dataset format for training new model

[alexander-ratzan]

Hi there, I'm trying to train a krakencoder model from scratch on a few HCP-YA parcellations that were not included in the original model. I have all the data stored in tsvs as in /example_data and have organized them into /example_data/mydata_4S456Parcels_FCcorrhpf.mat etc. I'm currently matching the format to /example_data/exampledata_10subj_fs86_inputs.mat but I'm noticing that this is not the expected format for the training input data. To validate this I tried using the /example_data/exampledata_10subj_fs86_inputs.mat to train a model and it indeed caused the same error.

Command:  run_training.py --subjectfile example_data/subject_splits_993subj_683train_79val_196test_retestInTest.mat --inputdata [FCcorr_fs86_hpf]@FC=example_data/exampledata_10subj_fs86_inputs.mat [SCifod2act_fs86_volnorm]@SC=example_data/exampledata_10subj_fs86_inputs.mat --dataflavors SCifod2act FCcorr --datagroups SCFC --latentsize 128 --latentunit --transformation pca256 --dropout .5 --losstype correye+enceye.w10+neidist+encdist.w10+mse.w1000+latentsimloss.w10000 --trainvalsplitfrac .8 --valsplitfrac .1 --outputprefix mykraken_example --epochs 10 --checkpointepochsevery 1 --displayepochs 2

Loading subject splits from example_data/subject_splits_993subj_683train_79val_196test_retestInTest.mat
	993 subjects
	683 subjidx_train
	79 subjidx_val
	196 subjidx_test
input_file: example_data/exampledata_10subj_fs86_inputs.mat
groupname: FC
inputfield: None
connfield: None
None of the following fields were found in the input file example_data/exampledata_10subj_fs86_inputs.mat: ['data', 'encoded', 'FC', 'SC', 'C', 'volnorm']
Traceback (most recent call last):
  File "/scratch/*/neuroinformatics/krakencoder/run_training.py", line 1185, in <module>
    run_training_command(sys.argv[1:])
  File "/scratch/*/neuroinformatics/krakencoder/run_training.py", line 516, in run_training_command
    conndata_alltypes[xc]=load_input_data(os.path.expanduser(input_file),group=groupname)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/scratch/*/neuroinformatics/krakencoder/krakencoder/data.py", line 1060, in load_input_data
    raise Exception("Input type not found")
Exception: Input type not found

Can you please provide the specific data format, or even better if you can add an example training data file so that I can match my inputs to that structure? Thanks for the help and great package!

[kjamison]

Those steps could definitely be documented more clearly! The example data file you are using is actually intended for use with the jupyter/colab notebook (which can be more flexible with loading data formats than the command line interface). I just updated the repo to fix/improve a few things, including adding two example files that can be used with run_training.py: example_data/exampledata_10subj_FCcorr_fs86_hpf.mat and example_data/exampledata_10subj_SCifod2act_fs86_volnorm.mat. Each file contains a 'data' field with a list of ROIxROI matrices for each subject, and a 'subjects' field with a list of subject IDs for each matrix. Using these files, you could say, for instance:

python run_training.py \
    --inputdata FCcorr_fs86_hpf@FC=example_data/exampledata_10subj_FCcorr_fs86_hpf.mat \
        SCifod2act_fs86_volnorm@SC=example_data/exampledata_10subj_SCifod2act_fs86_volnorm.mat \
    --latentsize 128 --latentunit --dropout .5 \
    --losstype correye+enceye.w10+neidist+encdist.w10+mse.w1000+latentsimloss.w10000 \
    --epochs 10 --checkpointepochsevery 1 --displayepochs 2 \
    --outputprefix mykraken_example \
    --transformation cfeat --trainvalsplitfrac 1 --valsplitfrac .2 --batchsize 4```

This example is training a new krakencoder model from scratch (not using any information from the existing pre-trained model). In this toy case there are only 10 subjects in the example data, so we can't create a 256-dimensional PCA, and I've just substituted "cfeat" as the transform type (it just centers each dimension of the input). Similarly, I modified the arguments --trainvalsplitfrac 1 --valsplitfrac .2 --batchsize 4 to work for such a small training set.

However, it sounds like you might be trying to side-train a NEW flavor to the EXISTING krakencoder model. If that is the case, you might instead want to focus on this section: Adding a new flavor to an existing model. The idea here is that you have mydata_4S456Parcels_FCcorrhpf.mat with a matrix for each of the original HCP subjects for which we already have a latent space representation, and train a single encoder and decoder to map from your data to that pre-computed latent space. For that, you would want to have a matrix for each subject in the subject_splits_993subj_683train_79val_196test_retestInTest.mat file. The file would have the same structure as example_data/exampledata_10subj_SCifod2act_fs86_volnorm.mat with a 'subjects' field and a 'data' field. Your training command might then look like:

python run_training.py \
    --inputdata FCcorr_4S456Parcels_hpf@FC=mydata_4S456Parcels_FCcorrhpf.mat \
    --encodedinputfile '{KRAKENDATA}/kraken_target_hcp_20240413_210723_ep002000_mse.w1000_encoded.mat' \
    --targetencoding --targetencodingname fusion --onlyselfpathtargetencoding \
    --subjectfile '{KRAKENDATA}/subject_splits_993subj_683train_79val_196test_retestInTest.mat' \
    --latentsize 128 --latentunit --transformation pca256 --dropout .5 \
    --losstype correye+enceye.w10+neidist+encdist.w10+mse.w1000+latentsimloss.w10000 \
    --outputprefix kraken_4S456Parcels_FCcorrhpf.fus \
    --epochs 500 --checkpointepochsevery 500 --displayepochs 100

The necessary {KRAKENDATA}/* files should be automatically downloaded when you run this.