Output unit number of Akita

[frostinassiky]

Thanks for the amazing job!

According to the Akita Tutorial, we need to specify model parameters json to have only two targets.

params_file   = './params.json'
with open(params_file) as params_file:
    params_tutorial = json.load(params_file)   
params_tutorial['model']['head_hic'][-1]['units'] =2

ref: https://github.com/calico/basenji/blob/master/manuscripts/akita/tutorial.ipynb

However, the original parameters json have 5 targets. What are the extra 3 targets?

[frostinassiky]

BTW, does the learning curve look good for a production Akita training? Early stopping happens on the 35-th epoch.

[davek44]

The tutorial chooses two of the datasets arbitrarily to demonstrate the code. The primary model that we studied in the paper was trained on the five target datasets described here: https://github.com/calico/basenji/blob/master/manuscripts/akita/data/targets.txt

Yes, those training curves look good. I'm guessing your showing the training set statistics since I don't think early stopping would have chosen to stop if that were the validation set statistics.

[frostinassiky]

Hi @davek44 Thanks for your response! Do you have direct links for the three datasets: GM12878, IMR90, and HCT116?

• they are not in the google storage https://storage.googleapis.com/basenji_hic/tutorials/coolers/GM12878_inSitu_MboI_all.hg38.2048.cool
• they are not in the location specified by https://github.com/calico/basenji/blob/master/manuscripts/akita/data/targets.txt ../../data

[davek44]

If you want all of the datasets, you should consider using the preprocessed dataset into TFRecords, which you can acquire with this script: https://github.com/calico/basenji/blob/master/manuscripts/akita/get_data.sh

If you want all of the cooler files, I added them to the cloud bucket here: https://console.cloud.google.com/storage/browser/basenji_hic/1m/data/coolers