This folder contains the model parameters of Avocado after it is trained on all 1,014 tracks of epigenomic data contained within the Roadmap compendium. There are multiple files that are all formatted as numpy arrays and most easily used in Python. The files are described below:

(1) celltype_embedding.npy
This file contains the learned cell type embeddings and is a matrix of shape (127, 32) because there are 127 cell types that are each represented using 32 latent factors. The ordering of these cell types can be found in `celltypes.npy`, which is an array of 127 strings showing the Roadmap names for each cell type.

(2) assay_embedding.npy
The same as `celltype_embedding.npy` except that each row corresponds to an epigenomic assay instead of a cell type and is of shape (24, 256). The ordering of these assays can be found in `assays.npy`, which is formatted in the same way as `celltypes.npy`.

(3) genome_embedding_chrX.npy
These files each contain the learned embedding for one chromosome. They are of shape (n, 110) where n is the number of 25 bp bins in that chromosome and 110 corresponds to the number of genomic latent factors. This means that each row in these matrices corresponds to the values of a 25 bp bin, not a single nucleotide, i.e., `X = numpy.load("genome_embedding_chr18.npy"); X[100:200]` will not return the representation of coordinates 100 to 200, but rather 2500 to 5000. The first 25 columns correspond to the 25 bp latent factors, and correspondingly change from one row to the next. The next 40 columns correspond to the 250 bp latent factors and so change only every 10 items. The last 45 columns correspond to the 5 kbp latent factors and so change only ever 200 items.