Skip to content

README.md

Latest commit

 

History

History
85 lines (49 loc) · 3.33 KB

README.md

File metadata and controls

85 lines (49 loc) · 3.33 KB
Raw

Classification

Files in the root directory are used for the experiments in the paper.

  • main.py – main function for training.

  • usage: main.py [-h] [--dataset {1,2,3}]
               [--architecture {ResNet50,DenseNet201}] [--pretrain {0,1}]
               [--fold {1,2,3}] [--index INDEX] [--mode {train,test}]
               [--savefile {0,1}]

Framework for training and evaluation.

optional arguments:
  -h, --help            show this help message and exit
  --dataset {1,2,3}, -d {1,2,3}
                        1 -- smear baseline, 2 -- smear pipeline, 3 -- LBC
                        pipeline
  --architecture {ResNet50,DenseNet201}, -a {ResNet50,DenseNet201}
                        choose a network architecture
  --pretrain {0,1}, -p {0,1}
                        use pre-trained weights on ImageNet
  --fold {1,2,3}, -f {1,2,3}
                        Dataset 1&2: 3 folds; Dataset 3: 2 folds.
  --index INDEX, -i INDEX
                        index for multiple training to get STD
  --mode {train,test}, -m {train,test}
                        train or test
  --savefile {0,1}, -s {0,1}
                        if save results to csv files

  • collect_results.sh – bash script to run the experiments.

  • make_results.py – make results table in the paper, and table for the plots in the paper.

In each dataset, the ResNet_aug{1-3}.hdf5 in each ./logs{1-{2|3}}/ represent 3 models trained for 2 or 3 folds.

Here are some common files exists in all 3 datasets' directory.

  • data_split.sh – to split data from ../Patches/Z_focused/ into ../Patches/data_train{1-3}/ and ../Patches/data_test{1-3}/, should be adjusted to each dataset before using.
  • randomfilerenamer.py – rename files in a directory randomly. This should be used on the training data before training. It is called in the data_split.sh.

Dataset 1: old smear data

Dataset 1 is relatively small. It can be trained and evaluated fast on local laptops. Here contains the code to use locally and also for some attempts:

  • ./slide_mosaic1_oldsmear/ – mosaics consists of 10 * 10 patches from each glass, in order to give an overview of the dataset.

Dataset 2: new smear data

Dataset 2 is specifically used for evaluating the impact of

  • focus level
  • color channel

on the classification results. Here contains several files:

  • focus_augmentation.py – to generate out-of-focus patches in separate folders like ./Z_focused_n/.
  • data_split_focusaug_train.sh – to copy data at several focus levels into training folder, and rename the data randomly by calling randomfilerenamer.py.
  • data_split_focusaug_test.sh – to split data in ./Z_focused_n/s into different testing folders like ./FocusAugEvaluation/data_test_n/.
  • focusaug_evaluate.py – to evaluate the performance of a trained model on test sets at different focus levels and write the results into ./focus_results_data2fold1.csv .
  • extractG.py – to exact Green channel of the data, and save the original RGB data to backup folders.
  • ./slide_mosaic2_newsmear/ – mosaics consists of 16 * 16 patches from each glass, in order to give an overview of the dataset.

Dataset 3: LBC data

  • ./slide_mosaic3_LBC/ – mosaics consists of 16 * 16 patches from each glass, in order to give an overview of the dataset.