed
  0d92ca2

  Copy full SHA for 0d92ca2

• Implemented the full pipeline, from preprocessing, training to evaluating the trained model, in TensorFlow 2.4.

  andreped

  committed
  e63dd62

  Copy full SHA for e63dd62

• Added relevant scripts, and added useful tips to README.

  andreped

  committed
  3a4c75f

  Copy full SHA for 3a4c75f

Loading