Deep-Learning-AI-on-HPC-clusters

1 Image patch extraction
2 Prediction
3 Heatmap stitching
4 Retreiving run-time statistics

1 Image patch extraction

• Adjust configuration parameters in files config_testing.txt, config_normal.txt and config_tumor.txt.

• Run the commands listed in the following subsections to launch Son of Grid Engine (SGE) jobs to extract, group patches in HDF5 files and create a lookup table for every HDF5 file.

  1.1 Extract and group

• qsub split_main.sh ./config_testing.txt
  -- split_main.sh in turn submits split_grp.sh which in turns runs split_grp.py in array job fashion. Every task in the array job processes one slide.

• qsub split_main.sh ./config_normal.txt

• qsub split_main.sh ./config_tumor.txt

1.2 Create lookup tables

• bash create_lookup_grp.sh ./config_testing.txt
  
• bash create_lookup_grp.sh ./config_normal.txt
  
• bash create_lookup_grp.sh ./config_tumor.txt
  

The lookup tables are created only once and used at Prediction stage for launching array job tasks. These tasks are run in parallel and scalable manner - if there are not enough resourcs for running all tasks then they are queued up automatically and started as resources become available. Each task processes only one group.

1.3 Retreiving wall-clock time statistics for extraction and grouping

Run time_stats_sg_V2.sh Linux scrip with two arguments: a) type of slides (test, normal or tumor) and b) the root directory of the results:

• time bash time_stats_sg_V2.sh [test | normal | tumor] ROOT_DIR
  

An example run:

• time bash wall_clock_time_split_group_V2.sh test /scratch/mikem/UserSupport/weizhe.li/runs_split_group/448_400_7691440
  

2 Prediction

• Additionally adjust configuration parameters in files, config_testing_cn_true.txt, config_normal_cn_true.txt, config_tumor_cn_true.txt.

• Run commands listed in the below subsections to launch SGE jobs to generate prediction matrices.

  2.1 With color normalization

• qsub process_main.sh ./config_testing_cn_true.txt
  -- process_main.sh in turn submits a number of SGE jobs using process_array.sh which in turn runs process_images_grp_normalization_wli.py in the array jobs generated for every slide. Number of tasks in an array job determined automatically based on the number of groups in the corresponding HDF5 file.

• qsub process_main.sh ./config_normal_cn_true.txt

• qsub process_main.sh ./config_tumor_cn_true.txt

2.2 Without color normalization

• qsub process_main.sh ./config_testing.txt
  
• qsub process_main.sh ./config_normal.txt
  
• qsub process_main.sh ./config_tumor.txt
  

3 Heatmap stitching

After the predictions matrices have been generated an SGE job using heatmap_main.sh SGE scrip could be launched to genertae heatmaps. Two arguments for this launch are: a) type of the slides (test, normal or tumor); b) the root directory of the results, like in below example run:

• qsub heatmap_main.sh test /scratch/mikem/UserSupport/weizhe.li/runs_process_cn_True/testing_wnorm_448_400_7690953
  

4 Retreiving run-time statistics for prediction

4.1 CPU time

In time_all_stats_pred.sh file adjust job results root directory, DIR and slides type, PREFIX (normal, test or tumor), like below:

• DIR=/scratch/mikem/UserSupport/weizhe.li/runs_process_cn_False/normal_wnorm_448_400_7691563
  
• PREFIX=normal
  Then run:
  

• time bash ./time_all_stats_pred.sh

  4.2 Wall-Clock time

  In wall_clock_time_stats_pred.sh file adjust job results root directory, DIR and slides type, PREFIX (normal, test or tumor), like below:

• DIR=/scratch/mikem/UserSupport/weizhe.li/runs_process_cn_V2_False/testing_wnorm_448_400_7694088

• PREFIX=test
  Then run:

• time bash ./wall_clock_time_stats_pred.sh